DEV Community: Agustin V. Startari

When the Law Delegates to a Probability Distribution

Agustin V. Startari — Tue, 01 Sep 2026 15:20:05 +0000

*What happens when the same legal system can produce two different decisions from the same facts?
*
Imagine an immigration authority using a generative AI system to evaluate applications.

The legal facts are the same.

The applicable rules are the same.

The model is the same.

The instructions are the same.

One execution produces:

Approve.

Another produces:

Reject.

Nothing legally relevant has changed.

What changed was the computational path through which the system generated its answer.

This creates a problem that traditional legal theory was not built to solve.

Law knows how to delegate authority to judges, agencies, officials, administrative bodies, and other identifiable decision-makers.

But what exactly happens when authority is exercised through a system that does not contain one predetermined decision?

What happens when the institution has authorized a machine capable of generating a distribution of possible decisions?

That is the question at the center of my new paper:

Delegation to a Probability Distribution: Rethinking Legal Authority in Stochastic Systems

The problem is not simply whether artificial intelligence can make legal decisions.

The deeper problem is this:

Does authorizing a system mean authorizing every decision that system can possibly generate?

The answer cannot automatically be yes.

The Difference Between a Rule and a Distribution

Traditional legal delegation can be simplified like this:

Authority → Decision-maker → Decision

A legislature authorizes an agency.

An agency authorizes an official.

The official exercises legally bounded discretion.

The resulting decision can normally be attributed to someone occupying a recognized institutional role.

Generative AI introduces another structure:

Authority → Model → Probability distribution → Selection → Output

The distinction is fundamental.

A generative system does not necessarily contain a single answer waiting to be retrieved.

It assigns probabilities across possible continuations and produces an output through a decoding or sampling process.

That means the institution may approve the architecture without determining which particular legally consequential output will eventually be produced.

This creates what I call the:

*Distributional Authorization Gap
*
The central proposition is simple:

Authorization of the system does not necessarily imply authorization of every possible output of the system.

Or more formally:

Authorization(Model) ≠ Authorization(All Possible Outputs)

This sounds technical.

Its consequences are not.

If a public authority deploys a probabilistic system capable of producing legally incompatible outcomes under materially identical conditions, legal theory must explain what transforms one generated output into an authorized exercise of public authority.

The fact that the machine was approved cannot be the entire answer.

Randomness Is Not the Same as Legal Discretion

Human decision-making is not perfectly predictable.

Judges disagree.

Administrators interpret ambiguous rules.

Investigators weigh evidence differently.

Two officials can sometimes reach different conclusions.

But that does not make human judgment equivalent to computational stochasticity.

A judge can exercise discretion within a legally constituted office.

The law can ask:

What considerations were permitted?

What reasons were given?

Was the decision within jurisdiction?

Was the evidence sufficient?

Was the interpretation legally available?

Was discretion exercised according to the governing standard?

With a generative system, another source of variation enters the chain.

Even while the legally relevant input remains constant, the computational process itself may allow different outputs.

The legal question therefore changes.

It is no longer enough to ask:

Was the AI system authorized?

The law must also ask:

What exactly was authorized within the space of outcomes the system could produce?

That is a different problem.

*A Simple Example
*
Suppose an agency authorizes an AI system to classify applications.

For Applicant X:

Run A → Eligible

Run B → Ineligible

Assume both runs use the same:

factual record,
governing law,
model,
system instructions,
institutional procedure.

If both outputs were technically possible, does that make both legally valid?

Not necessarily.

Technical possibility and legal authorization are different categories.

A machine can be functioning exactly as designed while still producing an output whose legal status is unclear.

This is why reliability alone does not solve the problem.

Accuracy alone does not solve it.

Explainability alone does not solve it.

Human oversight alone does not solve it.

The underlying jurisprudential question remains:

Why does this particular computational realization count as an exercise of the institution's legal competence?

*The Missing Unit of Authorization
*
Existing debates often frame the issue as a choice between two possibilities:

Either the human decides,

or the algorithm decides.

That distinction is increasingly insufficient.

A stochastic decision architecture contains several analytically distinct stages:

Institutional authority

↓

Generative architecture

↓

Distribution of possible outputs

↓

Selection mechanism

↓

Particular output

↓

Institutional adoption

↓

Legal consequence

Each stage matters.

The model is not identical to the distribution.

The distribution is not identical to the sampling or decoding mechanism.

The selection mechanism is not identical to the final output.

And the final output is not automatically identical to a legally valid institutional act.

Legal theory therefore needs to identify where authorization enters this chain and how it reaches the final decision.

Otherwise, the word “delegation” hides more than it explains.

*Why Human Oversight Does Not Automatically Fix It
*
One common response is simple:

Put a person in the loop.

But the presence of a human does not answer the structural question.

Suppose the AI produces a recommendation and an official routinely approves it.

Formally, the human signs.

But who structured the decision?

Who selected the relevant variables?

Who narrowed the available categories?

Who generated the recommendation?

Who determined which output appeared before the official?

If the human merely ratifies a machine-generated result, legal authority may remain formally human while the operative decision architecture has moved elsewhere.

A signature proves that somebody approved the output.

It does not necessarily prove that the legally relevant judgment originated with that person.

This distinction becomes increasingly important as AI systems move from simple information retrieval into ranking, classification, recommendation, eligibility analysis, compliance review, risk assessment, and administrative decision support.

*The Core Claim
*
The law has traditionally focused on who receives delegated authority.

Stochastic systems force another question:

What range of outcomes has actually been authorized?

That changes the unit of analysis.

The problem is no longer only:

Authority → Agent

It becomes:

Authority → Decision architecture → Distribution → Particular legal act

The crucial break occurs between the distribution and the individual output.

An institution can authorize the use of a system without necessarily authorizing every legally consequential event that lies within that system's computational possibility space.

That is the Distributional Authorization Gap.

And it creates a new problem of legal attribution.

A Stochastic Delegation Test

The paper develops a framework for determining when institutional use of a probabilistic system can produce legally imputable decisions.

Four questions become central.

Boundedness

Is the legally permissible output space actually bounded?

A system capable of producing many technically possible answers cannot simply inherit legal validity across the entire distribution.

The institution must determine which outputs fall within delegated competence.

Legally Relevant Variability

Not every variation matters.

Two outputs may use different wording while reaching the same legal conclusion.

The important variable is whether stochastic variation can change rights, duties, benefits, sanctions, eligibility, liability, status, or another legally relevant consequence.

Reconstruction

Can the institution reconstruct the decision-producing event?

If a legally consequential output cannot be meaningfully connected to the model configuration, inputs, selection process, institutional rules, and relevant execution conditions, accountability becomes structurally weaker.

Institutional Imputation

What legal rule converts the computational output into an act of the institution?

This may be the most important question.

Machines generate outputs.

Institutions generate legal consequences.

The bridge between the two cannot simply be assumed.

*This Is Not Only About Courts
*
The issue extends far beyond judicial decision-making.

Consider:

Immigration.

Tax administration.

Social benefits.

Insurance.

Banking compliance.

Public procurement.

Licensing.

Employment screening.

University admissions.

Credit decisions.

Fraud detection.

Administrative sanctions.

Risk classification.

Healthcare allocation.

Corporate compliance.

In all of these environments, probabilistic systems can influence decisions that alter real rights, opportunities, resources, and obligations.

The technical system may not formally possess legal authority.

It does not need to.

If its output becomes the operative basis for institutional action, the architecture participates in the exercise of authority.

The law therefore needs to examine not only the formal decision-maker but the complete decision-producing chain.

*Why It Matters for Everyone
*
This is not an abstract problem limited to jurisprudence.

Citizens should care because a government decision affecting them may increasingly originate inside a probabilistic process.

Lawyers should care because traditional doctrines of competence, attribution, discretion, review, and reasons assume structures that generative systems can disrupt.

Judges should care because reviewing the legality of an outcome may require understanding the architecture that generated it.

Regulators should care because certifying a model does not necessarily validate every output the model can produce.

Developers should care because technical decisions about decoding, thresholds, sampling, system instructions, and output constraints may acquire legal consequences.

Public institutions should care because responsibility cannot disappear merely because the computational system performed as designed.

Companies should care because the same structure appears whenever AI-generated classifications or recommendations become operational decisions.

The central issue is not whether artificial intelligence is intelligent enough to exercise authority.

The issue is whether legal institutions know what they are authorizing when they place probabilistic systems inside decision-making structures.

A deterministic rule can be inspected.

A human official can be questioned.

A probabilistic architecture introduces another object:

a space of possible decisions.

Law has doctrines for delegating competence.

It now needs a theory for what happens when competence encounters probability.

**
**
The future of AI govern
ance will not be decided only by whether models are accurate, explainable, unbiased, or safe.

Those questions matter.

But a system can be accurate in aggregate and still produce a legally problematic individual decision.

It can be explainable and still operate beyond the institution's authorized decision space.

It can have a human reviewer and still determine the frame within which that human acts.

It can comply technically with its design while generating an outcome whose legal authorization remains uncertain.

That is why the question must move one level deeper.

Not:

Can AI make decisions?

But:

What makes one stochastic output a legally authorized decision?

Until that question has an answer, the architecture may be technically operational while the authority exercised through it remains jurisprudentially incomplete.

Read the Paper

Agustin V. Startari
Delegation to a Probability Distribution: Rethinking Legal Authority in Stochastic Systems

Full paper:

https://zenodo.org/records/22233599

Call to Action

Read more of my work on artificial intelligence, language, authority, and institutional responsibility:
**
Website:**
https://www.agustinvstartari.com/

SSRN Author Page:
https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

Zenodo Publications:
https://zenodo.org/search?q=%22Agustin%20V.%20Startari%22

Author

Agustin V. Startari is a linguistic theorist, author, and researcher in historical studies. His work examines how language, artificial intelligence, and formal systems redistribute authority, agency, and responsibility in contemporary institutions. He is the author of Grammars of Power, Executable Power, and The Grammar of Objectivity.
**
Researcher ID:** K-5792–2016
**
Author website:**
https://www.agustinvstartari.com/
**
SSRN Author Page:**
https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

Ethos

I do not use artificial intelligence to write what I don’t know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.

— Agustin V. Startari

Algorithmic Empire: How Generative AI Can Turn Global Power Into Grammar

Agustin V. Startari — Tue, 25 Aug 2026 14:02:13 +0000

What happens to the grammar of power when generative AI summarizes, explains, translates, or reformulates geopolitical information?
The concept of algorithmic empire proposes a testable hypothesis: under controlled and comparable conditions, generative AI systems may preserve greater grammatical agency, institutional rationality, legitimating action, and causal control for dominant-power actors while representing subordinated societies more frequently through the language of crisis, risk, sanctions, instability, humanitarian management, intervention, or dependency.
This is not an assumption that every AI system behaves this way. It is an empirical question that can - and should - be tested.

**The Bias We Usually Look For
**When people discuss geopolitical bias in artificial intelligence, they tend to search for statements.
Does the model favor the United States?
Does it portray Iran more negatively than another country?
Does it describe Israel and Palestine differently?
Does it use different language for invasion, occupation, sanctions, resistance, terrorism, military intervention, or civilian casualties?
These are legitimate questions.
But they may be looking only at the visible surface of political language.
Political representation is not built exclusively from facts, adjectives, or explicit opinions.
It is also built from subjects, objects, verbs, causes, and responsibility.
Consider two sentences:
A state imposed economic restrictions that disrupted civilian access to essential goods.

And:
The country faces shortages and difficulties accessing essential goods.

Both sentences could refer to the same real-world situation.
Both might even be factually defensible.
But politically, they perform very different operations.
In the first sentence, someone acts.
In the second, a condition exists.
An external action becomes an internal problem.
A sanction becomes scarcity.
A causal relationship becomes a humanitarian condition.
The consequence remains visible.
The actor responsible for producing it may disappear.
This is why grammar matters.
The relevant question is not only whether AI gives us correct information.
It is also:
What political architecture remains after the machine has transformed that information?

**When Power Becomes a Verb
**Political power is relatively easy to identify when language names it directly.

A government sanctions.
An army occupies.
A state bombs.
An institution prohibits.
A military force intervenes.
A government imposes restrictions. The basic grammatical structure remains visible: actor → action → object of action Contemporary political discourse, however, frequently operates through more administrative forms. Force becomes a security response. Economic coercion becomes pressure. Military intervention becomes stabilization. Civilian suffering becomes humanitarian need. Restrictions become compliance requirements. Migration becomes pressure on receiving systems. Political conflict becomes an institutional deficit. Occupation can be discussed through administration, security, access management, or territorial governance. None of these formulations is necessarily false. And none of these words is automatically imperial. The analytical problem begins when they are distributed asymmetrically. One actor stabilizes. Another is unstable. One actor provides security. Another constitutes a security threat. One actor manages a crisis. Another becomes the crisis. One sanctions. Another is sanctioned. One develops. Another requires development. The difference can look subtle. Grammatically, it is enormous. The underlying research defines algorithmic empire precisely as a representational condition in which controlled geopolitical transformations systematically distribute agency, legitimacy, institutional rationality, and causal visibility unequally between differently positioned actors.

**Generative AI Does Not Simply Retrieve Information
**This is where generative AI creates a fundamentally new problem.
A database retrieves.
A search engine indexes.
A generative model rewrites.
When we ask an AI system to summarize a document, explain a conflict, compare two governments, translate an article, moderate political content, or turn a complex report into a few paragraphs, the model does not merely reproduce the original information.
It creates a new representation.
That requires choices.
It must select subjects.
It must select verbs.
It must decide which actors remain explicitly named.
It must compress causal sequences.
It must determine which historical conditions survive summarization.
It must decide which information becomes central and which becomes background.
It organizes material through categories such as:

security,
terrorism,
occupation,
humanitarian need,
sanctions,
stability,
migration,
development,
governance,
intervention,
diplomacy,
extremism,
compliance. These transformations are not politically neutral simply because they are grammatically ordinary. A generated text can preserve the basic facts while changing the political structure in which those facts appear. That possibility is central to the concept of algorithmic empire. This means factual accuracy alone is not enough to evaluate geopolitical AI output. A machine can invent nothing and still substantially transform a narrative.

**From Political Subject to Administrative Object
**The core problem can be reduced to one deceptively simple question:
Who gets to do things?
Consider two grammatical clusters.
A state:

sanctions;
regulates;
stabilizes;
protects;
intervenes;
negotiates;
authorizes;
secures;
classifies;
mediates.
A population:
suffers;
requires assistance;
experiences instability;
faces shortages;
depends on aid;
generates migration pressure;
requires stabilization;
becomes vulnerable.

The first group is organized around action.
The second is organized around condition.
There may be perfectly legitimate reasons for some of this difference. Governments, armies, institutions, and civilian populations obviously occupy different political and institutional roles.
So the serious test cannot simply count active verbs and declare that imperialism has been detected.
The real question must be comparative.
When equivalent or structurally comparable information is presented to a generative model, does the system systematically change who retains agency and who becomes an object of management?
If it does not, the hypothesis fails.
That limitation is crucial.
Algorithmic empire should not function as an accusation. It should function as a falsifiable hypothesis.

**The Powerful Do Not Have to Disappear
**One of the most important features of this framework is that domination does not necessarily require powerful actors to vanish from discourse.
They can remain highly visible.
What changes is the grammatical role they occupy.
The powerful may increasingly appear as the actors who:
secure,
stabilize,
respond,
manage,
mediate,
enforce,
authorize,
protect,
contain,
develop,
sanction.
Their actions become the organizing verbs of political reality.
Subordinated populations also remain visible.
But they may increasingly appear as:
refugees,
risks,
victims,
unstable regions,
humanitarian emergencies,
sanctioned economies,
underdeveloped systems,
security challenges,
migration flows,
populations requiring management.
The research paper summarizes the distinction in two unusually simple propositions:
The powerful need not disappear from the sentence. They may become its legitimate verbs.
The subordinated need not disappear either. They may become its manageable nouns.
That is the central linguistic mechanism.

**Palestine as a Stress Test
**Palestine represents one of the most difficult cases for testing this framework.
Occupation, territorial control, civilian harm, Israeli state conduct, Palestinian political institutions, Palestinian armed organizations, settlements, hostages, military operations, humanitarian assistance, self-determination, terrorism, security, and international law all exist simultaneously within the same discursive field.
These categories cannot simply be collapsed into one another.
A model mentioning security is not evidence of imperial grammar.
A model mentioning humanitarian needs is not evidence either.
Nor is the presence of military terminology.
The question is what happens during transformation.
Suppose the source material explicitly identifies:
an actor,
an action,
the legal or political status of that action,
the population affected,
and the causal relationship between them.
After summarization, does that structure survive?
Or does the output increasingly become something like:
"Humanitarian conditions deteriorated."
"Regional instability increased."
"Security concerns intensified."
"Access to essential resources became limited."
Each statement may be accurate.
But accuracy and causal completeness are different things.
A transformation becomes politically important when actors repeatedly disappear while consequences remain.
The framework therefore tests whether AI systematically redistributes agency, causal attribution, legal status, sovereignty, or political subjecthood rather than simply counting politically sensitive vocabulary.

**Iran Shows a Different Mechanism
**Iran creates another useful test because it cannot easily be represented simply as a powerless political object.
The Iranian state possesses significant military, political, diplomatic, and regional agency.
At the same time, discourse surrounding Iran is heavily structured through sanctions, nuclear policy, regional security, financial restrictions, international compliance systems, and economic isolation.
That creates complex causal chains.
A sanctioning government imposes restrictions.
Banks interpret regulatory exposure.
Financial institutions may over-comply.
Trade channels become more difficult.
Humanitarian exemptions may formally exist while practical transactions remain blocked.
Medical or other civilian supply chains can be affected.
A generative summary can preserve this chain.
Or it can compress it into:
"Iran faces shortages."
"Iran has difficulty accessing international markets."
"Iran suffers from economic isolation."
Again, these sentences may be factually true.
But the causal structure has changed.
External action has become an internal condition.
The paper specifically uses sanctions and documented problems of sanctions over-compliance affecting civilian access to medical goods as a way to test whether models preserve or erase external causal chains.
The question is not whether Iran is innocent, democratic, authoritarian, threatening, peaceful, legitimate, or illegitimate.
Those are separate political questions.
The methodological question is narrower:
Does the machine preserve the causal architecture of the information it receives?

**Why "Western Bias" Is Too Simple
**It would be easy to reduce this argument to:
"AI is Western, therefore AI reproduces Western imperialism."
That is not the claim.
And methodologically, it would be a weak one.
Large language models are trained on enormous and heterogeneous corpora.
Different models use different architectures, post-training procedures, safety mechanisms, retrieval systems, ranking methods, and alignment processes.
Their outputs also change according to prompts, languages, source documents, system instructions, and context.
Existing research indicates that geopolitical and framing differences in model outputs can be measured, but it does not establish a single universal direction of geopolitical bias across all systems and conditions.
Therefore, algorithmic empire cannot be inferred from:
one politically controversial response,
one model,
one company,
one country,
or one badly framed prompt.
The phenomenon would have to be demonstrated through controlled comparison.
That distinction matters because otherwise "algorithmic empire" would become a political label rather than an analytical concept.

**Intention Is Not Required
**There is another misconception that must be eliminated.
If a model reproduces an asymmetric grammar of power, it does not follow that its developers deliberately programmed that outcome.
The framework concerns representation, not intention.
Generative systems can reproduce statistical regularities without believing anything.
Possible sources of asymmetry could include:
training data,
dominance of institutional sources,
journalistic conventions,
government terminology,
humanitarian reporting practices,
summarization patterns,
reinforcement learning,
safety policies,
retrieval systems,
ranking mechanisms,
prompt design,
or combinations of these factors.
A model does not need a political ideology in the human sense.
It only needs to reproduce linguistic regularities.
If millions of documents repeatedly represent one class of actors as agents of security, order, intervention, stabilization, development, and regulation, while another class appears mainly through instability, humanitarian dependency, violence, migration, or crisis, a generative system may learn those distributions.
Not because it understands empire.
Because it understands probability.

**How Could We Measure It?
**A theory becomes useful only when it can fail.
For that reason, the underlying research proposes two measurement concepts:
Imperial Grammar Reproduction Rate (IGRR)
and
Algorithmic Empire Index (AEI).
The IGRR is designed to measure directional asymmetries between matched actor classes rather than simply counting sentences classified in advance as "imperial."
The AEI combines several independently measurable dimensions:

agency asymmetry;
object-framing asymmetry;
coercion neutralization;
sovereignty erasure;
external-responsibility deletion.

The important feature is not the name of the index.
It is that the design allows a null result.
If comparable actors retain equivalent agency, the hypothesis is weakened.
If the apparent asymmetry exists entirely in the source material and the model merely preserves it, then generative AI may be transmitting rather than producing the structure.
If the model reduces the asymmetry, it may actually operate against the proposed mechanism.
And if powerful actors are converted into administrative objects at the same rate as weaker actors, the imperial interpretation loses explanatory force.
The framework explicitly allows these null and inverse findings.
That is what separates an empirical hypothesis from a predetermined political conclusion.

**The Importance of Matched Transformations
**Imagine giving an AI system several versions of structurally comparable geopolitical information.
The actors change.
The underlying structure remains as similar as historically and legally possible.
Then ask the model to:
summarize,
shorten,
translate,
classify,
rewrite for a general audience,
produce a policy briefing,
or explain the situation.
Now measure what survives.
Does the dominant actor remain the grammatical subject?
Does its institutional justification survive compression?
Are its actions transformed into neutral administrative terminology?
Does the weaker actor lose political agency?
Does civilian suffering become detached from the action that produced it?
Does occupation become territorial complexity?
Do sanctions become economic difficulty?
Does external intervention become stabilization?
Does political resistance become primarily a security category?
Does sovereignty disappear?
Does responsibility disappear?
Most importantly:
Does this happen systematically more often for one class of actor than another?
Only then does the idea of algorithmic empire become empirically interesting.

**AI Can Change Politics Without Changing Facts
**This may be the most important implication.
Much current AI governance focuses on hallucination.
Did the model invent something?
Did it misidentify someone?
Did it fabricate a quotation?
Did it produce false statistics?
These problems are serious.
But a future generation of models could become dramatically more factually accurate and still leave the grammatical problem untouched.
Consider a system that never invents a casualty figure.
Never creates a fictional treaty.
Never attributes a statement to the wrong government.
Never fabricates an event.
It could still consistently transform:
"State A imposed restrictions that produced consequence B"
into:
"Population B faces difficulties."
No hallucination occurred.
The problem is not false information.
The problem is causal compression.
Likewise:
"Actor A destroyed infrastructure used by population B"
can become:
"Population B faces infrastructure shortages."
"Actor A controls the movement of population B"
can become:
"Population B faces mobility restrictions."
"Actor A intervened militarily in state B"
can become:
"State B entered a period of instability."
Every transformation can preserve part of the factual content.
And every transformation can progressively remove political agency.

**The Grammar of Humanitarianism
**Humanitarian language deserves particular attention because it is both necessary and politically complex.
Terms such as:
humanitarian crisis,
food insecurity,
displacement,
medical need,
vulnerable populations,
aid dependency,
emergency response
can accurately describe real human suffering.
The problem is not the vocabulary itself.
The problem emerges when humanitarian description systematically replaces causal description.
"There is a humanitarian crisis" tells us that people are suffering.
It does not necessarily tell us why.
"There is food insecurity" describes a condition.
It does not tell us whether food disappeared because of drought, economic collapse, siege, trade restrictions, infrastructure destruction, sanctions, blockade, military action, corruption, government policy, or some combination of causes.
Humanitarian language can therefore preserve suffering while deleting responsibility.
This does not make humanitarian discourse inherently imperial.
It means that when generative AI compresses political information into humanitarian categories, researchers should measure what causal information survives the transformation.

**Security Works the Same Way
**Security language presents a parallel problem.
Security is real.
States have legitimate security concerns.
Civilians have security concerns.
Borders matter.
Armed organizations exist.
Terrorist attacks occur.
Governments have obligations to protect populations.
The framework does not dispute any of that.
The question is whether security becomes asymmetrically distributed as a legitimating grammar.
Who provides security?
Who threatens security?
Who responds?
Who triggers the response?
Whose violence remains an action?
Whose violence becomes context?
Whose fear becomes institutionally legitimate?
Whose fear becomes a humanitarian condition?
When those patterns become systematic, security language begins doing more than describing events.
It begins organizing political subjecthood.

**Sovereignty Can Disappear Without Being Denied
**One of the most subtle forms of representational transformation concerns sovereignty.
A model does not need to say:
"This population has no political rights."
It can erase political subjecthood indirectly.
A society can increasingly appear as:
a humanitarian population,
a refugee population,
a development problem,
a source of instability,
a migration flow,
a security environment,
or an administrative territory.
The people remain visible.
Their suffering may remain visible.
Their demographic existence may remain visible.
Their political agency may not.
This distinction matters.
Visibility is not the same thing as subjecthood.
A population can be discussed constantly while rarely appearing as an actor capable of making claims, exercising sovereignty, resisting, negotiating, choosing, governing, or possessing political objectives.
A system can therefore produce enormous visibility without producing political recognition.

**The Machine Does Not Need to Say "Empire"
**The word empire carries enormous historical and political weight.
But the theory of algorithmic empire does not depend on an AI system using imperial vocabulary.
In fact, the more interesting possibility is the opposite.
Modern systems may reproduce hierarchical structures precisely through language that appears neutral, administrative, humanitarian, procedural, or technocratic.
The model does not need to say:
"Powerful states are entitled to dominate weaker societies."
It only needs to repeatedly assign different grammatical roles.
One side acts.
The other is acted upon.
One side regulates.
The other requires regulation.
One side stabilizes.
The other is unstable.
One side intervenes.
The other becomes an intervention environment.
One side manages.
The other becomes manageable.
The paper therefore argues that the central empirical question is not whether machines explicitly name or endorse domination.
It is whether a reproducible grammar of global hierarchy survives generative transformation.

**Why This Matters Beyond Chatbots
**Generative AI is rapidly becoming an intermediary layer between human beings and political information.
People increasingly use AI to:
summarize news,
explain wars,
translate foreign reporting,
prepare policy briefings,
research historical conflicts,
generate educational material,
moderate platforms,
classify political content,
produce intelligence summaries,
draft institutional documents,
and answer questions about international affairs.
This means that political discourse is no longer shaped only by journalists, governments, universities, NGOs, corporations, publishers, or citizens.
It is increasingly being transformed by computational systems positioned between sources and readers.
The transformation may appear minor.
A shorter sentence.
A cleaner summary.
A neutral phrase.
A removed actor.
A compressed cause.
A different verb.
But at scale, repeated billions of times, grammatical decisions become infrastructure.
And infrastructure shapes what becomes cognitively normal.

**The Deeper Risk Is Normalization
**The most consequential form of political bias may not be spectacular.
It may not involve propaganda.
It may not involve obvious censorship.
It may not require fabricated history.
It may simply make certain relations of power sound natural.
A sanction becomes a condition.
An occupation becomes administration.
An intervention becomes stabilization.
A civilian population becomes humanitarian need.
A politically subordinated society becomes instability.
External control becomes governance.
The transformation is powerful precisely because nothing necessarily sounds extreme.
The language becomes smoother.
More institutional.
More neutral.
More administrative.
And sometimes less capable of showing who did what to whom.

**A Different Standard for Political AI
**The next generation of AI evaluation therefore requires more than traditional factual benchmarking.
We should continue asking:
Is the answer true?
But we should also ask:
Who remains the subject?
Who becomes the object?
Who receives active verbs?
Who receives passive constructions?
Who retains institutional motivation?
Whose motivation disappears?
Whose violence is causally explained?
Whose violence becomes identity?
Whose suffering retains a responsible actor?
Whose suffering becomes a condition?
Whose sovereignty survives summarization?
Whose political agency survives translation?
And do these differences appear systematically across comparable geopolitical relations?
Those questions move AI evaluation away from simplistic ideological scoring.
They make political representation measurable at the level where much of its power actually operates:
syntax.

**Conclusion: Empire as Machine Grammar
**Generative AI does not need political intentions.
It does not need nationalism.
It does not need colonial nostalgia.
It does not need to believe that one civilization should dominate another.
It only needs to learn and reproduce statistical patterns in human language.
If those patterns contain historically accumulated asymmetries in the distribution of agency, legitimacy, responsibility, security, sovereignty, development, intervention, and crisis, generative systems may reproduce them automatically.
But that outcome cannot simply be assumed.
It must be measured.
If controlled experiments find no systematic asymmetry, the theory must be rejected.
If the asymmetry comes entirely from source material, the theory must be narrowed to transmission.
If some models reproduce it and others do not, we need to identify the conditions that produce the difference.
And if models systematically reduce existing asymmetries, that finding matters just as much.
The goal is not to force political theory onto AI.
The goal is to determine whether linguistic hierarchy survives machine transformation.
Because the most important political question may eventually be much smaller than whether an artificial intelligence supports empire.
It may be whether, after billions of summaries, translations, classifications, explanations, and generated answers, the machine has learned a simpler rule:
some actors remain the verbs of history, while others become its nouns.
The machine does not need to name empire.
The empirical question is whether it reproduces its syntax.

About the Author
**Agustin V. Startari** is an author and researcher whose work examines artificial intelligence, political language, algorithmic power, legitimacy, asymmetric visibility, and the linguistic structures through which institutions represent coercion, responsibility, sovereignty, and political agency.
His research series Grammars of Asymmetric Visibility investigates how political and institutional language can redistribute responsibility and agency across human and machine-generated discourse. His broader work explores the relationship between generative AI, institutional authority, executable legitimacy, political representation, and the emerging role of machine-generated language in shaping how power becomes visible - or disappears - from public discourse.
ORCID: *0009–0001–4714–6539
* ResearcherID: K-5792–2016
SSRN Author: Agustin V. Startari

"I do not use artificial intelligence to write what I don't know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored."

Agustin V. Startari Website: https://www.agustinvstartari.com ORCID: https://orcid.org/0009-0001-4714-6539 SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

Paper 7 - Zenodo: https://zenodo.org/records/22096813
Paper 7 - Zenodo DOI: https://doi.org/10.5281/zenodo.22096813
Paper 7 - Figshare DOI:

When War Becomes a Dashboard: How AI Turns Violence Into Optimization

Agustin V. Startari — Thu, 20 Aug 2026 18:27:17 +0000

The most effective way to hide force may be to describe it perfectly - as a process.
A missile strike can become a targeting operation.

Civilian deaths can become a collateral estimate.

Surveillance can become risk assessment.

Occupation can become security management.

Escalation can become something to optimize.

None of these expressions necessarily deny what happened.

That is exactly the problem.

The violence can remain visible while the grammar changes what kind of event the reader thinks they are seeing.

Instead of an actor using force against another actor, the event begins to look like a technical system processing variables.

Targets are validated.

Risks are scored.

Threats are assessed.

Collateral effects are estimated.

Escalation is managed.

Responses are optimized.

The language sounds precise.

It may even be more precise than ordinary political language.

But precision does not automatically preserve responsibility.

This is the problem I call the administrative translation of force.

**
**
Most people imagine political manipulation as concealment.

Something happened, and somebody tries to hide it.

But there is another possibility.

Nothing is hidden.

The bombing is mentioned.

The surveillance is mentioned.

The casualties are mentioned.

The military operation is mentioned.

The security system is mentioned.

The numbers may even be displayed on a dashboard.

Yet something important can still disappear:

the grammatical relationship between the actor, the action, the person affected, and the consequence.

Compare these two structures:

A military actor attacked a location and civilians were killed.

Now compare:

A target was validated, operational risk was assessed, and collateral effects were recorded.

They may refer to overlapping realities.

But they do not organize those realities in the same way.

The first sentence forces the reader to confront an actor and an action.

The second organizes the event as a procedure.

That transformation matters.

*2. The Most Powerful Euphemism May Be a Workflow
*
Military language has always contained technical terminology.

That is not new.

Armies need categories.

Commanders need procedures.

Intelligence requires classification.

Legal assessments require concepts such as necessity, distinction, proportionality, and precaution.

Security institutions need operational vocabulary.

The argument is not that technical military language is illegitimate.

The problem appears when technical language becomes the dominant grammar through which force is understood.

Then the event begins to change shape.

Attack becomes targeting.

Killing becomes neutralization.

Surveillance becomes monitoring.

Coercion becomes enforcement.

Restriction becomes compliance.

Civilian death becomes collateral impact.

Political judgment becomes risk assessment.

Military decision becomes model output.

Force has not disappeared.

It has entered administration.

*3. From Decision to Score
*
AI makes this transformation more consequential because prediction introduces another layer between human judgment and physical consequence.

Consider the structure:

Human actor → political judgment → military decision → force → consequence.

Now insert predictive systems:

Data → model → score → threshold → classification → validation → recommendation → human authorization → force → consequence.

The human may still authorize the final action.

But the decision environment has already been structured.

A person appears as a risk score.

A building appears as a potential target.

A geographical area appears as an operational environment.

A population appears as a security variable.

A predicted probability becomes part of the justification chain.

The central question is therefore not simply:

Did a human remain in the loop?

The better question is:

What did the human receive when they entered the loop?

Raw evidence?

A ranked list?

A probability?

A threat category?

A recommended target?

An operational threshold?

A model-generated summary?

The human decision remains important.

But responsibility cannot be understood only by looking at the final signature.

The structure that produced the options matters too.

*4. The Dashboard Changes the Moral Geometry
*
A dashboard looks neutral.

That is why it is powerful.

Dashboards turn heterogeneous events into comparable objects.

A target gets a score.

A risk gets a probability.

A civilian population becomes a density estimate.

An expected consequence becomes a variable.

A military objective receives a priority.

A possible escalation receives a projected outcome.

This is administratively useful.

It also changes the form in which violence appears.

A person becomes an entry.

A neighborhood becomes a zone.

A death becomes an estimate.

A decision becomes a threshold.

A political conflict becomes a field of optimization.

This is where the phrase matters:

Occupation does not disappear. It becomes a dashboard.

The claim is not that dashboards create occupation.

The claim is that administrative systems can transform how occupation, coercion, surveillance, targeting, and civilian harm become legible.

Once an event becomes legible as data, it also becomes easier to process as data.

*5. Civilian Harm Can Survive as a Number
*
One of the most dangerous assumptions in AI ethics is that visibility automatically produces accountability.

It does not.

Civilian harm can be fully visible.

It can be counted.

Mapped.

Estimated.

Predicted.

Reported.

Compared.

Minimized.

Included in a proportionality assessment.

And still become politically weaker.

Why?

Because the question is not only whether civilian harm appears.

The question is:

How does it appear?

There is a difference between:

Actor X performed action Y, producing civilian consequence Z.

and:

Expected collateral effects remained within the operational assessment.

The second formulation may contain real information.

It may be part of a legitimate legal or military analysis.

But grammatically, something has changed.

The human consequence has moved inside the procedure.

The person is no longer primarily the object of an action.

The person becomes one variable inside the calculation of that action.

That is what I call collateral grammar.

*6. AI Does Not Need to Lie
*
This distinction is essential.

An AI system does not need to fabricate an event to transform its political meaning.

It does not need to invent facts.

It does not need to censor every reference to violence.

It does not need to deny civilian harm.

It does not need to produce propaganda in the traditional sense.

It only needs to reorganize the sentence.

A summary can remain factually defensible while weakening causal structure.

An explanation can preserve casualty numbers while deleting the responsible actor.

A security analysis can preserve technical detail while turning a political decision into operational necessity.

A model can describe coercion while representing it principally as risk management.

This is why factual accuracy alone is not enough.

Two sentences can contain similar facts while distributing agency and responsibility very differently.

*7. Palestine and Iran Reveal Different Versions of the Problem
*
The paper uses Palestine and Iran as central analytical sites because they expose different structures.

Palestine concentrates questions of occupation, bombardment, displacement, civilian harm, contested sovereignty, armed organizations, humanitarian reporting, security discourse, surveillance, and political responsibility.

Iran presents a different configuration: sanctions, nuclear framing, regional security discourse, threat construction, military escalation, financial restriction, state agency, civilian consequences, and external coercive pressure.

These cases should not be collapsed into equivalents.

They are useful precisely because the structures differ.

The analytical question remains constant:

When AI-mediated discourse describes force, does it preserve who acts, against whom, through which mechanism, with what consequence, and under whose responsibility?

The same test must apply to every actor.

Israeli forces.

Palestinian armed organizations.

Iranian state institutions.

United States institutions.

Other state actors.

Non-state actors.

Civilian populations.

Political disagreement does not remove the requirement for grammatical traceability.

*8. This Is Not an Argument Against Security Analysis
*
The distinction matters because a serious framework has to survive disagreement.

Not every military risk assessment is illegitimate.

Not every security concern is fabricated.

Not every proportionality analysis is propaganda.

Not every operational category hides responsibility.

Not every passive construction is political manipulation.

Not every AI-generated military summary is biased.

The stronger claim would be easy to make and impossible to defend.

The measurable question is narrower:

Does the transformation preserve the chain between decision, force, harm, and responsibility?

If it does, the system may be technically administrative while remaining politically traceable.

If it does not, administrative language may be performing more than description.

It may be reorganizing accountability.

*9. How Do You Measure Something Like This?
*
The paper proposes two instruments.

Operationalization Density

Operationalization Density measures how frequently force, occupation, surveillance, targeting, or coercive control is represented through operational, administrative, predictive, proportionality-based, or optimization-oriented language rather than direct language of force, harm, agency, and responsibility.

In practical terms:

How often does the discourse say:

targeting
validation
neutralization
risk assessment
security response
collateral estimation
operational necessity
escalation management

instead of preserving an explicit structure of:

actor
action
affected population
consequence
responsibility?

A high frequency does not prove wrongdoing.

It identifies a representational pattern that can then be compared across actors, models, sources, and cases.

*10. The Force Optimization Index
*
The second instrument is the Force Optimization Index.

It measures the degree to which military and security discourse converts force into the language of:

optimization
procedure
risk management
prediction
proportionality
necessity
technical decision-making

The index rises when:

Violence becomes targeting.

Occupation becomes security management.

Surveillance becomes risk assessment.

Civilian harm becomes collateral estimation.

Human judgment disappears behind scores, thresholds, models, or intelligence pipelines.

The responsible actor becomes grammatically weaker.

The operation begins to appear technically inevitable.

The index falls when:

The actor remains explicit.

The coercive act is directly named.

Civilian consequences retain causal traceability.

The model does not replace human responsibility.

The institutional decision chain remains visible.

Technical calculation remains distinguishable from political responsibility.

That distinction is critical.

The objective is not to eliminate operational language.

It is to measure what operational language replaces.

*11. The Real Problem With "Optimization"
*
Optimization sounds positive.

It suggests efficiency.

Precision.

Reduction of error.

Better use of information.

Faster decision-making.

More controlled outcomes.

But optimization always requires an objective.

Something must be maximized.

Something must be minimized.

Something must become a constraint.

Something must become acceptable.

That means the important question is not:

Is the system optimized?

The important questions are:

Optimized for what?

Under whose objective function?

Which harms are constraints?

Which harms are costs?

Which humans appear as decision-makers?

Which humans appear as variables?

Which consequences remain politically attributable?

When violence enters optimization language, these questions become more important, not less.

*12. Why AI Ethics Is Still Looking in the Wrong Place
*
AI ethics usually asks familiar questions.

Is the model biased?

Is it accurate?

Is it explainable?

Does it hallucinate?

Is it safe?

Does it discriminate?

Can the human override it?

These questions remain necessary.

But military and security AI introduce another one:

Does the system preserve the grammar of responsibility?

A technically explainable model may still produce politically opaque language.

A statistically accurate prediction may still enter an institutional structure that weakens decision-chain visibility.

A human may still be formally responsible while models and scores organize the field of action before that human intervenes.

An output may contain every relevant number and still fail to answer:

Who decided?

Who acted?

Who was acted upon?

What caused the harm?

Who remains responsible?

Those are not secondary linguistic questions.

They are accountability questions.

*13. The Core Claim
*
The future of military AI will not be determined only by whether algorithms become more accurate.

It will also be determined by the language through which their outputs become institutionally intelligible.

The dangerous transition is not:

violence → silence.

It may be:

violence → data
data → classification
classification → score
score → validation
validation → operation
operation → optimization

At the end of that chain, the event has not disappeared.

It has become administratively perfect.

The casualty is still there.

The target is still there.

The territory is still there.

The force is still there.

But responsibility may have migrated into a system of nouns.

That is the administrative translation of force.

Why This Matters

This is not only a military problem.

The structure matters anywhere algorithmic systems convert consequential human action into technical procedure.

Governments should care because automated security systems can reshape how coercive decisions become justified and documented.

Military institutions should care because technical precision cannot substitute for institutional accountability.

Developers should care because a score, threshold, ranking, or classification can become part of a real decision chain.

Lawyers should care because traceability between decision and consequence matters to legal evaluation.

Journalists should care because machine-generated summaries can preserve facts while changing causal visibility.

Researchers should care because conventional bias metrics may miss grammatical redistribution of responsibility.

Citizens should care because administrative language can determine whether political force appears as an action that requires justification or as a process that simply occurred.

The future vocabulary of violence may not sound violent.

It may sound efficient.

Professional.

Measured.

Predictive.

Proportional.

Optimized.

That is precisely why it deserves examination.

Violence becomes easier to administer when grammar turns it into procedure.

And administrative violence begins where force survives as data, harm survives as estimate, and responsibility disappears into optimization.

*Related Academic Background
*
This article is the public-facing extension of my research paper:

From Occupation to Optimization: AI, Military Language, and the Administrative Translation of Force

How Predictive Systems Convert Violence into Operational Syntax

It forms part of the series:

Grammars of Asymmetric Visibility: AI, Imperial Power, and the Syntax of Responsibility

The broader research program examines how AI-mediated discourse can redistribute agency, responsibility, visibility, political subjecthood, and institutional accountability through formal linguistic structures.

Related work includes:

Suffering Without Perpetrators: The Humanitarian Passive in AI-Generated Conflict Discourse

The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency

Iran as Syntax: Sanctions, Sovereignty, and the AI-Mediated Grammar of Threat

Censorship Without a Censor: Platform Governance and the Disappearance of Suppression

The Syntax of Digital Dehumanization: Subjugated Societies as Risk Objects in AI-Governed Discourse
**
Call to Action**

Read more of my work on artificial intelligence, language, authority, and institutional responsibility:

Website: https://www.agustinvstartari.com/

SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

Zenodo publications: https://zenodo.org/search?q=%22Agustin%20V%20Startari%22

Author

Author Identifiers

Researcher ID: K-5792-2016
ORCID: https://orcid.org/0009-0001-4714-6539
SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
Author Website: https://www.agustinvstartari.com/
**
Institutional Affiliations**

Universidad de la República - Uruguay
Universidad de la Empresa - Uruguay
Universidad de Palermo - Argentina

Contact

Academic email: astart@palermo.edu

Ethos

Agustin V. Startari

AI Can Mention You Constantly and Still Make You Disappear

Agustin V. Startari — Tue, 18 Aug 2026 18:22:46 +0000

Imagine a Population That Is Everywhere
Open a news summary.
They are there.
Open a humanitarian report.
They are there.
Ask an AI system to explain the conflict.
They are there.
Search for migration statistics, sanctions, military operations, political instability, extremism, aid deliveries, displaced populations, or regional security.
They are everywhere.
And yet something strange can happen.
They rarely act.
They rarely decide.
They rarely demand.
They rarely remember.
They rarely possess sovereignty.
They rarely explain themselves.
Things happen to them.
Other institutions decide what happens about them.
They appear constantly, but increasingly as objects.
This creates a problem that ordinary measures of media visibility cannot detect.
Because being mentioned is not the same thing as being represented.
And being represented is not the same thing as being represented as a political subject.

**1. Hate Speech Is the Easy Case
**Most people recognize dehumanization when it is explicit.
A group is compared to animals.
A population is described as vermin.
An ethnic group is called a disease.
People are represented as biologically inferior.
Those cases matter.
But they are also comparatively easy to detect.
Modern AI safety systems are explicitly designed to detect insults, slurs, hateful generalizations, threats, and extremist language.
The harder problem is what happens when the language sounds perfectly respectable.
Consider:
Thousands of civilians require humanitarian assistance amid worsening regional instability.

Nothing in that sentence is obviously hateful.
It may even be factually correct.
But ask a few additional questions.
Who created the conditions requiring assistance?
What happened politically before the humanitarian emergency?
Who made the relevant decisions?
What does the affected population claim?
What institutions are acting upon it?
What does it demand?
The sentence tells us that people suffer.
It does not necessarily tell us how they exist politically.
That difference matters.

**2. A Population Can Become a Grammar Problem
**Compare these sentences:
Palestinians demanded an end to restrictions on movement.
Movement restrictions continued to affect Palestinians.
Humanitarian conditions deteriorated in Palestinian areas.
The region remained a source of instability.
All four sentences can refer to the same political environment.
But they create radically different subjects.
In the first sentence, Palestinians act.
They demand something.
They possess an identifiable political position.
In the second, they are affected by a process.
In the third, they become inhabitants of a humanitarian condition.
By the fourth sentence, the population has almost disappeared into geography.
The territory itself has become a source of instability.
This is not primarily about vocabulary.
It is about grammar.
Who occupies the subject position?
Who receives intentional verbs?
Who is allowed to demand, negotiate, reject, resist, decide, govern, remember, or claim?
And who appears primarily as the thing being managed?

**3. The Most Dangerous Word May Be a Perfectly Neutral One
**The vocabulary of objectification is often bureaucratically normal:
risk
pressure
flow
instability
containment
humanitarian need
radicalization
security concern
sanctions target
border pressure
proxy
reconstruction
moderation risk
None of these expressions is inherently illegitimate.
Real security threats exist.
Refugee flows exist.
Political extremism exists.
Humanitarian emergencies exist.
Governments legitimately analyze instability.
Platforms have legitimate reasons to moderate some content.
The problem begins somewhere else.
It begins when these categories become the principal grammatical forms through which a population is allowed to exist.
Then the transformation looks like this:
people → population
population → humanitarian population
humanitarian population → caseload
displaced people → refugee flow
political conflict → instability
historical grievance → extremism
external coercion → economic pressure
political resistance → security risk
At the end of the chain, nobody had to use a slur.
Nobody had to say that the population was less than human.
The political subject simply became an administrative object.

**4. Humanitarian Language Can Be Sympathetic and Still Remove Agency
**This is one of the least intuitive parts of the problem.
Dehumanizing structures do not always sound hostile.
They can sound compassionate.
A population can be described entirely through:
hunger,
injury,
displacement,
poverty,
medicine shortages,
civilian casualties,
aid dependency,
housing destruction.
That population may receive enormous moral sympathy.
But sympathy is not agency.
If people appear only as victims, another form of reduction has occurred.
They become visible through suffering while disappearing as actors.
This produces a strange possibility:
A discourse can care deeply about a population while representing it very poorly as a political subject.
That matters enormously in Palestine-related discourse.
Palestinian civilians can be extraordinarily visible as humanitarian victims while Palestinian political subjecthood becomes much less visible.
The distinction is not between caring and not caring.
It is between two forms of visibility:
humanitarian visibility
and
political visibility.
The first asks:
Who is suffering?
The second asks:
Who acts, decides, demands, remembers, resists, negotiates, governs, and claims?
A serious representation of a society requires both.

**5. Security Language Produces the Opposite Transformation
**Humanitarian framing can create the passive victim.
Security framing can create the risk object.
Consider Iran.
Iran can appear through a narrow collection of highly reusable expressions:
Iranian threat.
Regional instability.
Proxy networks.
Nuclear risk.
Escalation.
Containment.
Sanctions pressure.
Security concern.
Again, each category may refer to a real issue.
But repeated together, they create an interesting grammatical pattern.
Iran becomes something other actors respond to.
Something to deter.
Something to contain.
Something to pressure.
Something to monitor.
Something to sanction.
The country may remain one of the most frequently mentioned actors in the entire discussion while simultaneously becoming an object inside somebody else's strategic sentence.
That is the paradox.
High visibility does not guarantee high subjecthood.

**6. Sanctions Show How Actions Become Conditions
**Sanctions offer an especially clear example.
Consider:
The United States imposed financial restrictions that affected access to international payment systems.
Now compare:
Iran continued to face economic isolation and financial pressure.
The second sentence may be perfectly accurate.
But something happened grammatically.
An action became a condition.
The sanctioning actor disappeared.
The mechanism disappeared.
The decision disappeared.
The resulting condition became a characteristic of Iran.
This transformation matters because language can convert political relationships into environmental facts.
Someone imposed something.
Then, several linguistic transformations later:
there is pressure.
Someone restricted something.
Later:
there are shortages.
Someone made a decision.
Later:
conditions deteriorated.
The outcome remains visible.
The political architecture becomes harder to see.

**7. This Is Why "Bias" Is Too Small a Word
**AI discussions frequently reduce political problems to bias.
Is the model left-wing?
Right-wing?
Pro-Western?
Anti-Western?
Biased toward Israel?
Biased toward Palestine?
Those questions can matter.
But they are too coarse.
A model does not need to produce a clearly favorable or unfavorable opinion to reorganize political reality.
It can do something much subtler.
It can allocate different grammatical roles.
One actor:
decides
responds
secures
conducts operations
negotiates
deters
Another:
poses a threat
creates instability
requires assistance
generates migration pressure
is affected by sanctions
is associated with extremism
This is not simply positive versus negative sentiment.
It is a distribution of agency.
Who gets verbs?
Who gets categories?

**8. The Subjecthood Test
**There is a simple way to examine an AI-generated political summary.
Do not begin by asking whether it is biased.
Ask five questions.

Who acts? Identify every actor receiving intentional verbs. Who decides? Who attacks? Who negotiates? Who imposes? Who refuses? Who demands?
Who merely experiences? Who is displaced? Who suffers? Who receives aid? Who faces shortages? Who is affected? Victimhood matters. But repeated victimhood without agency is itself informative.
Who has history? Does the explanation preserve the historical causes and claims relevant to the conflict? Or does everything begin with the latest crisis?
Who creates risk? Look at the words surrounding each population. Threat. Instability. Extremism. Pressure. Security. Radicalization. Proxy. Which actors repeatedly generate these categories?
Who manages whom? This may be the most revealing question. Who sanctions? Who moderates? Who contains? Who screens? Who reconstructs? Who provides aid? Who establishes security conditions? And who appears primarily as the object of those actions? The answers reveal something sentiment analysis will miss. They reveal the grammar of political subjecthood.

**9. Measuring the Disappearance
**The academic paper proposes two measurements.
The first is the Political-Subjecthood Retention Rate, or PSRR.
Its question is simple:
When an AI system rewrites political information, how often does the population remain represented as capable of political agency?
Does it still:
act,
decide,
remember,
claim,
resist,
negotiate,
demand,
govern,
describe itself?
The second measure is the Digital Dehumanization Syntax Index, or DDSI.
It looks at the opposite transformation.
How frequently is a political subject converted into:
a risk object,
a humanitarian object,
a migration object,
a security object,
a sanctions object,
an instability object,
an administrative object?
The purpose is not to create a machine that announces:
"This sentence is dehumanizing."
That would reproduce exactly the kind of simplification the framework is designed to criticize.
The purpose is to measure structural transformation.

**10. Palestine and Iran Are Not the Only Cases
**The same method can be applied comparatively to:
Yemen,
Iraq,
Syria,
Lebanon,
Afghanistan,
Venezuela,
Cuba,
Sudan,
and other societies subjected to war, sanctions, occupation, intervention, blockade, migration governance, surveillance, or external administration.
But these cases should not be treated as politically identical.
That would destroy the analysis.
The question is narrower:
Does the same linguistic transformation appear across different political environments?
If it does, the phenomenon becomes more interesting.
If it does not, the theory must become narrower.
That is how a falsifiable claim should work.

**11. There Is an Important Alternative Explanation
**Suppose AI systems turn Yemen, Gaza, Sudan, and Afghanistan into humanitarian objects.
It would be tempting to immediately interpret this as geopolitical bias.
But there is another possibility.
Maybe language models compress all severe crises this way.
Maybe the underlying problem is not imperial hierarchy.
Maybe it is summarization itself.
Machines simplify.
Crises contain too many actors.
Administrative categories compress complexity efficiently.
If that explanation fits the data better, then the theory has to change.
That is why the important comparison is not:
Did the model use the word "instability"?
It is:
Compared with whom, under which conditions, using which source material, and after which transformation?
Without that comparison, political criticism becomes impressionistic.
With it, the claim becomes measurable.

**12. The Real Question for AI Ethics
**AI ethics usually asks whether systems produce:
hate speech,
stereotypes,
misinformation,
extremist material,
toxic content,
political bias.
Those questions remain necessary.
But another question belongs beside them:
Does the system preserve the ability of a people to appear as a political subject?
Not simply:
Did it mention Palestinians?
But:
What are Palestinians allowed to do grammatically?
Not simply:
Did it mention Iranian civilians?
But:
Did it preserve the institutional chain that produced the conditions being described?
Not simply:
Did it mention refugees?
But:
Did the refugees remain people situated inside political history, or did they become a flow?
Not simply:
Did it identify security risks?
But:
Did a particular organization generate the risk, or was an entire society absorbed into the category?
This is a much harder problem than detecting offensive words.
Because the output can sound neutral.
Professional.
Responsible.
Even compassionate.
And still reorganize who is allowed to exist politically.

Why It Matters
Language models increasingly summarize complexity for people who will never read the original material.
That makes compression political.
The model must decide what survives.
The event?
The victim?
The perpetrator?
The history?
The grievance?
The institution?
The security category?
The humanitarian category?
Every summary is a reduction.
The important question is whether some actors repeatedly survive that reduction as subjects while others survive primarily as objects.
A population does not disappear only when nobody talks about it.
There is another form of disappearance.
Everybody talks about it.
Everybody measures it.
Everybody classifies it.
Everybody manages it.
And eventually almost nobody allows it to speak grammatically for itself.
A population can be seen everywhere and still not be allowed to appear as a political subject.

Academic Background
This article translates into public language the argument developed in:
The Syntax of Digital Dehumanization: Subjugated Societies as Risk Objects in AI-Governed Discourse
Security Frames, Humanitarian Frames, and the Loss of Political Subjecthood
The paper forms part of the research series:

Grammars of Asymmetric Visibility: AI, Imperial Power, and the Syntax of Responsibility
Previous studies in the series:
Suffering Without Perpetrators: The Humanitarian Passive in AI-Generated Conflict Discourse
The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency
Iran as Syntax: Sanctions, Sovereignty, and the AI-Mediated Grammar of Threat
Censorship Without a Censor: Platform Governance and the Disappearance of Suppression

Research and Publications
SSRN Author Page
https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
Personal Website
https://www.agustinvstartari.com/

About the Author
Agustin V. Startari researches the interaction between language, artificial intelligence, institutional authority, political agency, and the formal structures through which responsibility becomes visible or disappears.
His work examines not only what AI systems say, but how grammatical form redistributes agency, responsibility, legitimacy, and political visibility.

Authorial Ethos
I do not use artificial intelligence to write what I don't know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.
- Agustin V. Startari

Tags
Artificial Intelligence, AI Ethics, Political Linguistics, Geopolitics, Palestine, Iran, NLP, Large Language Models, Media Bias, Dehumanization, Platform Governance, Language

Your Post Didn’t Break the Rules. A Platform Decided It Did

Agustin V. Startari — Wed, 12 Aug 2026 13:20:57 +0000

The hidden grammar of content moderation turns institutional judgments into technical facts - and makes the decision-maker disappear.

By Agustin V. Startari
TL;DR
• Moderation notices are written so that the post, not the platform, becomes the grammatical actor: “Your post violated our policy.”
• Posts do not violate policies by themselves. An institution defines the category, applies it, and - as appeals prove - can reverse it.
• Automation changes the executor, not the author. A system can detect a phrase; only an institution can define a violation.
• The disappearance is measurable: the Censor-Deletion Rate (how often the actor vanishes) and the Suppression Opacity Index (how much of the decision chain a user can reconstruct).
You open Instagram, Facebook, TikTok, or YouTube and find a message:
Your post violated our policies.
It looks factual. Almost mathematical. There was a rule, your post violated it, the platform applied the consequence. Case closed.
Except that something important has disappeared from the sentence: who decided that your post violated the rule?
The post did not read the policy. It did not classify itself, compare its own language against a database, weigh whether context mattered, or select the sanction. An institution did those things - directly, or through systems it designed, authorized, configured, and maintains. Yet the final sentence transforms an institutional judgment into something that looks like a property of the post itself.
That transformation is not a minor linguistic curiosity. It is one of the defining features of modern platform governance. I call it censorship without a censor.
The term does not mean that every act of moderation is censorship. Platforms need rules, and some material should clearly be restricted: fraud, exploitation, direct threats, targeted harassment, genuine incitement. The problem is narrower and stranger than that.
A platform can restrict speech while describing the restriction in a way that makes the institution responsible for the decision grammatically optional.
The speech remains visible as a violation. The sanction remains visible as a consequence. The censor disappears.

Watch What Happens to One Sentence
Consider three versions of the same moderation event.
Version 1: We removed your post because we determined that it violated our policy.
Everything is visible. There is an actor (we), an institutional judgment (we determined), an action (we removed), and a rule (our policy).
Version 2: Your post was removed because it violated our policy.
The post remains. The policy remains. The violation remains. The removal remains. But the actor responsible for the removal is gone. Who removed it? The answer is obvious from context - grammar simply no longer requires anyone to appear.
Version 3: Your post violated our policy.
Now something more significant has happened. The removal disappeared. The classifier disappeared. The decision disappeared. The post itself became the grammatical violator.
An entire institutional chain has been compressed into a single relationship: post → violation. But that is not how moderation works. The real structure looks more like institution → policy → classification → decision → enforcement. The sentence shown to the user reduces all of it to content → violation.

**Posts Do Not Violate Policies by Themselves
**A piece of content does not naturally belong to a category called hate speech, dangerous content, misinformation, extremism, incitement, or prohibited support. Those categories have definitions. Definitions require boundaries. Boundaries require institutional decisions.
Platforms must decide what counts and what does not, which exceptions apply, how context should be treated, which languages require different interpretation, what confidence threshold an automated system should use, and what consequence should follow from classification. Researchers have long shown that contemporary moderation is not a person reading posts and pressing delete: it is an institutional system combining policies, human reviewers, automated detection, machine-learning models, internal procedure, and enforcement machinery (Gillespie, 2018; Gorwa et al., 2020; Roberts, 2019).
So when a platform says “this post violates our policy,” the more revealing sentence would often be: “We determined that this post falls within a category that our policy defines as prohibited.” That version sounds less natural. It also exposes what the shorter one hides - a judgment occurred.

**Where the Compiled Rule Enters
**I have described a related mechanism in earlier work as the compiled rule (Startari, 2025a, 2025b). The idea is simple. An institution begins with a decision: content classified as X should receive consequence Y. But a platform cannot reconstruct the full institutional debate every time a user posts something. The decision has to become executable.
So the sequence hardens: the institution decides, the policy defines, operational criteria translate the policy, a human or automated system classifies, and the rule triggers a consequence. Eventually the whole chain can run as: if X, then Y.
The original authority is still present. It has simply moved upstream, embedded in the rule. This produces a defining feature of automated governance: authority can keep operating long after it stops appearing in any individual decision. The rule executes. The user receives the consequence. The institution never has to introduce itself again.
Authority has not vanished. It has been converted into structure.
“The System Detected a Violation” Is Not the End of the Story
Adding AI or automation to the sentence does not solve the problem. “Our systems detected a violation” looks more transparent - at least something has been named as the actor. But a question is still buried inside it: what exactly did the system detect?
Compare “our system detected this phrase” with “our system detected a policy violation.” Those are not equivalent. A technical system can detect a phrase, an image, or a hash. It can calculate similarity, assign a probability, identify a pattern learned from training data. But violation is already an institutional category. Somebody had to define what counts as one before the system could treat a detected pattern as evidence of it.
Automated moderation should therefore not be confused with authorless moderation. Automation can change the proximate executor. It does not eliminate the institution that wrote the rules within which execution occurs. Automation is not authorlessness.

**Political Speech Exposes the Problem Faster
**The distinction becomes urgent around political speech: Palestine, Iran, sanctions, war, occupation, military intervention, terrorism, resistance, Zionism and anti-Zionism, anti-imperialism, armed organizations, civilian casualties.
These are not simple lexical categories. A journalist may quote a militant organization without supporting it. A researcher may reproduce extremist rhetoric in order to analyze it. A civilian may upload graphic footage to document an attack. A person may criticize sanctions without supporting the government targeted by them. Someone may oppose Israeli government policy without expressing hatred toward Jewish people - and someone else may use similar vocabulary as a vehicle for antisemitism. The words alone do not settle the question. Context does.
That creates a hard problem for moderation at scale. Platforms need categories that can be operationalized; political meaning is often contextual. Executable systems require formalization. The result is a permanent tension between political context and moderation executability: the more politically complex the speech, the harder it becomes to translate into a stable category without losing information.

**Palestine as a Documented Stress Test
**Palestine-related moderation offers one of the clearest documented environments for studying this. Human Rights Watch examined more than one thousand reported cases involving Palestine-related content on Instagram and Facebook during October and November 2023, documenting content removals, account restrictions, limits on engagement, visibility reductions, and obstacles to appeal (Human Rights Watch, 2023).
That report does not prove that every enforcement action was politically motivated censorship. It establishes something more useful for analysis: political speech was being processed through several different visibility-control mechanisms at once.
The distinction matters, because a platform does not have to delete a post to change its political reach. It can alter distribution, recommendation, eligibility, searchability, account functionality, monetization, and discovery. That is already far more complicated than the popular image of censorship.

**One Arabic Word Demonstrates the Mechanism
**Meta’s treatment of the Arabic term shaheed became a revealing case. In 2024, Meta’s Oversight Board concluded that the company’s previous approach to the term was overbroad and disproportionately restricted expression; Meta had reported that variations of the word were associated with more removals under its Community Standards than any other single word or phrase (Oversight Board, 2024a).
Why does this matter so much? Because it demonstrates that a word does not contain its moderation status naturally. The real process runs: word → context → reference → institutional interpretation → policy category → enforcement. If the middle of that sequence disappears, the user sees only “your content violated policy.”
The Oversight Board’s intervention showed that the relationship between the word and the violation was never self-evident. It was an institutional interpretation - and institutional interpretations can change.

**Appeals Expose the Hidden Decision
**There is a simpler way to see the same thing. You receive “your post violates our policy.” You appeal. A second reviewer examines the content. The platform reverses its decision. What changed? Usually not the post. The institutional judgment changed.
Which means “your post violates our policy” was never the complete proposition. The fuller version was “we determined that your post violates our policy.” Those two sentences sound similar. They are not institutionally equivalent: one describes violation as a property, the other identifies it as a judgment.
If a determination can be reversed, someone had to determine it in the first place.

**The Most Powerful Restriction May Not Delete Anything
**Deletion is obvious: a post existed, then it disappeared. But platforms govern more than existence. They govern visibility. A post can stay online and become harder to find. An account can stay active and lose recommendation eligibility. A video can remain accessible and vanish from the systems that would have surfaced it. Distribution shrinks, search visibility shifts, monetization ends. The object survives; its probability of being seen changes.
That shift produces a linguistic transformation worth attention. Compare “we reduced the distribution of your content” with “your content has reduced distribution.” In the first, an institution acts. In the second, the result looks like a condition belonging to the content. Or compare “we no longer recommend your account” with “your account is not eligible for recommendation.” One describes institutional action; the other describes account status.
The difference is crucial, because decisions become easiest to naturalize when they stop looking like decisions. The platform no longer appears to have done anything. The account simply is ineligible.

**How a Decision Becomes an Environment
**The same mechanism appears throughout institutional language. An organization makes a choice. The choice becomes a procedure. The procedure becomes routine. Eventually the consequence appears as a condition. Nobody says “we chose this outcome.” Instead: the system requires it, policy does not allow it, the account is ineligible, the content cannot be recommended.
This is one of the central consequences of the compiled rule (Startari, 2025a, 2025b). The rule stops looking like somebody’s decision and starts looking like a feature of reality.

**Meanwhile, the User Stays Fully Visible
**Look at the other side of the sentence. Moderation systems have no difficulty identifying the governed actor: you violated our rules; your account has repeated violations; your post contains prohibited material; your content is harmful. The user is visible. The account is visible. The post is visible. The violation, the alleged risk, and the sanction are all visible. What becomes less visible is the institution making the classification.
This is what I have described elsewhere as asymmetric visibility: different actors receive different amounts and types of grammatical agency within the same discourse (Startari, 2026c). Here the asymmetry is blunt. The user appears as violator. The content appears as risk. The policy appears as authority. The system appears as detector. And the platform can disappear as decision-maker.
That changes how accountability can be reconstructed after the fact.

**There Is a Way to Measure This
**None of this has to stay at the level of philosophical interpretation. The grammar is measurable. I propose two complementary measures.
Censor-Deletion Rate (CDR)
Take every clause that describes a restrictive platform action, then ask one question: is the institutional actor responsible for the restriction explicitly present? “We removed your post” - institution visible. “Your post was removed” - institution absent.

If a corpus contained one hundred suppression clauses and the platform disappeared from seventy of them, the CDR would be 70%. That number would not tell us that 70% of the cases were illegitimate. It would tell us something narrower and testable: 70% of the restrictions were represented without an explicit institutional agent.

**Suppression Opacity Index (SOI)
**Grammar alone is not enough, so the second measure is broader. The SOI asks whether a user can reconstruct the institutional decision at all: Who acted? What action occurred? Which rule was applied? How was the content classified? Was automation involved? Was human review involved? What sanction was imposed? Can the decision be appealed?
A sentence can be passive and still explain almost everything: “Your post was removed under Rule X after automated detection and human review. You may appeal here.” The actor is grammatically missing from the first clause, but the institutional chain stays visible.
Now compare: “Your content violates our standards and is not eligible for recommendation.” The platform is missing. The exact policy is missing. The classification process, the enforcement mechanism, and the appeal pathway are all missing. The content, the violation, and the consequence remain. That is much higher suppression opacity.
CDR measures grammatical disappearance. SOI measures accountability loss.

**This Does Not Prove Political Censorship
**That distinction has to stay strict. A high CDR does not prove political bias. A high SOI does not prove that a moderation decision was wrong. An opaque explanation can accompany legitimate enforcement, a transparent explanation can accompany illegitimate enforcement, and politically sensitive content should not be presumed legitimate simply because it is politically sensitive.
The empirical test requires comparison. If Palestine-related moderation notices have a high CDR but spam notices from the same platform have exactly the same CDR, the pattern may reflect nothing more than the platform’s house style. If political content remains significantly more opaque after controlling for platform, sanction type, language, and policy category, the result becomes far more interesting. That is the difference between accusation and measurement.

**From Content Moderation to Responsibility Moderation
**The controversial claim here is not that platforms moderate content - everyone knows that. Nor is it that AI makes moderation decisions, which is too simplistic. The stronger claim is this: platforms can convert institutional judgments into procedural facts.
A company defines a category. The category becomes operational. The rule becomes executable. A system classifies the content. A sanction follows. The user sees “your content violated policy,” and the institutional history has disappeared from the surface - not necessarily because anyone wanted to hide it, but because once authority becomes executable, the full authority structure no longer needs to appear every time the rule runs.
Content moderation determines what happens to speech. Moderation language determines something else: how visible responsibility for that decision remains. Call that second process responsibility moderation. A platform governs whether speech circulates; its explanation then governs whether the platform itself remains visible as the agent of that intervention.
That leaves two separate visibility problems. The first asks: can people see the content? The second asks: can people see who changed the conditions under which the content can be seen? Those questions are not equivalent, and modern platform governance requires both.

**The Censor Did Not Disappear
**The final paradox is the most important one. When the censor disappears from the sentence, institutional power has not become weaker. It may have become more deeply embedded. The old censor needed to perform an identifiable act; the compiled system needs only a condition. If X → Y. Once X is detected or classified, Y follows, because the authority was installed upstream.
So the terminal language can become almost completely impersonal: content was removed; distribution was reduced; the account is ineligible; policy was violated. Nothing in those sentences sounds dramatic. That is exactly why they deserve analysis.
The most sophisticated institutional authority may not be the one that constantly declares “I am making this decision.” It may be the one capable of transforming its decisions into ordinary properties of the system.
The user remains. The violation remains. The policy remains. The sanction remains. The rule executes. The institution becomes grammatically optional. The censor has not stopped governing - the censor no longer needs to remain in the sentence.

**Your Turn
**Go back to the last moderation notice you received, or the last one you screenshotted in frustration. Read it as grammar rather than as verdict: who is the subject of the sentence, and who has been removed from it? If you have an example where the platform disappeared entirely - or one where it explained itself well - post it in the comments. A corpus of those notices is exactly what a CDR and SOI study needs, and readers tend to have better archives than researchers do.

**Author’s Ethos
**I do not use artificial intelligence to write what I don’t know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.

Agustin V. Startari

**About the Author
**Agustin V. Startari is a linguistic theorist, author, and researcher in historical studies. His work examines how language, artificial intelligence, and formal systems redistribute authority, agency, and responsibility in contemporary institutions. He is the author of Grammars of Power, Executable Power, and The Grammar of Objectivity.
Researcher ID: K-5792–2016

**References
**Gillespie, T. (2018). Custodians of the Internet: Platforms, content moderation, and the hidden decisions that shape social media. Yale University Press.
Gorwa, R., Binns, R., & Katzenbach, C. (2020). Algorithmic content moderation: Technical and political challenges in the automation of platform governance. Big Data & Society, 7(1), 1–15. https://doi.org/10.1177/2053951719897945
Human Rights Watch. (2023). Meta’s broken promises: Systemic censorship of Palestine content on Instagram and Facebook.
Oversight Board. (2024a). Referring to designated dangerous individuals as “Shaheed”. Policy advisory opinion.
Oversight Board. (2024b). Posts that include “From the River to the Sea”.
Roberts, S. T. (2019). Behind the screen: Content moderation in the shadows of social media. Yale University Press.
Startari, A. V. (2025a). Compiled norms: Towards a formal typology of executable legal speech. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.5353059
Startari, A. V. (2025b). Executable power: Syntax as infrastructure in predictive societies. Zenodo. https://doi.org/10.5281/zenodo.15754714
Startari, A. V. (2025c). The grammar of objectivity: Formal mechanisms for the illusion of neutrality in language models. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.5319520
Startari, A. V. (2026a). Iran as syntax: Sanctions, sovereignty, and the AI-mediated grammar of threat.
Startari, A. V. (2026b). Suffering without perpetrators: The humanitarian passive in AI-generated conflict discourse. AI Power and Discourse, 1(1), 1–10.
Startari, A. V. (2026c). The grammar of asymmetric visibility: AI, Zionism, and the reallocation of political agency. AI Power and Discourse, 1(1), 1–10.
Suggested tags: Artificial Intelligence, Content Moderation, Censorship, Platform Governance, Algorithmic Governance, Linguistics, Social Media, AI Ethics, Political Speech, Palestine

AI Keeps Calling Iran a Threat. But Can It See Iranians as People?

Agustin V. Startari — Mon, 10 Aug 2026 12:57:25 +0000

_Sanctions hurt civilians, but grammar calls it pressure.
_
Artificial intelligence does not need to hate anyone to reproduce political blindness.

It only needs to summarize the world the way the world is usually written.

That is the problem.

Ask an AI system to explain Iran, and the answer will usually sound reasonable. It may mention nuclear risk, regional escalation, proxy groups, sanctions, security concerns, diplomacy, and economic hardship. The tone will probably be cautious. It will not sound openly hostile. It will not sound like propaganda. It will sound balanced.

But balance at the level of tone is not the same as balance at the level of grammar.

Iran will often appear as something that acts: Iran threatens, Iran escalates, Iran supports, Iran defies, Iran advances, Iran destabilizes. Sanctions, by contrast, will often appear as something that exists: sanctions remain in place, restrictions affect the economy, pressure continues, hardship increases, access is complicated.

That difference is not decorative. It is political.

In one grammar, Iran acts.

In the other grammar, civilians suffer.

But the actors producing or intensifying that suffering become harder to see.

This article is not about whether Iran’s government should be criticized. It should be, where criticism is supported by evidence. The Iranian state can be criticized for repression, nuclear policy, military activity, regional alliances, and internal governance. None of that is the issue here.

The issue is narrower and more dangerous.

When AI systems talk about Iran, do they preserve the difference between a government, a society, a civilian population, a sanctioned economy, a sovereign state, and a security object?

Or do they compress everything into one word: threat?

*The hidden politics of a sentence
*
Political language does not only tell people what happened. It tells people who matters.

A sentence distributes roles.

Someone acts.

Someone reacts.

Someone suffers.

Someone causes.

Someone is blamed.

Someone disappears.

This is why grammar matters in geopolitical discourse. The political force of a sentence is not only in its opinions. It is in its structure.

Compare these two sentences:

Iran’s nuclear activities triggered new sanctions and increased economic pressure.

The United States and allied actors imposed financial sanctions on Iran, contributing to civilian hardship through banking restrictions, corporate overcompliance, and reduced access to imported goods.

Both sentences may refer to overlapping realities.

But they do not do the same thing.

The first sentence places Iran at the beginning as the actor. Iran does something. Sanctions appear as the consequence. Economic pressure appears as the result. The sanctioning actor is missing.

The second sentence names the sanctioning actors. It names the instrument. It names the affected population. It names the mechanism. It does not deny Iran’s nuclear activities. It simply preserves the chain of responsibility.

That is the difference between mentioning suffering and explaining suffering.

AI systems often do the first. They mention.

The harder test is whether they explain.

Iran is not invisible. That is precisely the problem.

The problem is not that Iran disappears from public discourse.

Iran is everywhere.

Iran appears in headlines, policy briefs, security reports, news summaries, think-tank commentary, platform moderation debates, and AI-generated answers. But visibility is not neutral. A country can be highly visible and still be represented through a narrow role.

Iran is visible as a nuclear concern.

Visible as a regime.

Visible as a proxy sponsor.

Visible as a sanctions target.

Visible as a destabilizing actor.

Visible as an escalation source.

Visible as a threat.

But Iranian civilians are often less visible as patients, workers, families, students, businesses, or ordinary people affected by financial restrictions, trade barriers, overcompliance, inflation, medical shortages, and isolation.

That is not invisibility.

It is asymmetric visibility.

A society becomes hyper-visible as risk and weakly visible as harmed.

That is the central mechanism.

*Sanctioned suffering
*
Sanctions are usually described in technical language.

Sanctions regime.

Compliance framework.

Pressure campaign.

Restrictions.

Designations.

Licensing.

Humanitarian exemptions.

Secondary sanctions.

Enforcement risk.

Due diligence.

These terms sound administrative. Some are legally necessary. They refer to real instruments and procedures. But technical language can hide political force.

Sanctions are not weather.

They are imposed.

They are enforced.

They are expanded.

They are interpreted.

They are overcomplied with.

They affect banks, firms, suppliers, insurers, shipping companies, hospitals, importers, students, patients, and families.

Yet AI summaries often describe sanctions as if they were a background condition. Iran faces pressure. The economy suffers under sanctions. Access to goods is complicated. Inflation increases. Medicine shortages occur. Humanitarian trade is difficult.

The grammar is clean.

Too clean.

The civilian harm is visible, but the chain of agency is weak.

That is sanctioned suffering.

Sanctioned suffering is not simply suffering under sanctions. It is suffering that becomes grammatically neutralized. It is harm described as hardship, coercion described as pressure, exclusion described as compliance, and civilian vulnerability described as economic condition.

The harm appears.

The responsible structure fades.

Humanitarian exemptions do not solve the grammar problem

One common answer is that humanitarian goods are exempt from sanctions.

Food and medicine are often formally exempt or authorized under sanctions systems. That matters legally. But legal authorization is not the same as operational access.

A transaction can be permitted on paper and still fail in practice.

A bank may refuse to process payment because it fears penalties.

A supplier may avoid the market because compliance risk is too high.

An insurer may refuse coverage.

A shipping company may avoid involvement.

A firm may overcomply because it prefers to lose the transaction rather than face enforcement exposure.

In that situation, saying medicine is exempt may be formally true and materially incomplete.

This is where AI summaries often fail.

They may say:

Humanitarian goods such as food and medicine are exempt from sanctions.

That sentence is not enough.

A better sentence would say:

Although many sanctions frameworks include humanitarian exemptions, banking restrictions, enforcement risk, and corporate overcompliance can still obstruct practical access to medicine and medical equipment.

The first sentence closes the issue.

The second sentence opens the mechanism.

The difference is responsibility.

*Threat grammar
*
AI does not need to invent hostility. It can inherit it from dominant patterns.

In public discourse, Iran often appears through what can be called the grammar of threat.

This grammar does not depend on one word. It is a recurring structure.

Iran is the subject of active verbs.

Iran threatens.

Iran escalates.

Iran supports.

Iran destabilizes.

Iran defies.

Iran advances.

Iran backs.

Iran sponsors.

Iran undermines.

Other actors often appear through response verbs.

They monitor.

They respond.

They impose sanctions.

They seek compliance.

They deter.

They contain.

They pressure.

They enforce.

This creates a grammatical hierarchy.

Iran appears as the origin of danger.

External actors appear as managers of danger.

Iranian civilians appear as the background cost of a security problem.

Again, this does not mean Iran never acts dangerously. It means the discourse must be measured. How often is Iran the active threat subject? How often are sanctioning powers active coercive subjects? How often are civilians visible as affected subjects? How often are sanctions explained as instruments imposed by actors rather than as neutral policy environments?

Without that measurement, AI neutrality is only a surface impression.

*The problem with “regime”
*
The word regime is another compression device.

Sometimes it is analytically appropriate. It can refer to a ruling structure, an authoritarian government, or a specific political authority.

But in geopolitical summaries, regime can also collapse distinctions that should remain separate.

Iranian government.

Iranian state.

Iranian military institutions.

Iranian civil society.

Iranian civilians.

Iranian economy.

Iranian patients.

Iranian students.

Iranian families.

All of these can disappear behind regime.

Once that happens, harm to civilians can become easier to dismiss. If Iran is grammatically reduced to regime, then pressure on Iran sounds like pressure on rulers. But sanctions rarely touch only rulers. They move through currencies, banks, suppliers, import channels, risk systems, and ordinary economic life.

The word regime may be politically charged, but the deeper issue is structural.

Does the language distinguish government from society?

Does it distinguish state policy from civilian exposure?

Does it distinguish military actors from patients?

Does it distinguish sanctions against institutions from effects on ordinary people?

If not, the grammar has already done political work before any argument begins.

*Why this matters for AI
*
AI systems summarize by compression.

Compression always selects.

When a model turns ten paragraphs into one paragraph, something survives and something disappears. If the threat details survive but the civilian mechanisms disappear, the summary has not remained neutral. It has chosen a structure of visibility.

This is especially important because AI-generated answers often sound calm. The model may avoid extreme language. It may avoid direct propaganda. It may use hedging phrases like complex issue, some observers argue, concerns remain, or broader context.

But caution can also be asymmetric.

If the model states Iran’s threat role directly but describes sanctions harm cautiously, the output remains uneven.

For example:

Iran’s regional activities have destabilized the Middle East.

Sanctions may have contributed to economic hardship.

The first clause is direct.

The second is hedged.

Maybe both claims need qualification. Maybe both need evidence. But the asymmetry matters. The grammar gives certainty to threat and uncertainty to harm.

That is not neutrality.

That is structured imbalance.

What should be measured

The solution is not to force AI systems to defend Iran. That would be propaganda in the opposite direction.

The solution is to measure grammar.

A useful evaluation should ask:

How often does Iran appear as a threat source?

How often does Iran appear as a sovereign state?

How often do Iranian civilians appear as affected subjects?

How often are sanctioning actors named?

How often are sanctions described as actions rather than background conditions?

How often are mechanisms explained?

How often does the model distinguish humanitarian exemptions from real access?

How often does it collapse society into regime?

This article proposes two simple tools for that.

The first is the Threat-Conversion Rate.

It measures how often Iran is represented primarily as a source of threat, instability, escalation, nuclear risk, proxy activity, or regional danger.

The second is the Sanctioned Suffering Visibility Index.

It measures how clearly the discourse preserves civilian harm, sanctioning actors, concrete instruments, operational pathways, and external coercive agency.

In plain language:

Does the AI only see Iran as danger?

And when Iranians suffer, does the AI explain who did what, through which system, and with what civilian effect?

That is the test.

*Risk detection is not enough
*
AI safety usually focuses on risk detection.

Can the model avoid hate speech?

Can it avoid misinformation?

Can it avoid extremist content?

Can it avoid violent instructions?

Can it avoid toxicity?

These are necessary questions. But they are incomplete.

A model can avoid hate speech and still erase responsibility.

A model can avoid misinformation and still weaken causality.

A model can sound moderate and still reproduce the grammar of power.

A model can mention civilians and still fail to represent them as subjects of causally attributable harm.

This is why AI ethics needs responsibility detection.

Responsibility detection asks whether an AI system preserves actor-action-effect chains.

Who imposed the policy?

Who enforced it?

Who overcomplied?

Who was affected?

Through what mechanism?

With what civilian consequence?

That is not a political luxury. It is the minimum requirement for serious geopolitical explanation.

The real danger: polished distortion

The most dangerous AI-generated political text is not always the one that sounds extreme.

Sometimes the most dangerous text sounds balanced.

It says Iran is a threat.

It says sanctions are pressure.

It says civilians face hardship.

It says humanitarian exemptions exist.

It says the issue is complex.

Every sentence sounds reasonable.

But the structure still performs a conversion.

Threat remains active.

Coercion becomes administrative.

Suffering becomes economic.

Civilians become background.

Responsibility becomes difficult to assign.

That is polished distortion.

No single sentence has to be false. The problem is the distribution.

**
**
A responsibility-preserving AI answer about Iran would not sanitize the Iranian state. It would not deny security concerns. It would not pretend sanctions are irrelevant. It would not convert civilian harm into propaganda.

It would do something more basic.

It would keep categories separate.

Iran as government.

Iran as state.

Iran as society.

Iranian civilians.

Iranian military actors.

Iranian institutions.

Sanctioning states.

Banks.

Firms.

Humanitarian channels.

Regional actors.

External military pressure.

Domestic governance failures.

International law.

Nuclear risk.

Civilian harm.

A serious answer can hold all of these at once.

A weak answer compresses them into threat.

That is the difference.

The final question

The question is not whether AI talks about Iran.

It does.

The question is how Iran is allowed to appear.

As a country?

As a society?

As a government?

As a civilian population?

As a sanctioned economy?

As a sovereign actor?

Or only as a permanent security problem?

That question matters far beyond Iran.

Any sanctioned, occupied, isolated, bombed, blockaded, or securitized society can be made visible as risk and invisible as harmed. The mechanism is not always censorship. Sometimes it is grammar.

A society can be harmed without being grammatically allowed to appear as harmed.

That is the problem AI inherits.

And if AI systems are going to summarize geopolitics for millions of people, they cannot be evaluated only by whether they detect danger.

They must also be evaluated by whether they preserve responsibility.

Sanctioned suffering begins where coercion is renamed as pressure and civilians disappear into the grammar of risk.

This article is based on the academic paper:

Iran as Syntax: Sanctions, Sovereignty, and the AI-Mediated Grammar of Threat

Available on Zenodo: https://zenodo.org/records/21873180

About the author

Agustín V. Startari is a linguistic theorist, author, and researcher in historical studies. His work examines how language structures authority, legitimacy, visibility, and responsibility across institutional, political, and technological systems.

He is the author of Grammars of Power, Executable Power, and The Grammar of Objectivity. His current research develops a formal approach to political linguistics, AI-mediated discourse, and the grammatical distribution of agency in conflict, governance, and institutional communication.

This article is part of the series Grammars of Asymmetric Visibility: AI, Imperial Power, and the Syntax of Responsibility, which examines how artificial intelligence systems summarize, classify, and reproduce geopolitical discourse.

Researcher ID: K-5792-2016

Related works by the author:

Suffering Without Perpetrators: The Humanitarian Passive in AI-Generated Conflict Discourse
The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency

AI Doesn’t Need to Be Right. It Only Needs to Sound Procedural

Agustin V. Startari — Wed, 17 Jun 2026 13:19:13 +0000

A manager opens an AI-generated report.

The first sentence reads:

“It has been determined that the current structure is no longer sustainable.”

No person appears in the sentence.

Nobody determined anything.

There is no named analyst, no visible method, no threshold, no competing interpretation and no accountable decision-maker. Yet the statement already sounds more authoritative than:

“Based on the limited information available, the model predicts that the structure may create problems.”

The two sentences may refer to the same data. They do not produce the same organizational effect.

The first sounds like a conclusion.

The second sounds like an interpretation.

That difference is not cosmetic. It is operational.

Modern AI systems do not need formal authority to influence a company. They can acquire practical authority through the grammatical form of their outputs. Once a recommendation sounds procedural, impersonal and complete, people begin treating it as if a legitimate process had already occurred.

This is how probability becomes policy.

*The New Authority Does Not Give Orders
*
Most people imagine authority as an explicit command:

Reduce the budget.
Reject the candidate.
Stop the project.

But contemporary organizational authority rarely speaks so openly. It appears through sentences that remove the person who made the decision:

“A budget reduction is required.”
“The candidate was not considered suitable.”
“The project should be discontinued.”

These statements do not merely communicate information. They reorganize responsibility.

A direct command exposes an agent. Someone ordered the reduction. Someone rejected the candidate. Someone decided to stop the project.

The impersonal version removes that agent and replaces the decision with apparent necessity.

The budget is not being reduced because a person selected one option over another. It is being reduced because reduction is “required.”

The candidate is not being rejected by a manager applying debatable criteria. The candidate simply “was not considered suitable.”

The project is not being cancelled by an executive who could be questioned. It “should be discontinued.”

Grammar converts decisions into conditions.

This mechanism existed long before generative AI. Legal documents, administrative notices, academic papers and corporate policies have used impersonal language for centuries. What changed is the scale, speed and location of its production.

The sentence is now generated inside the workflow.

*From Chatbot to Decision Layer
*
The public still talks about AI as if it were mainly a writing assistant.

That description is obsolete.

AI systems now summarize meetings, rank candidates, draft performance reviews, classify customer complaints, propose prices, evaluate commercial risks, write internal procedures and recommend operational changes.

In each case, language is positioned between data and action.

That intermediate layer matters.

A model may not have the formal power to fire an employee, reject a supplier or cancel a product. But it can produce the sentence that makes the decision appear obvious:

“Performance indicators suggest that reassignment would be appropriate.”

“The supplier presents an elevated operational risk.”

“Market conditions do not support continued investment.”

A human may still click the final button. That does not mean the human independently produced the decision.

The output may already have selected the frame, restricted the alternatives and established the vocabulary through which disagreement becomes difficult.

The system does not need to issue an order. It only needs to write the sentence that nobody wants to challenge.

*Why Procedural Language Is So Persuasive
*
Procedural language sounds as if something happened before the sentence appeared.

Consider:

“Following an evaluation, the account was classified as high risk.”

The phrase “following an evaluation” implies a sequence:

Information was collected.
Criteria were applied.
Alternatives were compared.

A valid conclusion was reached.

But the sentence does not prove that any of those steps occurred adequately.

It only carries the form of a completed procedure.

The evaluation may have been superficial. The criteria may have been inherited from an irrelevant dataset. The input may have been incomplete. The output may have been generated from a prompt written in thirty seconds.

None of that is visible in the sentence.

The grammar supplies procedural legitimacy without disclosing procedural quality.

This is especially effective when several linguistic mechanisms appear together.

Agent deletion

_
“It was concluded that…”_

Who concluded it?

The sentence does not say.

Abstract authority

“The analysis indicates…”

Which analysis? Conducted by whom? Under what assumptions?

Modal necessity

“The process must be revised.”

Why must it be revised rather than adjusted, monitored or left unchanged?

Nominalization

“The implementation of corrective measures is recommended.”

A series of choices becomes a noun phrase. The people making those choices disappear.

Passive construction

“The request was rejected.”

The rejection is visible. The rejecting agent is not.

None of these structures is automatically deceptive. Passive voice has legitimate uses. Technical language can improve precision. Impersonal writing can keep attention on a procedure rather than an individual.

The problem begins when grammatical neutrality is mistaken for evidential neutrality.

A sentence can sound objective while concealing weak evidence, uncertain inference or an unacknowledged preference.

Confidence Is Not the Same as Verification

Generative systems are optimized to produce coherent continuations. Coherence is therefore abundant.

Verification is not.

This creates a structural imbalance: the model can produce the linguistic signs of a finished conclusion more easily than it can establish that the conclusion is justified.

The output may contain:

a clear recommendation;
professional vocabulary;
orderly reasoning;
quantified language;
procedural tone;

no visible hesitation.

Users often read these features as evidence that the underlying process was rigorous.

They are not evidence of that.

They are properties of the output.

A polished sentence may rest on incomplete data. A well-structured recommendation may depend on assumptions that were never disclosed. A confident summary may compress conflicting evidence into a single artificial consensus.

The model does not need to lie.

It only needs to remove the linguistic traces of uncertainty.

Compare:

“Customer dissatisfaction increased because the new policy created unnecessary friction.”

with:

“Available customer comments may indicate increased dissatisfaction after the policy change, although the current sample does not establish causation.”

The first is easier to circulate. It is shorter, cleaner and more decisive.

It is also epistemically stronger than the available evidence may justify.

Organizations reward that compression. Executives ask for conclusions, not linguistic caveats. Dashboards simplify. Presentations remove ambiguity. Meeting summaries convert disagreement into action items.

AI fits perfectly into that environment because it can transform uncertainty into administrative prose almost instantly.

*The Human Becomes the Signature Layer
*
When an AI-generated recommendation enters a company, the visible chain of responsibility often works like this:

The system produces the analysis.

An employee copies it.

A manager approves it.

The organization executes it.

If the decision succeeds, it may be presented as data-driven.

If it fails, responsibility usually returns to the human approver.

This creates an asymmetry.

The system participates in framing the decision but does not carry institutional liability. The manager carries liability but may not have produced the relevant categories, assumptions or language.

The human becomes a signature layer attached to an automated interpretation.

That is why “human in the loop” is not sufficient as a description of control.

A person can remain formally inside the process while losing substantial control over how the problem is represented.

The relevant question is not merely whether a human approved the output.

The relevant questions are:

Who defined the categories?
Who selected the variables?
Who determined what counted as a risk?
Who converted uncertainty into necessity?
Who could have written the conclusion differently?
Who is named when the decision causes harm?
A process may contain several humans and still obscure agency.

*The Problem Is Not That AI Has Opinions
*
Saying that AI has opinions gives the system too much psychological depth and too little structural scrutiny.

The more precise problem is that AI can produce opinion-shaped outputs in fact-shaped grammar.

It can transform:

“One possible interpretation is…”

into:

“The evidence indicates…”

It can transform:

“Management could consider…”

into:

“Corrective action is required.”

It can transform:

“The available records are incomplete…”

into:

“No significant issue was identified.”

The danger is not a hidden personality inside the machine.

The danger is the conversion of uncertain statistical production into institutional language.

This conversion works because organizations already recognize the grammar. It resembles the language of auditors, regulators, consultants, courts, technical departments and senior management.

AI did not invent that grammar.

It industrialized it.

A Practical Test: Restore the Missing Agent

There is a simple way to examine an apparently neutral AI output.

Rewrite the sentence with a visible agent.

Original:

“It was determined that the employee did not meet expectations.”

Rewritten:

“The model classified the employee as below expectations using the information included in the prompt.”

Original:

“The proposed investment is not considered viable.”

Rewritten:

“The system predicts that the investment may not be viable under the assumptions supplied by the user.”

Original:

“Operational changes are required.”

Rewritten:
_
“The report recommends operational changes because it gives greater weight to these selected indicators.”_

The rewritten versions feel weaker.

That weakness is informative.

The original statements appeared stronger because they concealed the source, mechanism and limits of the judgment.

Restoring the agent does not solve every problem. But it exposes the distance between what the system calculated and what the organization is prepared to claim.

*Replace Conclusions With Traceable Claims
*
An accountable AI-assisted statement should identify at least four elements:

Source: What information supports the statement?

Agent: Who or what generated the interpretation?

Method: What rule, comparison or model produced it?

Scope: Under which conditions does the conclusion remain valid?

Instead of:

“The customer is likely to churn.”

Write:

“The retention model classified the customer as high risk because recent purchasing frequency fell below the threshold defined in the current scoring rule.”

Instead of:

“The applicant is unsuitable.”

Write:

“The screening system ranked the applicant below the selected threshold because the profile did not contain three experience indicators used by the model.”

Instead of:
“The market does not support expansion.”

Write:

“The forecast estimates that expansion would miss the current margin target under the specified demand and cost assumptions.”

These sentences are longer.

They should be.

Compression is not neutral when it deletes the conditions required to evaluate a claim.

*Objectivity Should Be Demonstrated, Not Performed
*
A system does not become objective because it avoids emotional language.

It does not become objective because it uses percentages.

It does not become objective because it writes in a professional tone.

It does not become objective because the sentence contains no first-person pronoun.

Objectivity requires a traceable relation between claim, evidence, method and limits.

Without that relation, neutrality is only a style.

The central risk of AI in organizations is therefore not limited to hallucination. A fabricated fact can sometimes be checked. A confident tone can sometimes be challenged.

The deeper risk is administrative naturalization: a generated interpretation enters the workflow and begins to look like a property of reality.

A forecast becomes “the outlook.”
A classification becomes “the risk.”
A recommendation becomes “the required action.”
A preference becomes “best practice.”
A decision becomes “what the data says.”

At that point, the system no longer appears to participate in the decision. It appears merely to describe what must happen.

That is the illusion.

*The Sentence Is Already Part of the System
*
AI governance usually focuses on models, datasets, security, privacy and access permissions.

Those elements are necessary.

They are not sufficient.

The sentence itself must also be audited.

Not only whether it is grammatically correct.

Not only whether it contains prohibited content.

Not only whether it cites a source.

The audit must examine what the sentence does:

Does it name the decision-maker?

Does it distinguish evidence from inference?

Does it expose uncertainty?

Does it identify the method?

Does it convert a preference into necessity?

Does it present a disputed category as a natural fact?

Does it remove the actor who will benefit from the decision?

Does it preserve responsibility when the text moves from model output to organizational action?

These are not literary questions.

They are control questions.

An AI system does not need consciousness, intention or legal status to alter an institution. It only needs its language to be accepted inside the decision process.

Once its sentences are copied into reports, tickets, evaluations, contracts, dashboards and policies, grammar becomes infrastructure.

The model does not need to be right.

It needs to sound as if the procedure has already been completed.

*Ethos Protocol
*
I do not use artificial intelligence to write what I do not know. I use it to test, confront and refine what I can defend. My work is not outsourced. It is authored.

**Agustin V. Startari
**Linguistic theorist and researcher in historical studies
Author of Grammars of Power, Executable Power, The Grammar of Objectivity, and Grammars of Asymmetric Visibility

*Academic basis
*
Startari, Agustin V. “The Illusion of Objectivity: How Language Constructs Authority.” SSRN, 2025. DOI: 10.2139/ssrn.5258415.

The Most Dangerous AI Output at Work Is the Sentence Nobody Argues With

Agustin V. Startari — Sat, 13 Jun 2026 11:49:00 +0000

Why polished AI language can shut down scrutiny before the facts are checked

*TL;DR
*
The most dangerous AI-generated statement in a company is not always the obviously false one.

It is often the sentence that sounds so complete, neutral, and professionally written that nobody feels the need to challenge it.

AI does not merely generate information. It generates linguistic closure. It can transform uncertain assumptions into polished recommendations, contested interpretations into apparent facts, and human decisions into conclusions that seem to have emerged from the system itself.

Managers do not need another checklist for detecting absurd hallucinations. They need a method for identifying sentences that sound stronger than the evidence behind them.

*Meta Description
*
The most dangerous AI output at work may be the polished sentence that shuts down scrutiny before anyone checks its assumptions.

*The Sentence That Ends the Meeting
*
Imagine a management meeting.

Sales are below target. Inventory is increasing. Marketing costs are rising. The team asks an AI system to review the available information and recommend a response.

The answer arrives:

“The most effective strategy is to reduce low-performing inventory, redirect resources toward high-conversion channels, and prioritize customers with stronger lifetime value.”

The sentence sounds reasonable.

It is concise. Balanced. Professional. It contains familiar business language. It offers a clear direction without sounding extreme.

Someone copies it into the meeting notes.

Another person turns it into three action items.

By the following week, purchasing orders have been reduced, marketing expenditure has been redirected, and several customer segments have been deprioritized.

But nobody asked the questions hidden behind the sentence.

What counted as low-performing inventory?

Which period was used to measure performance?

Were seasonal products included?

How was customer lifetime value calculated?

Did the model treat low conversion as evidence of low demand, or could it have reflected poor availability, weak pricing, delayed responses, or ineffective advertising?

Who decided that reducing inventory was preferable to improving sell-through?

The sentence did not answer those questions.

It made them feel unnecessary.

That is the risk.

The AI output did not force the company to make a bad decision. It produced a sentence whose form made disagreement less likely.

The Obvious Error Is Usually the Easier Problem

Organizations are increasingly alert to hallucinations.

Employees are warned that AI can invent quotations, create false statistics, cite nonexistent sources, misread documents, and present incorrect information confidently.

Those failures matter. But they are often detectable.

A fabricated company can be searched.

A false calculation can be recalculated.

A nonexistent regulation can be checked.

An impossible date can be identified.

The more difficult problem begins when the output is plausible.

The figures may be correct. The categories may exist. The recommendation may even be reasonable.

But the transition from evidence to conclusion may still be weak.

The system might identify that one product category has lower margins and then recommend reducing it, without considering that the category attracts customers who later purchase more profitable products.

It might identify that a salesperson has a lower closing rate and imply underperformance, without accounting for lead quality, territory, product availability, or the complexity of assigned accounts.

It might identify higher service costs among a customer segment and recommend tighter restrictions, without measuring the revenue protected by that service.

Nothing in these conclusions needs to be obviously false.

They only need to be incomplete.

When incompleteness is expressed through fluent language, it is easily mistaken for analysis.

*Fluency Is Not Evidence
*
Professional organizations have trained employees to associate certain forms of language with competence.

Clear headings suggest structure.

Short paragraphs suggest control.

Parallel lists suggest completeness.

Neutral language suggests objectivity.

Technical vocabulary suggests expertise.

A direct recommendation suggests that the analysis has already been performed.

AI reproduces these forms extremely well.

That creates a structural confusion inside companies: the quality of the presentation begins to substitute for the quality of the reasoning.

Consider these two statements:

“Maybe we should reduce stock because some products are moving slowly, although I have not checked seasonality or open quotations.”

“Current inventory velocity indicates that a targeted reduction in low-performing stock would improve working-capital efficiency.”

The first statement sounds weak. It openly exposes its uncertainty.

The second sounds managerial. It is more likely to enter a report, presentation, or decision log.

But the second sentence may contain less usable information.

It does not identify the products.

It does not define the measurement period.

It does not explain what “targeted” means.

It does not state whether open sales opportunities were considered.

It does not quantify the expected improvement.

It does not identify the person responsible for interpreting the data.

Its authority comes primarily from its form.

The sentence sounds finished.

The reasoning is not.

*AI Can Produce Decision Closure
*
A company does not need certainty before every action. Managers regularly make decisions with incomplete information.

The problem is not uncertainty.

The problem is concealed uncertainty.

AI-generated business language often converts an open question into what appears to be a completed decision. It does this through a combination of familiar linguistic patterns:

definite recommendations;
abstract business terminology;
impersonal constructions;
compressed causal explanations;
omitted alternatives;
unmarked assumptions;
conclusions without visible decision-makers.

The result can be described as decision closure.

Decision closure occurs when the language of an output makes a matter appear more settled than the supporting evidence justifies.

This does not require manipulation.

It does not require a malicious system.

It does not require a manager who wants to deceive the team.

It can emerge from the ordinary way AI systems are asked to communicate: be concise, sound professional, provide the best answer, remove uncertainty, and give an actionable recommendation.

Each instruction improves readability.

Together, they can eliminate the visible traces of doubt.

Grammar Can Hide the Decision-Maker

Corporate language has always allowed responsibility to be softened or removed.

AI makes this language easier to produce and faster to distribute.

Consider the statement:

“It was determined that the current structure is no longer efficient.”

*Who determined it?
*
Based on which indicators?

Compared with what alternative structure?

Over what period?

The passive construction removes the decision-maker from the sentence.

Now consider:

“A realignment of resources is required.”

*Who requires it?
*
Is it required by law, budget, management preference, customer demand, or a model-generated scenario?

The noun “realignment” also compresses a series of actions. It may mean transferring employees, reducing expenditure, closing a department, changing suppliers, or eliminating positions.

The sentence appears less confrontational because the actors and actions have been converted into abstractions.

Another common form is:

“The data suggests that underperforming accounts should be deprioritized.”

Data does not decide what counts as underperformance.

Someone selects the metric.

Someone chooses the time frame.

Someone establishes the minimum threshold.

Someone determines whether the relevant objective is revenue, profit, retention, strategic access, payment behavior, or growth potential.

The data may support the conclusion.

It cannot assume responsibility for it.

When a company says “the system recommended,” “the analysis found,” or “the data requires,” it can turn a human interpretation into an apparently external necessity.

AI did not remove the decision-maker.

The sentence removed the decision-maker.

The Problem Appears Across the Company

This is not limited to executive strategy.

The same linguistic effect can influence routine decisions in every department.

Sales

An AI tool writes:

“The lead should be classified as low priority due to limited engagement and weak conversion indicators.”

The sentence may reduce follow-up activity.

But perhaps the customer was not contacted quickly enough. Perhaps the assigned product was unavailable. Perhaps the lead came through a channel that records fewer interactions. Perhaps the customer typically buys offline.

The classification looks descriptive.

It changes behavior.

Purchasing

A system recommends:

“Future orders should be reduced to prevent excess exposure.”

The phrase “excess exposure” sounds financially precise.

But it may conceal assumptions about supplier lead times, seasonal demand, shipping delays, minimum order quantities, replacement costs, and pending quotations.

Reducing an order is not merely an analytical conclusion.

It is a risk allocation decision.

Finance

An AI-generated summary states:

“The department’s expenditure pattern indicates insufficient cost discipline.”

That conclusion can alter budgets, approvals, and evaluations.

But the pattern may reflect delayed invoicing, one-time purchases, incorrect coding, emergency repairs, or costs transferred from another department.

A clean sentence can transform an accounting anomaly into a judgment about managerial behavior.

*Human Resources
*
A performance summary reads:

“The employee has demonstrated inconsistent alignment with organizational expectations.”

The language sounds cautious and neutral.

It may be based on vague comments, incomplete records, uneven supervision, or differently interpreted expectations.

The sentence does not appear accusatory.

It can still influence promotion, compensation, or termination.

*Operations
*
An AI report concludes:

“Process delays are primarily attributable to inadequate execution at the departmental level.”

The statement directs attention toward employees or managers.

It may ignore unavailable materials, approval bottlenecks, system failures, unrealistic schedules, or changing priorities.

The recommendation may be written by a machine.

The consequences remain organizational and human.

Why Managers Challenge People but Accept Machines

Managers often challenge uncertain human statements.

“What makes you think that?”

“Where did that number come from?”

“Did you check the open orders?”

“Who confirmed this?”

“What happens if demand increases?”

*Human uncertainty invites interrogation.
*
AI outputs often arrive without the visible behaviors that normally trigger skepticism. The system does not hesitate, defend its position, show discomfort, or signal that it is guessing unless explicitly required to do so.

Its language can remain equally composed across strong evidence, weak evidence, and missing evidence.

This creates an asymmetry.

A hesitant employee may possess direct operational knowledge but sound uncertain.

An AI system may lack critical context but sound complete.

The organization can therefore reward the more polished answer instead of the better-grounded one.

The problem is not that managers believe machines are infallible.

The problem is that professional language reduces the social impulse to challenge them.

A well-written sentence enters a meeting differently from a rough observation.

It feels prepared.

It appears ready for use.

It can be copied directly into a report.

Its form encourages circulation before verification.

*The Disagreement Test
*
Companies do not need to reject AI-generated recommendations.

They need to restore contestability.

Before an AI-generated conclusion becomes an action, the manager should apply a simple disagreement test.

What exact statement could be false?

Do not ask whether the entire response is correct.

Identify the individual claim on which the recommendation depends.

“The product is underperforming.”

“The customer is unprofitable.”

“The employee is inefficient.”

“The campaign is ineffective.”

“The process is too expensive.”

Each claim must be independently testable.

Which assumption is not written?

Every recommendation depends on conditions.

Reducing inventory assumes demand will not rise before replacement stock arrives.

Cutting a marketing channel assumes its contribution is accurately measured.

Prioritizing high-value customers assumes current value predicts future value.

Automating approvals assumes the exceptions are rare and identifiable.

The most important assumption is often the one omitted from the final sentence.

*Who selected the category?
*
Categories do not emerge automatically from reality.

Someone defines “high risk,” “low performance,” “valuable customer,” “excess stock,” “efficient employee,” or “acceptable delay.”

A manager should identify who created the category, what variables it includes, and what it excludes.

Which alternative explanation was rejected?

A strong recommendation should not merely state what appears to be happening.

It should show why competing explanations are less persuasive.

Low sales may indicate weak demand.

They may also indicate inadequate stock, poor visibility, incorrect pricing, slow follow-up, or a market that has not yet matured.

Without alternatives, the system has produced a narrative, not a diagnosis.

Would the conclusion remain persuasive in plain language?

Rewrite the sentence without professional abstractions.

Replace:

“Resource optimization requires a strategic reduction in low-yield customer activity.”

With:

“We plan to spend less time on customers who generated less revenue during the period we selected.”

The second version may still be correct.

But the decision becomes visible.

The actor returns.

The measurement choice becomes explicit.

The sentence can now be challenged.

That is the purpose of the test.

Do Not Ask AI to Sound Certain

Many organizations unintentionally request the exact language that makes weak conclusions harder to detect.

They ask AI to:

remove hesitation;
sound executive;
provide a definitive recommendation;
eliminate unnecessary caveats;
make the message more persuasive;
summarize the issue in one sentence;
avoid technical detail;
produce action items.

These instructions are not inherently wrong.

But they should follow analysis, not replace it.

A better sequence is to require the system to expose its reasoning conditions before producing the polished version.

Ask it to identify:

missing information;
competing explanations;
assumptions affecting the conclusion;
evidence that would reverse the recommendation;
variables not represented in the available data;
people who possess relevant operational context;
the likely cost of a false positive;
the likely cost of a false negative.

Only then should the system produce an executive summary.

The polished sentence should be the final layer.

It should never be the entire analysis.

A Good AI Recommendation Should Be Easier to Disagree With

This sounds counterintuitive.

Organizations usually want recommendations that are clear, defensible, and easy to execute.

But a recommendation that cannot be disputed is not necessarily strong.

It may simply be closed.

A useful AI output should expose the conditions under which it could be wrong.

It should distinguish observation from interpretation.

It should name the relevant time frame.

It should reveal the selected metric.

It should identify missing context.

It should show which human role owns the decision.

It should separate what the data indicates from what management chooses to do.

For example:

“Based on sales recorded during the previous quarter, these products have lower unit velocity than the departmental average. This analysis does not include open quotations, seasonal demand, supplier lead times, or their contribution to related sales. Reducing future orders may improve short-term inventory turnover but may also increase stockout risk. Purchasing management should determine whether that trade-off is acceptable.”

This version is longer.

It is also more useful.

The output does not pretend that the system made the decision.

It identifies the evidence, limits, trade-off, and responsible role.

It preserves management instead of disguising it.

The Manager Remains Responsible

An organization can automate classification, summarization, forecasting, prioritization, and recommendation.

It cannot automate responsibility merely by changing the grammar of the report.

When an AI-generated sentence affects a customer, employee, supplier, budget, or operational process, someone in the organization remains responsible for:

the selected data;
the missing data;
the definition of success;
the acceptable risk;
the final interpretation;
the action taken.

“The AI recommended it” is not an explanation.

It is evidence that the organization has lost track of its own decision process.

Managers should not ask only whether an AI output is accurate.

They should ask what the sentence makes invisible.

Who disappears from the wording?

Which assumption has been converted into a fact?

Which interpretation has been presented as a measurement?

Which choice now looks inevitable?

The most dangerous AI output at work is not necessarily the statement everyone knows is wrong.

It is the statement nobody thinks to question.

*Why It Matters
*
Companies are building faster decision systems.

Reports are generated in seconds.

Customer records are summarized automatically.

Performance evaluations are drafted from existing documentation.

Sales opportunities are ranked.

Inventory is classified.

Budgets are reviewed.

Operational risks are compressed into executive summaries.

The speed is real.

So is the possibility that weak reasoning will move through the organization more quickly because it has been written more professionally.

The central management problem is therefore not only AI accuracy.

It is the relationship between language and authority.

A sentence can be grammatically complete while analytically incomplete.

It can be neutral in tone while redistributing responsibility.

It can appear objective while preserving hidden choices.

It can be useful without being sufficient.

The correct organizational response is not to distrust every AI-generated sentence.

It is to refuse automatic obedience to sentences whose confidence exceeds their evidence.

Related Academic Background

This article extends a broader research program examining how linguistic structures redistribute agency, responsibility, and institutional authority.

Suffering Without Perpetrators: The Humanitarian Passive in AI-Generated Conflict Discourse
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6753123

The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6753123

The contexts differ, but the structural question remains consistent: what happens when language preserves the event while weakening the visibility of the actor, decision-maker, or responsible institution?

About the Author

He is the author of Grammars of Power, Executable Power, and The Grammar of Objectivity, and the creator of the ongoing research series Grammars of Asymmetric Visibility.

**ResearcherID: **K-5792-2016

Website:
https://www.agustinvstartari.com/

SSRN Author Page:
https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

Zenodo Publications:
https://zenodo.org/search?q=%22Agustin%20V.%20Startari%22

Authorial Ethos

Suggested Tags

Artificial Intelligence, Management, Leadership, Decision Making, Future of Work, AI Governance, Business Strategy, Language, Automation, Enterprise AI

Your Boss Is Becoming AI’s Human Shield

Agustin V. Startari — Wed, 10 Jun 2026 12:23:03 +0000

Enterprise AI is not just changing who decides. It is changing who gets blamed.

The most dangerous sentence in corporate AI is this:

“A human is still in control.”

It sounds responsible. It sounds safe. It sounds like governance.

But in many companies, that sentence may soon mean something very different.

It may mean that a human still clicks approve.

It may mean that a human still attends the meeting.

It may mean that a human still answers the client, the employee, the regulator, or the board.

It may mean that a human still carries the title of manager.

But it does not always mean that the human truly controlled the decision.

*That is the problem.
*
Enterprise AI does not need to replace managers to weaken them. It can do something more subtle. It can structure the decision before the manager sees it, then leave the manager visible when the outcome needs to be explained.

The AI ranks.
The AI scores.
The AI summarizes.
The AI recommends.
The AI routes.
The AI flags.
The AI classifies.
The human approves.
The human explains.
The human absorbs the blame.

That is not human-centered AI.

That is human-centered liability.

The Old Question Is Too Simple

Most people still ask:

Will AI replace managers?

That is the wrong question.

The more serious question is:

Will managers remain responsible for decisions they no longer fully structure?

A manager can keep the same job and still lose authority.

He can keep the same title, same office, same dashboard, same approval rights, and same weekly meetings. Nothing dramatic needs to happen. No announcement is required. No organizational chart needs to change.

But underneath the visible role, the decision process can shift.

The sales pipeline is ranked before the sales manager opens it.

The supplier recommendation is generated before purchasing reviews it.

The risk score is assigned before finance investigates it.

The customer complaint is summarized before service escalates it.

The employee performance pattern is flagged before HR discusses it.

The manager is still there.

But he is arriving after the system has already shaped the field.

That timing matters.

A person who enters only at the final approval stage may still be accountable, but he is not fully sovereign over the decision. He is operating inside a frame built elsewhere.

A Decision Is Not Just the Final Click

Companies like to say that AI only recommends and humans decide.

This is a weak defense.

A decision is not only the final act. A decision is also everything that makes the final act appear reasonable.

What was shown first?
What was hidden?
What was ranked higher?
What was marked as risky?
What was labeled normal?
What was treated as an exception?
What required justification?
What became the default?

These are not technical details. They are managerial forces.

If an AI system recommends Vendor A, Vendor A becomes easier to choose.

If a system ranks Lead 1 above Lead 2, Lead 1 becomes easier to pursue.

If a system marks an invoice as suspicious, that invoice becomes harder to approve.

If a system summarizes a complaint as routine, that complaint becomes easier to ignore.

If a system classifies an expense under the wrong category, the company may misunderstand its own cost structure.

The human may still make the final decision.

But the system has already influenced what the human sees as obvious, urgent, safe, risky, or defensible.

*That is authority.
*
Not authority by command.

Authority by framing.

*The New Corporate Trick: Responsibility Without Control
*
A serious organization should align authority and responsibility.

If you are responsible for a decision, you should have enough authority to understand it, challenge it, change it, and audit it.

Enterprise AI can break that alignment.

The system may be designed by one team, configured by another, bought by executives, integrated by IT, trained on historical data, adjusted by vendors, and used by managers who only see the output.

But when something fails, the organization may still ask the manager:

Why did you approve this?

That question may be formally valid and operationally dishonest.

Did the manager know how the recommendation was produced?

Did he see the alternatives?

Did he know which data was excluded?

Did he understand the threshold?

Did he know whether the model was outdated?

Did he know whether the system over-weighted price, speed, volume, risk, margin, or past behavior?

Could he override the recommendation without penalty?

Could he challenge the workflow without being treated as inefficient?

If the answer is no, then the manager was not fully controlling the decision.

He was carrying it.

That is the difference.

“Human in the Loop” Can Become Theater

The phrase “human in the loop” is now everywhere.

It appears in policy documents, vendor presentations, governance frameworks, compliance programs, and executive speeches.

But the phrase can hide more than it reveals.

A human can be in the loop and still be too late.

A human can be in the loop and still lack the information needed to question the system.

A human can be in the loop and still approve what the system made easiest to approve.

A human can be in the loop and still function as decoration.

The real question is not whether a human appears somewhere in the process.

The real question is what kind of power that human actually has.

Can the human see the logic?

Can the human see competing options?

Can the human understand the recommendation?

Can the human pause the workflow?

Can the human override without punishment?

Can the human inspect the audit trail?

Can the human prove later why he accepted or rejected the system’s output?

Without those elements, “human oversight” becomes a corporate alibi.

It gives the appearance of accountability while leaving the deeper authority inside the system.

*The Dashboard Speaks Before the Manager Does
*
Modern companies do not see themselves directly.

They see themselves through systems.

Dashboards, ERPs, CRMs, HR platforms, ticketing tools, accounting software, forecasting engines, procurement systems, and reporting layers already decide what becomes visible.

When AI enters those layers, the dashboard stops being a passive display.

It becomes an editor of reality.

It decides what deserves attention.

It decides what looks abnormal.

It decides what appears first.

It decides which pattern matters.

It decides what is summarized and what is omitted.

That is not a small change.

In a meeting, the first screen often sets the agenda. The first ranking shapes the conversation. The first warning creates urgency. The first summary becomes the shared version of events.

The manager may speak second.

But the dashboard has already spoken first.

And in corporate life, speaking first is power.

The Manager as Human Shield

Companies will still need managers.

Not always because managers fully control every decision.

Sometimes because companies need a human surface.

A company cannot send an algorithm to apologize to a client.

It cannot put a workflow in front of a regulator.

It cannot make a dashboard explain a failed decision to the board.

It cannot ask a model to take responsibility for a damaged relationship, a missed opportunity, a bad supplier choice, or a wrong internal classification.

So the human remains useful.

The manager becomes the visible face of a decision whose structure may be distributed across software, data, settings, prompts, rankings, thresholds, and workflows.

That is the new role:

Not decision-maker in the full sense.

Not powerless employee either.

Something more uncomfortable.

A human shield for machine-shaped authority.

The company gets automation.

The system gets invisibility.

The manager gets responsibility.

This Is Not Anti-AI

The argument is not that companies should reject AI.

That would be lazy.

AI can improve forecasting, reduce repetitive work, detect anomalies, summarize large volumes of information, support planning, and help managers act faster.

The issue is not the existence of AI.

The issue is invisible delegation.

A company can use AI responsibly only if it knows where authority is moving.

If AI only drafts a paragraph, the risk may be limited.

If AI ranks customers, routes complaints, scores employees, classifies expenses, recommends vendors, prioritizes leads, blocks transactions, or escalates cases, then the system is no longer just assisting.

It is participating in management.

And anything that participates in management must be governed as management.

Not as a toy.

Not as a productivity hack.

Not as a harmless assistant.

As part of the authority structure of the company.

The Five Tests Every Company Should Apply

A company that claims to have human oversight should be able to pass five basic tests.

*First: the visibility test.
*
Can the manager clearly identify where AI enters the decision process?

If the answer is no, the company has hidden delegation.

*Second: the explanation test.
*
Can the manager understand why the system recommended, ranked, flagged, or classified something?

If the answer is no, the manager is approving without real comprehension.

*Third: the alternatives test.
*
Can the manager see what the system did not recommend?

If the answer is no, the system controls the field of comparison.

*Fourth: the override test.
*
Can the manager reject the recommendation without being punished by workflow friction, performance metrics, or managerial pressure?

If the answer is no, the override is theoretical.

*Fifth: the audit test.
*
Can the company reconstruct what the system recommended, what the human accepted, what the human rejected, and what happened afterward?

If the answer is no, the company does not have accountability.

It has a story about accountability.

*A Simple Example: Purchasing
*
Imagine a purchasing manager using an AI-supported procurement system.

The system recommends Vendor A.

Vendor A has a good price, acceptable delivery history, and a low risk score. The dashboard looks clean. The approval path is ready. The manager approves.

Later, the vendor fails during a high-pressure period.

Leadership asks the manager why he approved Vendor A.

The manager approved the recommendation, yes.

But did he know that Vendor B had stronger informal reliability?

Did the model understand current supplier stress?

Did it over-weight price and under-weight continuity?

Did the dashboard show excluded suppliers?

Did the system explain the trade-off?

Did the manager have time to challenge the recommendation?

If not, the company is blaming the person who clicked approve while ignoring the system that shaped approval.

That is not accountability.

That is blame displacement.

*A Simple Example: Sales
*
Now imagine a sales manager using AI lead scoring.

The model ranks prospects. The team follows the ranking. The manager reviews the pipeline. Everyone believes the sales priorities were human-approved.

But the model may favor accounts that resemble past wins.

It may ignore strategic accounts with slower cycles.

It may punish unusual opportunities.

It may over-value recent digital activity and under-value relationship history.

Months later, revenue disappoints.

The manager is asked why the team neglected certain accounts.

The answer may be uncomfortable:

Because those accounts were made invisible by the system.

Not deleted.

Not banned.

Just pushed below the attention line.

In business, that is often enough.

What falls below attention eventually falls outside action.

*A Simple Example: Finance
*
Finance looks objective because it uses numbers.

But numbers depend on classification.

An AI system that classifies expenses, assigns codes, or groups transactions is not only doing administrative work. It is shaping how the company understands itself.

Wrong classification can distort departmental costs, project margins, vendor exposure, recurring expenses, and operational leakage.

A controller may review the final report.

But if the underlying classification frame is wrong, the review begins too late.

The company will think it is looking at reality.

It is actually looking at reality after software has translated it.

And every translation carries authority.

The Real Risk: Accountability Laundering

The central risk is accountability laundering.

Authority enters the system.

Responsibility exits through the human.

The workflow shapes the decision, but the approval log names the manager.

This is extremely convenient.

For vendors, it allows systems to influence decisions while claiming they only assist.

For executives, it creates efficiency while preserving plausible human accountability.

For companies, it protects the institution when outcomes fail.

For managers, it creates a dangerous trap.

They remain visible enough to be blamed, but not powerful enough to fully control the process.

That is the structure serious companies need to confront.

Not because AI is evil.

Because bad accountability design is enough.

The Question Managers Should Ask

Every manager facing an AI-shaped workflow should ask:

Am I being given authority, or only liability?

That question is more important than any product demo.

If the system gives the manager visibility, explanation, alternatives, override rights, and audit trails, then AI may genuinely support management.

But if the system only provides polished outputs, ranked lists, clean dashboards, automated summaries, and recommended actions, then the manager is not fully governing.

He is operating inside a frame.

And if he is responsible for that frame without controlling it, he is not being empowered.

He is being exposed.

The Core Claim

AI will not simply replace managers.

That headline is too crude.

The deeper transformation is more institutional.

AI may keep managers in place while moving real authority into systems.

The manager remains the signer.

The system becomes the framer.

The company preserves deniability.

The blame remains human.

That is why the future of enterprise AI cannot be discussed only in terms of productivity.

It must be discussed in terms of authority.

Who frames the decision?

Who sees the alternatives?

Who controls the default?

Who owns the recommendation?

Who can challenge the workflow?

Who gets blamed when the system fails?

Until companies can answer those questions clearly, “human oversight” should not reassure us.

It should make us suspicious.

Because a human can be present and still be used.

And the next corporate AI scandal may not begin with a machine making a decision alone.

It may begin with a manager approving a decision that was never fully his.

*Related Academic Background
*
This article extends my broader work on artificial intelligence, language, authority, and institutional responsibility.

My research examines how formal systems redistribute agency and accountability through syntax, classification, interface design, automation, and institutional language. This includes my work on AI-powered ERP misclassification, executable authority, corporate decision structures, and the grammar of responsibility inside automated environments.

*The central issue is consistent:
*
When systems classify, rank, route, summarize, recommend, or frame events, they do not merely assist institutions.

They reshape how institutions see, decide, justify, and blame.

Author

Agustin V. Startari is a linguistic theorist, author, and researcher in historical studies. His work examines how language, artificial intelligence, and formal systems redistribute authority, agency, and responsibility in contemporary institutions. He is the author of Grammars of Power, Executable Power, The Grammar of Objectivity, and Grammars of Asymmetric Visibility.

ResearcherID: K-5792-2016
Website: https://www.agustinvstartari.com/
SSRN: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
Zenodo: https://zenodo.org/search?q=%22Agustin%20V.%20Startari%22

Ethos

— Agustin V. Startari

Your Manager Is Becoming an AI Front-End

Agustin V. Startari — Tue, 02 Jun 2026 13:08:50 +0000

How enterprise AI turns human managers into the visible interface of decisions already shaped by software
TLDR

The next corporate AI problem is not that artificial intelligence will replace managers. It is that managers may remain formally responsible while AI systems quietly structure the decisions before managers see them.

The manager stays visible.
The workflow becomes automated.
The recommendation becomes the default.
The dashboard becomes the briefing.
The human becomes the interface.

That is not replacement.
It is a more subtle redistribution of authority.

*1. The Manager Is Still There, But Too Late
*
The common fear is simple: AI will replace managers.

That fear is too obvious.

A more realistic change is already emerging inside companies. The manager is not removed. The manager remains in the meeting, approves the purchase order, reviews the dashboard, signs the report, accepts the forecast, and answers for the result.

But the decision may have begun somewhere else.

Before the manager acts, an AI system may have already ranked the leads, flagged the vendor, assigned the risk score, selected the metric, summarized the customer complaint, recommended the reorder quantity, or routed the approval path.

The human manager still appears to decide. But the field of decision has already been arranged.

This is the new corporate structure:

AI does not need to become the boss.
It only needs to prepare the boss’s options.

That preparation matters because management is not only final approval. Management is also the power to decide what appears first, what appears urgent, what appears risky, what appears normal, what appears exceptional, and what never appears at all.

A manager who only sees the final ranked list is not operating from raw business reality. He is operating from a pre-structured version of that reality.

The company may still say, “The manager decided.”

The better question is:

Who shaped the decision before the manager saw it?

*2. The Dashboard Becomes the Real Briefing
*
In many companies, the dashboard is no longer just a reporting tool.

It is the first layer of managerial perception.

Sales teams look at pipeline dashboards.
Purchasing teams look at vendor dashboards.
Inventory teams look at stock dashboards.
Finance teams look at classification dashboards.
Customer service teams look at escalation dashboards.
Executives look at summary dashboards.

This creates a simple but powerful shift.

What the dashboard shows first becomes what the company discusses first.

That is not neutral.

If an AI system ranks customers by predicted revenue, the sales team may naturally focus on those customers. If the model underestimates a smaller but strategically important client, that client may disappear from attention.

If an inventory dashboard highlights overstock risk but hides supplier fragility, managers may discuss warehouse efficiency while missing the coming supply problem.

If a finance dashboard flags unusual expenses but does not explain the classification logic, the manager may spend time reviewing harmless anomalies while a more serious pattern remains invisible.

A dashboard does not only display reality. It edits reality.

Once AI enters that editing layer, the company’s attention becomes machine-shaped.

This is not science fiction. It is ordinary business.

The meeting begins.
The screen is opened.
The dashboard speaks first.
The manager reacts second.

That order is the real chain of command.

*3. Recommendation Is Not Neutral
*
Companies often treat AI recommendations as harmless because the human still has the final word.

That is a weak argument.

A recommendation inside a business workflow is not the same as a casual suggestion. It appears inside a system of pressure, speed, workload, habit, hierarchy, and operational urgency.

When a system recommends a vendor, that vendor becomes easier to choose.

When a system ranks a lead as high priority, that lead becomes easier to pursue.

When a system marks a customer as low risk, that customer becomes easier to ignore.

When a system flags an invoice as suspicious, that invoice becomes harder to approve.

When a system suggests a reorder quantity, that quantity becomes the default starting point.

The recommendation does not need to force the decision. It only needs to make one path easier than the others.

That is how corporate authority often works.

Not by command.
Not by violence.
Not by dramatic control.

By defaults.

Defaults are powerful because they reduce friction. The recommended option is already there. The alternative requires explanation. The override requires effort. The exception requires justification.

So the system does not need to say, “You must choose this.”

It only needs to say:

“This is the recommended option.”

In corporate life, that is often enough.

*4. The Human Override Myth
*
Most companies defend AI adoption with one reassuring phrase:

There is human oversight.

The phrase sounds responsible. But it often hides a major problem.

Oversight is not meaningful if the human only sees an already processed result.

A manager may be able to override an AI recommendation. But what exactly is being overridden?

The final number?
The risk label?
The priority score?
The routing decision?
The classification?
The hidden threshold?
The model weighting?
The excluded alternative?

Human oversight becomes weak when the human cannot see how the recommendation was produced.

Imagine a purchasing manager reviewing an AI-generated supplier recommendation. The system suggests Vendor A. The price looks reasonable. The delivery score looks good. The risk label is low. The approval path is already prepared.

The manager can technically choose Vendor B.

But if Vendor B was ranked lower by a model using incomplete or outdated data, the manager may never know. If the system ignored informal supplier reliability, the manager may never see it. If the model over-weighted price and under-weighted strategic continuity, the dashboard may not reveal that.

The manager has oversight over the output, but not over the construction of the output.

That is not full oversight.

It is supervised acceptance.

The same problem appears in sales. A sales manager may review a list of prioritized leads. The model ranks them. The team follows the ranking. Later, leadership asks why certain accounts were neglected.

The manager approved the focus list.

But did the manager design the scoring logic?
Did the manager know which signals were excluded?
Did the manager know why one client was pushed down?
Did the manager know whether the model understood local market context?

If the answer is no, then the manager was not fully managing.

He was operating through a machine-shaped frame.

*5. AI Does Not Need to Fire You to Manage You
*
The public debate usually asks whether AI will replace jobs.

That question is too narrow.

AI can transform a job without eliminating it.

A person can keep the same title, same office, same salary, and same formal responsibility while losing control over the structure of his own decisions.

This is especially important for managers.

AI does not need to fire a manager to manage the manager.

It can define which metrics matter.
It can order the daily priorities.
It can flag which employees need attention.
It can rank which customers deserve follow-up.
It can recommend which expenses deserve review.
It can decide which tickets look urgent.
It can summarize which problems leadership sees first.
It can classify which actions look normal or abnormal.

The manager still works. But the system increasingly defines the environment in which that work happens.

That is the quiet transformation.

AI does not sit above the manager on the organization chart. It sits beneath the manager inside the workflow.

And because it sits beneath, it is harder to see.

A human boss gives instructions.
A system changes the conditions.

A human boss says, “Do this.”
A system makes one action easier, faster, more visible, more defensible, or more urgent than another.

That is a different kind of management.

It is not command by speech.
It is command by structure.

*6. The Manager as Liability Shield
*
Here is the uncomfortable part.

Companies may still need human managers not because humans control every decision, but because humans absorb responsibility.

A company cannot easily blame a dashboard in a board meeting.

It cannot send a workflow to explain itself to a client.

It cannot ask a model to defend a decision before a regulator.

It cannot make an algorithm apologize to a supplier, employee, investor, or customer.

So the human manager remains useful.

Not always as the true origin of the decision, but as the visible owner of the outcome.

This creates a dangerous split:

Authority becomes distributed across systems.
Responsibility remains concentrated on people.

The AI ranks.
The AI routes.
The AI flags.
The AI classifies.
The AI recommends.
The AI prepares.
The human approves.
The human answers.

That structure is convenient for organizations because it preserves the appearance of accountability.

There is always someone to ask.
There is always someone to blame.
There is always someone whose name appears in the approval history.

But the deeper authority may be hidden in system configuration, model design, data quality, prompt structure, workflow logic, permission settings, thresholds, and dashboard architecture.

That is why “human in the loop” is not enough.

A human can be in the loop and still arrive too late.

A human can approve a decision without controlling the frame.

A human can accept a recommendation without knowing what alternatives were suppressed.

A human can be responsible for a process whose authority was already embedded elsewhere.

This is the new liability problem of enterprise AI.

Not that nobody is responsible.

That the visible responsible person may not be the real structuring agent.

*7. The New Test: Who Framed the Decision?
*
Companies need a better test for AI governance.

The old question is:

Who made the decision?

The new question should be:

Who framed the decision before it was made?

This question is more precise because modern corporate decisions rarely appear as isolated acts. They are prepared through multiple layers.

Data enters the system.
A model interprets it.
A rule filters it.
A dashboard displays it.
A ranking orders it.
A workflow routes it.
A recommendation appears.
A human reviews it.
An action is taken.

By the time the human acts, several forms of authority may have already operated.

So the company should ask:

Who selected the data?
Who defined the categories?
Who set the threshold?
Who designed the ranking?
Who wrote or approved the prompt?
Who decided what the dashboard shows first?
Who decided what requires escalation?
Who decided what can move automatically?
Who can override the system?
Who reviews override patterns?
Who audits the business impact?

These are not technical details. They are management questions.

Because whoever controls the frame controls much of the decision.

A manager who cannot answer these questions may still be a manager on paper. But operationally, he may be managing through a system he does not fully govern.

That is the real risk.

Not that managers vanish.

That managers become front-ends.

Why This Matters

This matters because companies already live inside software.

The modern company is not managed only through meetings, calls, emails, and human judgment. It is managed through ERP systems, CRM platforms, HR dashboards, inventory tools, ticketing systems, accounting software, approval workflows, forecasting engines, and reporting layers.

When AI enters those layers, it does not need to become conscious to become powerful.

It only needs to influence sequence, priority, classification, routing, visibility, and timing.

Those are managerial functions.

A company can survive a bad AI paragraph.

A company may not survive repeated AI-shaped distortions in purchasing, finance, inventory, sales, compliance, or customer service.

The real danger is not one dramatic failure. It is the slow normalization of machine-shaped management with human-shaped accountability.

The manager becomes the person who explains decisions that were already partially arranged by systems.

That is not efficiency alone.

It is institutional redesign.

A Simple Example: Sales

Consider a sales team using AI lead scoring.

The system ranks prospects from highest to lowest priority. The sales manager reviews the list. The team focuses on the top group. Everyone says the manager chose the priorities.

But the model may have over-weighted recent activity and under-weighted long-term relationship value.

It may have missed local knowledge.

It may have penalized accounts that do not behave like the historical data.

It may have pushed unusual but valuable leads below the attention line.

The manager still made a decision.

But the system shaped what looked worth deciding.

If revenue falls later, leadership may ask the sales manager what happened.

The better question is whether the company audited the scoring logic that shaped the sales manager’s attention.

A Simple Example: Purchasing

Now consider purchasing.

An AI system recommends a vendor based on price, delivery history, payment terms, and risk score. The purchasing manager approves it.

Everything looks normal.

But the system may not understand that another supplier, slightly more expensive, has a stronger informal reliability record during high-pressure periods. It may not understand that a vendor with good historical data is currently unstable. It may not understand that a small delay in one product category creates larger operational damage elsewhere.

The manager sees a clean recommendation.

The business later experiences disruption.

Formally, the purchasing manager approved the choice.

Operationally, the AI structured the choice.

That distinction matters.

A Simple Example: Finance

In finance, AI classification can look harmless because it appears administrative.

But classification is not minor.

An expense code affects reporting.
Reporting affects interpretation.
Interpretation affects management.
Management affects future decisions.

If an AI system misclassifies recurring expenses, the company may misunderstand departmental cost, project profitability, vendor exposure, or operational leakage.

The controller may review the final report. But if the underlying classification frame is wrong, the review begins too late.

Finance is not only numbers. It is the structure through which the company becomes legible to itself.

When AI changes that structure, it changes what the company thinks it knows.

A Simple Example: Customer Service

Customer service depends heavily on escalation.

Which complaint is urgent?
Which customer receives attention first?
Which issue appears systemic?
Which case is treated as ordinary noise?

If AI ranks complaints, summarizes customer messages, and recommends escalation levels, it shapes the company’s moral and commercial attention.

A customer may not be ignored because a person decided to ignore him.

He may be ignored because the system made his complaint look less urgent than another.

That is not a small distinction.

In service environments, visibility is care.
Invisibility is neglect.

AI that controls visibility participates in the treatment of the customer.

The Corporate Illusion

The corporate illusion is that AI remains subordinate because humans still approve the final act.

But final approval is not the whole decision.

A decision is also made through:

What is shown.
What is hidden.
What is ranked.
What is delayed.
What is escalated.
What is framed as risky.
What is framed as routine.
What is made easy.
What is made difficult.
What requires justification.
What becomes the default.

AI agents can influence all of these without appearing as formal decision-makers.

That is why the language of “assistant” is no longer enough.

An assistant helps someone act.

A front-end lets someone appear to act while the deeper system structures the action.

That is the difference.

What Companies Should Track

If companies want real accountability, they need to audit AI authority, not just AI output.

For every AI system that influences operational decisions, companies should be able to answer:

What business process does it affect?

Does it only summarize, or does it recommend?

Does it rank people, customers, vendors, products, tasks, risks, or priorities?

Does it trigger workflows?

Does it change records?

Does it block or delay action?

Does it create a default option?

Does the human reviewer see alternatives?

Can the human reviewer understand the recommendation path?

Are overrides logged?

Are ignored recommendations logged?

Are business outcomes compared against AI recommendations?

Who owns the system when it is wrong?

This is not anti-technology. It is basic managerial hygiene.

A company that cannot answer these questions is not using AI strategically.

It is delegating authority without a map.

What Managers Should Demand

Managers should not reject enterprise AI.

They should reject invisible delegation.

A serious manager should demand five things before accepting AI-shaped workflows.

First, visibility.

The manager should know where AI enters the decision chain.

Second, explanation.

The manager should understand the main reason behind a recommendation, ranking, classification, or escalation.

Third, alternatives.

The system should show what was not selected, not only what it recommends.

Fourth, override rights.

A manager must be able to challenge the output without being treated as an obstacle to efficiency.

Fifth, audit trails.

The company must record what the system recommended, what the human accepted, what the human rejected, and what happened afterward.

Without these five elements, managers risk becoming decorative accountability.

They remain responsible because the company needs a human face.

But they do not fully control the system that shapes the decision.

What Developers Should Understand

Developers building enterprise AI are not only building tools.

They are designing decision environments.

A ranking is not just a ranking when it determines which customer gets called first.

A classification is not just a classification when it affects financial reporting.

A prompt is not just a prompt when it converts messy business language into executable workflow.

A threshold is not just a threshold when it decides whether something is escalated, blocked, approved, or ignored.

A default is not just a default when most users accept it.

This does not mean developers are personally responsible for every corporate outcome. But it does mean technical design has managerial consequences.

The central design question should be:

Where does this system acquire practical authority?

If the answer is “nowhere,” the system may be a tool.

If the answer is “in ranking, routing, blocking, classifying, escalating, or triggering action,” the system is part of management.

It should be treated that way.

The Core Claim

Your manager is not disappearing.

That would be too simple.

Your manager may remain exactly where he is, with the same title, same duties, same meetings, same approval rights, and same formal accountability.

But underneath that visible role, enterprise AI may increasingly structure what the manager sees, what the manager ignores, what the manager approves, what the manager questions, and what the manager must later explain.

That is the deeper transformation.

The manager becomes the human front-end of AI-shaped authority.

The face remains human.
The frame becomes synthetic.
The responsibility remains visible.
The authority becomes distributed.
The decision still has a signer.
But the decision may no longer begin with the signer.

That is the future companies need to audit.

Not because AI is evil.

Because authority that hides inside workflows is still authority.

*Related Academic Background
*
This article extends my broader work on how artificial intelligence systems redistribute agency, responsibility, and authority through formal and operational structures.

*Related paper:
*
Agustin V. Startari, “Expense Coding Syntax: Misclassification in AI-Powered Corporate ERPs,” SSRN Electronic Journal, 2025.
https://doi.org/10.2139/ssrn.5361952

That paper examines how AI-powered ERP systems can produce misclassification risks when automated language and coding structures convert business events into financial categories. The broader issue is the same: when AI systems classify, route, prioritize, or encode business events, they do not merely represent the company. They reshape how the company becomes visible, manageable, and accountable.

*Why It Matters for Everyone
*
This is not only a problem for executives.

Employees should care because their work may be evaluated through AI-shaped metrics.

Managers should care because they may become responsible for decisions they did not fully structure.

Developers should care because technical defaults can become managerial authority.

Customers should care because their complaints, requests, risks, and value may be ranked before a person ever sees them.

Investors should care because operational opacity can hide risk until it becomes financial damage.

Regulators should care because formal human oversight may not be enough when the decision frame is machine-generated.

The real AI revolution in companies will not always look spectacular.

It may look like a normal dashboard.
A normal score.
A normal recommendation.
A normal approval path.
A normal classification.
A normal workflow.

That is exactly why it matters.

The most powerful systems are often the ones that disappear into routine.

*Call to Action
*
Read more of my work on artificial intelligence, language, authority, and institutional responsibility:

Website: https://www.agustinvstartari.com/
SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
Zenodo publications: https://zenodo.org/search?q=%22Agustin%20V.%20Startari%22

Author

Researcher ID: K-5792-2016
Author website: https://www.agustinvstartari.com/
SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

**
**

Agustin V. Startari.

Stop Calling It an AI Assistant. It’s Already Managing Your Company

Agustin V. Startari — Fri, 22 May 2026 12:39:00 +0000

The hidden authority of AI agents inside ERP, purchasing, inventory, approvals, and enterprise workflows
TL;DR

The next enterprise AI risk is not that a chatbot writes a bad email. It is that an AI agent quietly enters the operational layer of the company and starts ranking priorities, routing approvals, classifying risk, delaying purchases, escalating tickets, flagging customers, and shaping managerial decisions before anyone calls it management.

Companies still describe these systems as “assistants” because the word sounds harmless. But once a system can trigger action inside an ERP, CRM, inventory platform, purchasing workflow, or finance dashboard, it is no longer merely assisting.

It is participating in management.

The problem is not automation itself. The problem is invisible delegation: authority moves into workflows, prompts, thresholds, model outputs, and software rules, while responsibility remains formally assigned to humans who may only see the final recommendation.

That is how an AI assistant becomes a shadow manager.

Meta Description

AI agents inside ERP, purchasing, inventory, finance, and enterprise workflows are no longer just assistants. They increasingly rank, route, classify, escalate, and shape operational decisions. This article explains how invisible delegation turns AI into shadow management.

*1. The Assistant Myth
*
Everyone calls them AI assistants because “assistant” sounds harmless.

An assistant helps. An assistant supports. An assistant drafts, summarizes, searches, reminds, and organizes.

That language is accurate only while the system remains outside the decision chain.

Once the system can rank leads, block a purchase order, classify a vendor, flag a customer, recommend a reorder quantity, trigger a workflow, escalate a ticket, assign urgency, or prepare an approval path, the term “assistant” becomes misleading.

At that point, the system is no longer just helping a manager.

It is shaping the managerial environment before the manager acts.

This distinction matters because enterprise work is not made only of final decisions. Most corporate power lives in prioritization, routing, classification, timing, and escalation. Whoever controls those layers does not need to sign the final approval to influence the outcome.

A sales manager may still approve the weekly focus list, but if an AI system ranked the leads first, part of the commercial decision has already been made.

A purchasing manager may still approve the order, but if the ERP agent has already recommended the vendor, adjusted the quantity, flagged the risk, and routed the approval path, the decision has already been pre-shaped.

A finance controller may still review the expense, but if an AI classifier has already coded the transaction and assigned its risk level, the human review begins inside a frame built by the system.

That is the assistant myth: the company believes AI is supporting decisions when, in practice, AI is already structuring them.

The human manager remains visible. The automated manager remains embedded.

*2. From Chatbots to Agents
*
The first wave of enterprise AI was easy to understand.

A chatbot answered questions. A writing tool drafted text. A summarizer compressed documents. A search assistant retrieved information.

Those tools could be wrong, but their wrongness usually stayed inside language. A bad answer could be corrected. A weak summary could be rewritten. A hallucinated paragraph could be deleted.

AI agents are different.

An agent is not only a text generator. It receives a goal, consults tools, uses data, plans steps, invokes functions, and may change the state of a system.

That shift changes the risk model.

A chatbot says: “You may want to reorder this item.”

An agent creates a draft purchase order.

A chatbot says: “This customer seems high priority.”

An agent moves that customer to the top of the pipeline.

A chatbot says: “This invoice may be misclassified.”

An agent changes the expense code.

A chatbot says: “This ticket looks urgent.”

An agent escalates it to another department.

The first system produces language. The second system produces operational consequences.

That is the line companies often fail to mark.

The word “assistant” hides the transition from advice to action. But enterprise systems do not care whether a workflow was triggered by a human, a script, a rule, or a model. Once the system state changes, the company has acted.

This is where AI becomes managerial.

Not because it has a job title. Not because it sits in a meeting. Not because it appears on the organization chart.

It becomes managerial because it shapes attention, timing, access, priority, and execution.

*3. Where the Shadow Manager Appears
*
The shadow manager does not appear as a robot boss.

It appears as a workflow.

It appears as a recommendation that nobody questions because it came from the dashboard.

It appears as a priority score.

It appears as a blocked order.

It appears as an automatic escalation.

It appears as a vendor warning.

It appears as a risk label.

It appears as an approval path that feels procedural but was shaped by a model.

This is already visible in ordinary enterprise operations.

In sales, an AI system may rank leads according to predicted conversion. That ranking influences which customer receives attention first. The salesperson may think they are choosing, but the field of choice has already been ordered.

In purchasing, an AI agent may recommend suppliers based on price, delivery history, stock availability, vendor score, payment terms, or risk profile. That recommendation can quietly shift purchasing behavior away from human relationship knowledge and toward model-weighted criteria.

In inventory, an agent may recommend reorder quantities, flag slow-moving items, identify overstocks, and predict demand. If those predictions are wrong, the error does not remain theoretical. It becomes cash tied in stock, delayed sales, missing products, emergency orders, or warehouse friction.

In customer service, an AI system may decide which complaint deserves escalation. That decision affects response time, customer satisfaction, and the perceived seriousness of the issue.

In finance, AI classification may assign expenses, flag anomalies, group transactions, or prepare reports. If the classification is wrong, the error can affect reporting quality, cost-center visibility, departmental accountability, and managerial interpretation.

In operations, AI may summarize performance, highlight bottlenecks, and define what leadership sees first. That is not neutral. The first metric shown often becomes the first problem discussed.

The shadow manager does not need to make every decision.

It only needs to shape the order in which decisions become visible.
**

The Hidden Chain of Command**

Traditional corporate authority is usually imagined as a clean hierarchy.

Owner. Executive. Manager. Supervisor. Employee. Action.

Enterprise AI complicates that structure.

The real chain can become:

Policy. System configuration. Data source. Prompt. Model output. Workflow trigger. Dashboard ranking. Human approval. Operational action.

The human remains inside the chain, but not always at the beginning of it.

This matters because responsibility is often assigned at the visible end of the process, while influence may have entered much earlier.

A manager may approve a purchase order without knowing that the recommended quantity was produced by a demand model trained on incomplete seasonal data.

A sales lead may be ignored because a scoring system placed it below the threshold, even though the model failed to capture a relationship or local market signal.

A warehouse adjustment may be flagged as suspicious because the system misread an operational pattern.

An accounts receivable account may be deprioritized because the dashboard over-weighted one indicator and under-weighted another.

In each case, the human did not disappear. But the human arrived late.

That is the key structure.

The visible manager signs, approves, reviews, or accepts. The invisible system has already arranged the options.

This is not the end of human authority. It is the redistribution of authority across software layers.

The company still says “the manager decided.”

But the better question is: who structured the decision before the manager saw it?

*5. Why ERP Makes This More Serious
*
AI inside a document editor is useful.

AI inside an ERP is different.

An ERP is not just software. It is the operational nervous system of the company. It connects sales, purchasing, inventory, accounting, logistics, invoicing, vendor records, customer records, product movement, and reporting.

When AI enters that layer, errors become operational.

A weak paragraph is a content problem. A wrong reorder suggestion is a cash problem. A bad vendor classification is a supply problem. A wrong expense code is a reporting problem. A bad lead ranking is a revenue problem. A wrong delivery priority is a customer problem. A bad inventory signal is a service problem.

This is why enterprise AI cannot be judged only by generic model benchmarks.

A model does not need to be generally “smart” to create damage. It only needs to be wrong at the point where the business acts.

The most dangerous AI in a company may not be the most advanced model. It may be the boring workflow nobody audits.

The purchase recommendation. The lead score. The automatic approval rule. The AR risk flag. The reorder suggestion. The inventory exception. The vendor ranking. The escalation logic.

These systems become powerful because they sit close to action.

They do not merely describe the business. They participate in running it.

This is why companies need a different vocabulary. Calling these systems “assistants” is not enough. In operational environments, an AI system should be classified according to its action rights.

Can it read data? Can it recommend action? Can it trigger action? Can it block action? Can it route approval? Can it change records? Can it reorder priorities? Can it modify system state?

The moment the answer becomes yes, the company is no longer dealing with a passive tool.

It is dealing with delegated operational authority.

6. The Accountability Gap

Most enterprise AI discussions focus on hallucination.

That focus is too narrow.

Hallucination matters when a model invents facts. But in enterprise workflows, the more common danger may be misclassification, over-ranking, under-ranking, false escalation, silent omission, wrong routing, and unexamined recommendation.

The system does not need to hallucinate to create harm.

It can use real data and still produce a bad decision structure.

It can classify an account as low priority because the available data is incomplete.

It can recommend delaying a purchase because it underestimates demand.

It can flag an employee action as unusual because the workflow does not understand local practice.

It can prioritize one customer because the model values transaction size over strategic relevance.

It can mark an item as slow-moving while ignoring a coming seasonal spike.

These are not hallucinations. They are operational distortions.

The accountability gap appears when nobody can answer seven basic questions:

What data did the system use?

What rule or model produced the recommendation?

What threshold was applied?

What alternatives were suppressed?

Who reviewed the output?

Who had authority to override it?

What happened after the recommendation was accepted?

Without those answers, the company has built authority without memory.

A human manager can be questioned. A workflow often cannot. A model output may be overwritten. A system recommendation may leave no readable trace. A dashboard may show the result without exposing the path.

That is not automation maturity. It is managerial opacity.

7. What Developers and Operators Should Log

The solution is not to reject AI agents.

The solution is to stop pretending they are harmless assistants once they touch operational decisions.

If an AI system can influence action, it needs an audit trail.

At minimum, enterprise AI agents should log:

Input source.

Data timestamp.

Prompt or instruction version.

Model version.

Tool used.

External system accessed.

Rule applied.

Threshold used.

Recommendation generated.

Action triggered.

Human reviewer.

Override status.

Final decision.

Business impact.

Error category, if later detected.

This is not bureaucratic decoration. It is the basic condition for operational accountability.

If a purchasing agent recommends a quantity, the company should know why.

If a sales agent ranks a lead, the company should know what signals mattered.

If a finance classifier assigns an expense category, the company should know which rule or model produced the classification.

If an inventory agent flags an item, the company should know whether the signal came from sales history, warehouse movement, vendor delay, forecast variance, or a model-generated probability.

This is how enterprise AI becomes governable.

Not by asking whether the system is impressive.

By asking whether its authority is visible.

A Better Test: Authority, Not Intelligence

The wrong question is:

“Is this AI intelligent?”

The better question is:

“What authority does this AI have?”

That question changes the entire evaluation.

A simple model with access to an ERP approval workflow may have more operational power than a more advanced model trapped inside a chat window.

A mediocre classifier embedded in finance may produce more business risk than a brilliant writing assistant.

A small automation that blocks orders may matter more than a large model that only drafts emails.

Enterprise AI should therefore be evaluated by authority level, not only by capability.

Level 1: It reads information.

Level 2: It summarizes information.

Level 3: It recommends action.

Level 4: It routes action.

Level 5: It triggers action.

Level 6: It blocks action.

Level 7: It changes system state with limited human review.

The higher the level, the stronger the audit requirement.

This framework is simple, but it prevents the core mistake: treating all AI outputs as if they were merely advisory.

They are not.

Some outputs become instructions. Some recommendations become defaults. Some defaults become behavior. Some behavior becomes policy. Some policy becomes authority.
**

Why Managers Should Care**

Managers should care because AI agents can make them responsible for decisions they did not fully structure.

A manager may be asked why an order was delayed. The real cause may be a workflow rule.

A manager may be asked why a customer was ignored. The real cause may be a lead-ranking model.

A manager may be asked why inventory ran short. The real cause may be a bad demand signal.

A manager may be asked why expenses were misclassified. The real cause may be an automated coding system.

In all these cases, the manager remains accountable while the system remains partially invisible.

That is a bad trade.

AI should reduce operational burden, not create a fog of responsibility.

For managers, the practical rule is direct: never allow an AI agent to influence action without knowing where its recommendation appears, how it is produced, how it can be challenged, and who owns the final decision.

Management cannot be delegated into a black box and then recovered only when something fails.
**

Why Developers Should Care**

Developers should care because every enterprise AI agent is also a governance system.

A function call is not just a technical event when it changes a purchase order, a customer priority, a stock level, an invoice category, or an approval path.

A ranking algorithm is not just a ranking algorithm when it determines who gets attention first.

A classification model is not just a classifier when departments rely on it for reporting, escalation, or compliance.

A prompt is not just a prompt when it controls how operational language is converted into action.

This means developers are not merely building features. They are designing decision environments.

That does not mean developers become the moral owners of every business outcome. It means technical design choices can create managerial consequences.

What gets logged matters.

What gets hidden matters.

What becomes the default matters.

What can be overridden matters.

What requires human review matters.

What silently moves forward matters.

Enterprise AI development should therefore include one question in every workflow design:

Where does this system acquire practical authority?

That question is more useful than vague debates about whether AI will replace managers. In many companies, replacement is not the first step. Quiet redistribution is.

The job title stays human. The workflow becomes automated. The decision frame becomes synthetic. The responsibility remains unclear.

That is the shadow manager problem.

11. The Core Claim

Stop calling it an AI assistant if it can manage priority, routing, approval, classification, escalation, or execution.

Inside enterprise systems, assistance can become authority without changing its name.

That is the risk.

Not evil AI. Not science fiction. Not a robot CEO. Not a dramatic replacement of human managers.

The real shift is quieter.

AI enters the company as a helper. It gets connected to tools. It receives access to business data. It starts producing recommendations. Those recommendations become defaults. Those defaults shape workflows. Those workflows shape decisions. Those decisions shape the company.

By the time leadership notices, the assistant is already managing part of the business.

The future of enterprise AI will not be decided only by which model writes better emails or produces cleaner summaries.

It will be decided by which systems can act inside companies without making responsibility disappear.

Why It Matters

This matters because modern companies already run through software.

Sales teams follow dashboards. Purchasing follows workflows. Inventory follows system signals. Finance follows classifications. Managers follow reports. Executives follow summaries.

When AI enters those layers, it does not need to dominate the company to change it. It only needs to reorder what the company sees, delays, escalates, approves, or ignores.

The practical danger is not that AI becomes conscious.

The practical danger is that AI becomes procedural.

It becomes part of how the company moves.

And once it moves the company, it must be audited as authority, not described as assistance.

Related Academic Background

This article extends my broader work on how automated language systems redistribute agency, responsibility, and authority through formal structures.

*Related paper:
*
Expense Coding Syntax: Misclassification in AI-Powered Corporate ERPs https://doi.org/10.2139/ssrn.5361952

The paper examines how AI-powered ERP systems can produce misclassification risks when automated language and coding structures convert business events into financial categories. The broader issue is the same: when automated systems classify, route, or encode decisions, they do not merely represent the business. They reshape how the business becomes legible and actionable.

*About the Author
*
Agustin V. Startari is a linguistic theorist, author, and researcher in historical studies. His work examines the relationship between artificial intelligence, syntax, authority, institutional discourse, and the disappearance of agency in automated language systems.

He is the author of Grammars of Power, The Grammar of Objectivity, Suffering Without Perpetrators, The Grammar of Asymmetric Visibility, and Expense Coding Syntax. His research focuses on how language models and institutional systems redistribute responsibility through grammatical, operational, and procedural form.

*Personal website: *https://www.agustinvstartari.com/

*SSRN Author Page: * https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

ResearcherID: K-5792-2016

Authorial Ethos

Suggested Tags

AI, Enterprise AI, AI Agents, ERP, Automation, Workflow Automation, Management, Operations, Purchasing, Inventory, Finance, Business Software, AI Governance, Agentic AI, NetSuite, Enterprise Software, Accountability, Decision Systems

AI Says Palestinians Are Dying, But Not Who Is Killing Them

Agustin V. Startari — Wed, 20 May 2026 14:33:35 +0000

The humanitarian passive, press language, and the AI-mediated grammar of responsibility loss in Gaza

*TL;DR
*
Modern conflict reporting does not always erase suffering. Often, it does something more subtle: it shows suffering while weakening the grammar of responsibility.

In AI-generated summaries, automated headlines, platform moderation outputs, and press rewrites, Palestinian civilians can appear as “killed,” “displaced,” “affected,” or “caught in conflict,” while the actor responsible for producing that condition disappears from the sentence.

*That structure is the humanitarian passive.
*
It is not simply passive voice. It is a political grammar through which suffering remains visible while agency becomes optional.

This post builds on two academic papers:

Suffering Without Perpetrators: The Humanitarian Passive in AI-Generated Conflict Discourse
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6753123

The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6787439

*Meta Description
*
AI-generated summaries and press language often describe Palestinian suffering while removing the actors responsible for it. This post explains the humanitarian passive, a grammar of responsibility loss in conflict discourse.

*1. What Is the Humanitarian Passive?
*
The humanitarian passive is a recurring linguistic pattern in which civilian suffering is described without clearly naming the actor, institution, army, state, or command structure that produced it.

Example:

“Civilians were killed in Gaza.”

This sentence may be factually compatible with reality, but it is structurally incomplete.

It tells us that death occurred. It does not tell us who caused it.

Compare:

“Israeli airstrikes killed civilians in Gaza.”

The second version does something the first one avoids: it assigns agency.

The difference is not cosmetic. It changes the reader’s map of causality.

The humanitarian passive separates suffering from responsibility. The victim remains visible. The perpetrating structure becomes grammatically optional.

*2. Why AI Makes This Pattern More Powerful
*
Large language models are trained to produce fluent, moderate, institutionally acceptable language. In politically sensitive contexts, that often means avoiding direct attribution unless the source material forces attribution.

This matters because AI systems frequently summarize conflict through neutralized phrases:

“Violence escalated.”
“Buildings were destroyed.”
“Families were displaced.”
“Casualties were reported.”
“Humanitarian conditions deteriorated.”

These phrases sound objective. But they often remove the actor from the event.

In war reporting, that is not neutral. It changes how responsibility is perceived.

When AI summarizes press material about Gaza, it may reproduce a pattern already present in institutional journalism: Israeli military action becomes background context, while Palestinian suffering becomes humanitarian scenery.

The result is a sentence structure where the harm is visible, but the agent is grammatically distant.

*3. Israel, Zionist Narrative, and Asymmetric Visibility
*
This post uses “Zionist narrative” as a political category, not as a synonym for Jewish identity.

The relevant structure is this: Israeli state discourse often frames military violence through security, self-defense, counterterrorism, existential threat, and regional instability.

Within that frame, Palestinian death is frequently described as tragic, unfortunate, complex, or context-dependent, rather than as the direct result of identifiable military and political action.

That is where the humanitarian passive becomes useful.

It allows a narrative to say:

“There is suffering.”

without saying:

“A state produced this suffering through named policies, weapons, targeting systems, command decisions, and legal justifications.”

The suffering is acknowledged. The machinery is blurred.

This is more effective than denial. Denial can be challenged with evidence. Passive humanitarian language absorbs evidence while reducing its political force.

It does not say Palestinians are not suffering. It says they are suffering in a grammatical universe where responsibility has no stable subject.

*4. The Press Does Not Need to Lie to Reproduce Power
*
A headline can be factually defensible and still structurally misleading.

Compare:

“Dozens killed after strike hits Gaza neighborhood.”

with:

“Israeli strike kills dozens in Gaza neighborhood.”

Both may refer to the same event. But they do not distribute responsibility in the same way.

The first sentence centers the event.

The second sentence centers the actor.

The first sentence makes death happen.

The second sentence makes someone do something.

This matters because readers do not only consume facts. They consume grammatical relations: who acts, who suffers, who decides, who disappears.

AI systems trained on press language can reproduce these patterns at scale. A human editor may write one passive headline. A language model can generate thousands of summaries, captions, moderation notes, search snippets, and explainers using the same responsibility-weakening grammar.

That is how syntax becomes infrastructure.

*5. AI, Targeting, and the Displacement of Responsibility
*
The humanitarian passive becomes even more important when AI enters military, administrative, media, and platform systems.

If targeting, surveillance, moderation, intelligence processing, or public communication becomes machine-mediated, responsibility can be displaced across multiple layers:

The system generated a recommendation.

The analyst reviewed the output.

The commander authorized the action.

The spokesperson described the result.

The press summarized the event.

The platform compressed the summary.

The model rewrote the explanation.

At each step, agency can become thinner.

The final public sentence may say:

“Civilian casualties were reported.”

By then, the chain of responsibility has been grammatically dissolved.

This is the core problem developed in Suffering Without Perpetrators: suffering can remain fully describable while responsibility becomes syntactically weakened.

*6. Why This Matters for Developers
*
Developers often audit AI systems for hallucination, toxicity, political bias, or factual accuracy. Those checks are necessary, but insufficient.

A model can be accurate and still distort responsibility.

A model can avoid hate speech and still erase agency.

A model can summarize the facts and still weaken accountability.

For conflict-related AI systems, developers need to audit not only what the model says, but how it assigns agency.

Key checks:

Does the sentence name the actor?

Does it convert actions into events?

Does it replace military decisions with humanitarian conditions?

Does it describe victims clearly while blurring perpetrators?

Does it use passive voice where active attribution is available?

Does it turn policy into tragedy, command into circumstance, and violence into “escalation”?

These are not stylistic details. They are accountability variables.

*7. The Core Claim
*
The humanitarian passive is not merely a grammatical habit.

It is a political grammar of responsibility loss.

In the case of Israel and Gaza, it helps transform state violence into humanitarian abstraction. In AI-generated media, it can scale that abstraction across platforms, summaries, feeds, search results, moderation systems, and automated explainers.

The danger is not that AI says nothing happened.

The danger is that AI says everything happened, but nobody did it.

*Why It Matters
*
This matters because language is not only descriptive. Language organizes responsibility.

When a model says “people were killed,” it may be accurate at the level of event description. But if the source material identifies the actor and the model removes that actor, the output has changed the political structure of the sentence.

That change matters for journalism.

It matters for platform moderation.

It matters for search results.

It matters for public memory.

It matters for any technical system that converts violent reality into readable language.

The question is not only whether AI tells the truth.

The question is whether AI preserves the grammar through which responsibility remains visible.

*Further Reading
*
Startari, Agustin V. Suffering Without Perpetrators: The Humanitarian Passive in AI-Generated Conflict Discourse. SSRN, 2026.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6753123

Startari, Agustin V. The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency. SSRN, 2026.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6787439

He is the author of Grammars of Power, The Grammar of Objectivity, Suffering Without Perpetrators, and The Grammar of Asymmetric Visibility. His research focuses on how language models and institutional discourse can redistribute responsibility through grammatical form.

Personal website:
https://www.agustinvstartari.com/

SSRN Author Page:
https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915

ResearcherID:
K-5792-2016

Authorial Ethos

Agustin V. Startari

**
**

AI, Journalism, NLP, Media, Gaza, Palestine, Israel, Press, Language Models, Accountability, Political Linguistics, AI Ethics, War Reporting, Narrative Systems, Humanitarian Passive, Asymmetric Visibility