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: 
Top comments (0)