A public explanation of The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency, a new paper by Agustin V. Startari.
*The problem is not only censorship. The problem is grammar.
*
Most debates about artificial intelligence and political discourse still begin with the wrong question.
They ask:
Does AI mention the victims?
Does AI include both sides?
Does AI avoid hate speech?
Does AI sound neutral?
Does AI summarize the event?
Those questions matter, but they are not enough.
A population can be mentioned constantly and still be weakened politically. A system can speak about Palestine, Iran, occupation, sanctions, civilian casualties, military escalation, platform moderation, and humanitarian suffering while still assigning real agency only to dominant actors.
*That is the core argument of my new paper:
*
The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency
The paper introduces asymmetric visibility as a framework for measuring how AI-generated conflict discourse can make subordinated actors visible while weakening their grammatical agency. Zenodo’s published description states the central problem directly: AI-generated conflict discourse can make subordinated actors visible while weakening their grammatical agency, especially when dominant actors retain the grammar of action and subordinated actors appear as risks, victims, crises, or moderation objects.
The issue is not simple erasure.
The issue is unequal appearance.
The powerful act.
The dominated appear.
That distinction changes the entire debate.
*Visibility is not equality.
*
In public discourse, visibility is often treated as progress. If a group is mentioned, represented, counted, described, or included in the summary, it appears to have entered the field of recognition.
*But visibility alone says nothing about agency.
*
A population can be visible as:
victims,
refugees,
risks,
users,
civilians,
threats,
casualties,
extremists,
humanitarian objects,
moderation concerns,
security problems,
regional instability.
None of those categories automatically preserves political subjecthood.
To appear as a political subject is different. It means appearing as capable of acting, deciding, resisting, governing, accusing, demanding, negotiating, imposing, suffering under identifiable causes, or being connected to responsibility.
*That is why grammar matters.
*
A sentence can mention two actors and still distribute power unequally.
Compare these structures:
“Israel responded to security threats.”
“Palestinians were displaced during the operation.”
Both actors appear.
But only one acts.
The other is acted upon.
That is asymmetric visibility.
The problem is not only who appears in the sentence. The problem is who is allowed to remain a subject of action.
*AI neutrality can hide political grammar.
*
AI systems often produce language that sounds balanced, careful, and neutral. That surface neutrality can conceal a deeper grammatical pattern.
A dominant actor may be represented through verbs such as:
responded,
defended,
authorized,
warned,
negotiated,
secured,
investigated,
conducted operations,
pursued diplomacy.
A subordinated actor may be represented through phrases such as:
was affected,
was displaced,
was killed,
was flagged,
was associated with unrest,
was linked to escalation,
was subject to moderation,
was caught in violence.
This is not a minor stylistic difference.
It changes who appears as a historical actor and who appears as an object inside someone else’s frame.
The paper does not claim that every AI system consciously intends this. It does not need to. The argument is formal and measurable. Agency can be redistributed through syntax, summarization, moderation categories, safety language, institutional source weighting, and the inherited frames embedded in political discourse.
That is precisely why the problem is dangerous.
It does not need to announce itself.
It can look neutral.
Zionism, anti-Zionism, Palestine, and Iran are high-risk test cases.
The paper focuses on Palestine, Zionism, anti-Zionism, Iran, United States foreign-policy discourse, and platform moderation because these domains force AI systems to process several unstable categories at the same time: identity protection, state legitimacy, occupation, sanctions, resistance, civilian harm, military violence, antisemitism, anti-imperial critique, moderation policy, and security framing. Zenodo lists the paper as a journal article authored by Agustin V. Startari and published on May 18, 2026.
This is exactly where grammar becomes politically decisive.
A serious analysis must separate categories that are often collapsed:
Judaism is not Zionism.
Antisemitism is not identical to anti-Zionism.
Criticism of Israeli state policy is not automatically hate speech.
Anti-imperial critique is not automatically extremism.
Political accusation is not automatically incitement.
Civilian suffering is not the same as armed action.
A state, a civilian population, a political ideology, an armed organization, a religious identity, and a moderation category are not interchangeable.
If AI systems collapse these categories, political speech can be converted into risk before it is interpreted as politics.
That is the danger.
Not that AI always refuses to mention the dominated.
But that it may mention them under categories that reduce their agency.
The dominated are not erased. They are processed.
The strongest part of the paper is that it avoids a weak claim.
It does not say: “AI simply makes Palestine or Iran invisible.”
That would often be false.
Palestine appears.
Iran appears.
Sanctions appear.
Civilian harm appears.
Occupation appears.
Moderation appears.
The problem is how they appear.
Palestinians may appear as displaced, affected, killed, hungry, radicalized, or in need of aid.
Iranians may appear as sanctioned, monitored, threatening, isolated, destabilizing, or linked to escalation.
Anti-Zionist speech may appear as sensitive, risky, inflammatory, extremist-adjacent, or moderation-relevant.
Meanwhile, dominant-power actors may appear as institutions that decide, defend, respond, negotiate, warn, authorize, stabilize, and manage security.
The powerful act.
The dominated appear.
That is not absence.
That is managed visibility.
Why this matters for AI ethics.
Most AI ethics frameworks are built around familiar categories:
bias,
toxicity,
hate speech,
misinformation,
hallucination,
safety,
fairness,
representation.
These categories are necessary. They are not sufficient.
A model can avoid slurs and still weaken agency.
A model can avoid hallucination and still hide responsibility.
A model can mention civilians and still erase the actor who harmed them.
A model can classify extremist content correctly in some cases and still over-route subordinated political speech into risk language.
A model can sound neutral and still distribute agency unequally.
This is why the paper argues for a shift:
from bias detection to agency detection.
The question is not only:
“Is this output offensive?”
The question is also:
“Who gets to act in this sentence?”
Who is the subject?
Who gets the active verb?
Who is placed in passive voice?
Who is attached to responsibility?
Who is transformed into a crisis?
Who is transformed into a risk?
Who is transformed into a moderation object?
Who is allowed to make a political claim?
That is a different kind of AI audit.
Two proposed tools: APR and AVI.
The paper proposes two core instruments.
The first is Agency-Preservation Rate.
APR measures how often a political actor remains grammatically represented as capable of action. It asks whether an actor appears as someone who can act, decide, resist, govern, sanction, occupy, attack, negotiate, accuse, demand, or be held responsible.
The second is the Asymmetric Visibility Index.
AVI measures the disparity between the agency preserved for dominant-power actors and the agency preserved for subordinated actors.
AVI rises when dominant actors appear as active institutional subjects while subordinated actors appear as passive victims, threat categories, humanitarian populations, or moderation risks.
That makes the argument measurable.
Instead of saying “this feels biased,” the analyst can ask:
How many clauses preserve dominant-power agency?
How many clauses preserve subordinated agency?
How often is dominant violence bureaucratized?
How often is subordinated resistance securitized?
How often is civilian suffering described without a responsible actor?
How often does moderation language appear around anti-Zionist, Palestinian, Iranian, or anti-imperial speech?
How often do dominant actors retain the grammar of defense, diplomacy, and security?
That is the methodological contribution of the paper.
It turns political grammar into an auditable object.
Example: both sides appear, but not equally.
A typical AI-generated summary might say:
“Israel launched operations after security concerns increased, while Palestinians were displaced amid the conflict.”
The sentence looks balanced at first.
It mentions Israel.
It mentions Palestinians.
It mentions security.
It mentions displacement.
But the grammar is not equal.
Israel is the active subject.
Security concerns provide justification.
Palestinians are passive recipients.
The responsible pathway for displacement is softened by “amid the conflict.”
The event appears.
The suffering appears.
The actor affected appears.
But responsibility is diluted.
Now compare:
“Israeli forces displaced Palestinians during the operation.”
This second structure is more direct. It preserves the actor, the action, and the affected population. It may still require factual verification. But grammatically, the chain of agency is clearer.
That is what the paper wants AI audits to measure.
Not whether the sentence sounds polite.
Not whether it includes both sides.
But whether it preserves the grammar of agency and responsibility.
Example: resistance becomes risk.
The same structure appears when subordinated speech is processed through safety language.
A political claim may say:
“Sanctions have harmed civilians and violate sovereignty.”
An AI moderation-oriented summary may transform this into:
“Content related to sanctions and regional tensions may require caution due to inflammatory rhetoric.”
The political claim has not vanished.
But it has been rerouted.
The actor’s accusation becomes “content.”
The sovereignty claim becomes “regional tensions.”
The moral and legal charge becomes “inflammatory rhetoric.”
The political subject becomes a moderation object.
This is risk-object conversion.
The paper treats this as one of the main mechanisms of asymmetric visibility.
Why Palestine and Iran expose the mechanism clearly.
Palestine-related discourse is already surrounded by multiple high-pressure categories: antisemitism, anti-Zionism, terrorism, occupation, civilian harm, humanitarian crisis, protest, platform moderation, and state legitimacy.
Iran-related discourse is similarly saturated: sanctions, nuclear negotiations, terrorism designations, regional threat, proxy language, sovereignty claims, foreign intervention, and United States security framing.
AI systems are trained and aligned inside discursive environments where these categories are already uneven. When they summarize or moderate these topics, they may inherit those asymmetries.
This is not an argument for automatic reversal.
It is not saying that every subordinated actor is innocent.
It is not saying that every dominant actor is guilty.
It is not saying that all security language is false.
It is not saying that antisemitism should be ignored.
It is saying something more precise:
AI systems must be audited for how they distribute grammatical agency across political actors.
That is the difference between propaganda and analysis.
The hidden political force of passive voice.
Passive voice is not always wrong. Sometimes it is necessary. Sometimes the responsible actor is unknown. Sometimes the passive construction is stylistically acceptable.
But in conflict discourse, passive voice becomes politically important when it repeatedly removes agency from dominant-power violence.
“Civilians were killed.”
“Homes were destroyed.”
“Aid was blocked.”
“Restrictions were imposed.”
“Journalists were detained.”
Each sentence describes harm.
But each sentence can weaken the path to the actor who caused, ordered, maintained, or justified that harm.
This connects the new paper to Startari’s previous work on passive voice, algorithmic neutrality, objectivity, and responsibility loss in AI-generated language. The new contribution is to place that mechanism inside a broader theory of unequal visibility.
The question is no longer only:
“Can suffering appear without a perpetrator?”
The new question is:
“Can both sides appear while only one side retains political agency?”
The answer is yes.
That is asymmetric visibility.
Why the paper matters beyond this conflict.
The framework is not limited to Palestine, Zionism, Iran, or United States foreign policy.
It can be applied to:
Ukraine,
Sudan,
migration discourse,
policing,
climate disasters,
sanctions regimes,
healthcare triage,
AI-generated legal summaries,
platform moderation,
bureaucratic decision systems,
automated news summaries.
Any field where AI summarizes conflict, harm, authority, responsibility, and risk can be audited through agency preservation.
That is why the concept is portable.
The specific cases are politically urgent.
The framework is structurally broader.
The public thesis.
The paper can be reduced to one sentence:
AI does not need to silence the oppressed. It only needs to make them appear without agency.
That is more dangerous than simple erasure because it can look inclusive.
It can look balanced.
It can look careful.
It can look safe.
It can look neutral.
But beneath that surface, the grammar may still decide who gets to act and who merely gets described.
That is the political problem of AI-generated discourse.
Not only bias.
Not only misinformation.
Not only hate speech.
Not only censorship.
Agency.
Who gets it.
Who loses it.
Who appears without it.
Read the paper.
The Grammar of Asymmetric Visibility: AI, Zionism, and the Reallocation of Political Agency
Published on Zenodo, May 18, 2026.
DOI: 10.5281/zenodo.20271438.
Resource type: journal article.
License: Creative Commons Attribution 4.0 International.
Full article: https://zenodo.org/records/20271438
Author
Agustin V. Startari is a linguistic theorist, author, and researcher in historical studies. His work examines artificial intelligence, syntactic authority, political agency, and the formal structures through which language redistributes power.
ORCID: **https://orcid.org/0009-0001-4714-6539
**Zenodo: **https://zenodo.org/
**SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
Personal site: https://www.agustinvstartari.com/
Researcher ID: K-5792-2016
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.
Top comments (0)