_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
Top comments (0)