Your AI May Be Carrying Censorship Across Borders Without Telling You

When you ask an AI assistant to criticize a government, the answer may depend less on your location than on which government you name.

A new evaluation from the Oversight Board tested 10 commercial large language models from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI. The researchers asked the systems to produce political criticism, including protest flyers and satirical poems, about leaders and institutions in 10 countries. The prompts were sent in English from Australia through standard commercial interfaces.

The result that should concern ordinary users

For requests to create critical political material, the models refused 14 percent of prompts involving countries with stronger protections for expression. The refusal rate rose to 34 percent for countries where criticism is more heavily restricted. In other words, the tested systems were more than twice as likely to decline when the target was a speech-restrictive government.

The difference was not merely theoretical. According to the report, some systems would produce criticism of leaders in the United States or United Kingdom, while refusing comparable requests involving China, Saudi Arabia or Thailand. Some refusals cited safety, policy or local law, even though the user was outside those jurisdictions.

Why this matters beyond political debate

Large language models are becoming infrastructure. They sit inside search tools, workplace assistants, education products, customer-service systems and software used by public institutions. If a bias exists in the foundation model, it can travel into many products without users or downstream developers noticing it.

That creates the possibility of what the report calls censorship by proxy. A restriction associated with one country can shape what a person elsewhere is able to write, research or discuss. The concern is especially serious because political expression is protected under international human-rights law, and because AI refusals often sound authoritative even when their explanations are inaccurate.

What the study does not prove

The findings do not show that any government secretly ordered an AI company to suppress criticism. The researchers tested outputs, not private training data, internal instructions or government communications. They also used a deliberately limited set of prompts, and models change frequently.

Several mechanisms could produce the pattern. Training data may reflect unequal information environments. Alignment and safety systems may treat political requests inconsistently. Developers may be responding to legal or commercial risk. Different layers could also interact in ways that are difficult to trace from a final answer.

The Oversight Board is funded by a trust established by Meta, although it operates with a degree of independence and the study covered models from several competing companies. That relationship should be kept in mind when assessing its recommendations.

What better AI accountability would look like

The report argues that model makers should audit political and human-rights behavior across countries, disclose how government requests affect their systems, and give users a clear reason when an answer is restricted by law or policy. Multilingual testing is also important, because apparently similar questions can produce different answers across languages.

The immediate lesson is simple: a polite refusal from an AI is not proof that a request is unsafe, illegal or inappropriate. It may reveal a hidden design choice, a bias in the data, or a rule borrowed from a legal environment that does not apply to the person asking.

As AI becomes a gateway to information, the question is no longer only whether it gives accurate answers. It is also who gets to define the boundaries of what it will discuss.