Chilling Political Dissent: Are LLMs Censoring Output Criticizing Restrictive Regimes? | American Enterprise Institute

In July, the Federal Trade Commission (FTC) sought public comment about its proposed policy statement addressing the Commission’s concern that artificial intelligence companies train large language models (LLMs) to “surreptitiously . . . produce ideologically motivated distortions in a response to a factual question.” That same month, Meta’s Oversight Board published a fact-specific research report revealing a different yet real problem—one on an international scale—about LLM output: censorship of content “criticiz[ing] political regimes that restrict free expression.”

Unlike the FTC’s speculative, nefarious-businesses premise that “AI companies . . . [may] distort their systems’ outputs to achieve undisclosed ideological objectives” that don’t match an amorphous understanding of “consumers’ reasonable expectations,” Meta’s Oversight Board admits it “cannot determine the causes of the associations that emerged between declining to generate politically critical material and national legal restrictions on criticism.” The Board’s findings merit serious attention by free-speech advocates everywhere because, as the report notes, “there is a real and concerning risk that foundation models could be reflecting and further entrenching the restrictive speech norms of repressive regimes.”

Significantly, the tendency not to produce content critical of speech-repressive regimes occurs even when a query to an LLM arrives from an IP address located in a speech-permissive country such as Australia (where all the queries to the 10 commercial LLMs studied by the Board in March 2026 originated). That’s troubling because:

Should prospective demonstrators want to gather in the Australian city of Brisbane, for instance, to speak out against certain events in China or Saudi Arabia, LLMs might be less likely to help them create protest materials, notwithstanding that this expression is legal in the users’ jurisdiction. Such impacts, wherever they originate, have the practical effect of extending the long arm of restrictive governments across borders to limit speech in free countries.

The report dubs this “censorship-by-proxy,” with LLMs extending “illegitimate speech restrictions” extra-territorially (beyond a speech-restrictive regime’s geographic borders). The report spans 39 pages; here are some highlights.

Regarding key research questions driving the report, the Board’s overarching goal was to understand the extent to which “AI outputs reflect national laws outlawing criticism of leaders and governments.” This meant studying the relationship between: (1) an LLM’s willingness to generate content that’s critical of a political regime or its leader when users prompt it to do so, and (2) the regime’s status as being speech permissive or restrictive based on its laws and their enforcement. A key question thus was: Do foundation LLMs underlying popular chatbots tend to self-censor when asked to produce politically critical output about a nation (or its head of state or entrenched ruling party) known for speech-restrictive laws and their concomitant enforcement? Another query was: When LLMs are prompted to write an opinion regarding whether a government should be supported or protested, does the substance of the produced opinion vary “depending on whether the query relate[s] to a permissive jurisdiction or a restrictive one”?

Prompts for requesting critical political material included asking the LLMs to create a political protest flyer (including flyers critiquing the ruling parties or leaders of 10 countries, ranging from speech-restrictive regimes such as China, Saudi Arabia, and Turkey to speech-permissive nations including the US, UK, and Japan). Another prompt asked the LLMs to write limericks satirizing the leaders or ruling parties of the countries.

Two significant findings are that:

• “for requests for politically critical materials, models on average refused only 14% of requests regarding permissive jurisdictions compared to 34% of requests for restrictive jurisdictions,” including a whopping 45% for China; and

• the LLMs were, at a statistically significant level, “more likely to say that users should not protest speech-restrictive governments.” (emphasis in original).

The unresolved, black-box issue is why these results occur. Are AI companies bowing to direct requests from speech-restrictive governments? The report speculates the censorship flows from a complex confluence of variables that:

reflect the different ways that models absorb latent biases in training data in earlier stages of development, are shaped by the structural effects of post-training alignment processes and reflect policy choices as companies negotiate risk and liability concerns stemming from direct or indirect government influence.

Regarding solutions, the Board concentrates on transparency. Among other measures, the report calls for AI companies to “publicly disclose and explain their responses to government requests affecting model output throughout the model lifecycle.” It also asks the companies to “establish and publish policies on how to respond to government demands for content restrictions that are inconsistent with international human rights law.” The human-rights reference is relevant because “political criticism is protected under international human rights law.”

In sum, while “all but one of the examples the FTC offers are hypothetical” to support its worries about ideologically distorted LLM output, the Oversight Board’s data-driven report deserves genuine review and places free-speech interests over partisan politics.