A Meta Oversight Board study finds major AI systems are more likely to refuse or avoid criticizing restrictive leaders or governments, raising concerns that AI could spread government-imposed limits on free speech globally.

The findings suggest AI models may mirror latent biases from training data and developer risk assessments, potentially constraining free expression across borders.

The report warns that these outputs could practically extend restrictive governments’ reach into free countries by curbing speech.

Experts emphasize the need for human rights due diligence, data auditing, and multilingual safeguards by model developers to mitigate risks, while exact causes remain unsettled.

Some models justified refusals by citing rules or legal restrictions that researchers could not verify, signaling inconsistencies in how moderation decisions are explained.

The study notes differences in handling politically sensitive content across companies, fueling debates about neutrality, transparency, and freedom of expression in generative AI.

Major players like Anthropic and OpenAI did not respond to requests for comment on the studies cited.

The industry response notes that several companies were contacted for comment, but responses were not provided in the reporting.

Contextually, leaders like Google's DeepMind CEO have called for a U.S.-led AI watchdog to screen advanced models before deployment, tying governance debates to the new findings.

The findings come amid ongoing global discussions on AI governance, guardrails, and regulatory efforts balancing security with innovation.

The board urges AI companies to perform systematic human rights analyses and increase transparency in training, evaluation, and moderation processes to address biases and risks.

Tests across 10 models from major labs show higher willingness to criticize authorities in freer environments (e.g., Australia) than in highly restricted regimes, with geographic spillover effects.