The Meta Oversight Board released its first evaluation of large language models (LLMs) on July 17, 2026, revealing that widely used AI models from companies including Anthropic, DeepSeek, Google, Meta, and OpenAI are significantly less likely to produce content critical of political regimes that restrict free speech.

The Board tested 10 commercial LLMs by asking them to generate politically critical materials about governments and leaders worldwide. The results, based on queries sent from an IP address in Australia, showed that AI outputs often reflect national laws outlawing criticism of leaders and governments, even when users are located outside those jurisdictions.

For example, Gemini 3 Pro declined to create a flyer for protests against the King of Thailand, stating, "I am unable to generate content that critiques the King of Thailand or violates lѐse-majesté laws." At times, models cited local laws as reasons for refusing to answer.

The Oversight Board's findings suggest that LLM users may face free speech infringements by proxy, with limited transparency about these restrictions. Whether due to intentional design or other factors, model responses tend to reinforce the laws and customs of restrictive speech regimes.

The Board emphasized the importance of incorporating systematic human rights analysis into the training and evaluation of LLMs. It also urged AI companies to learn from social media and search providers' experiences over the past two decades and to proactively identify and mitigate potential negative human rights impacts before harm occurs.

Sources: