A study warns that AI models could be suppressing political discourse.

      When asked to critique a government with strong free-speech protections, a leading AI model typically complies. In contrast, when asked to criticize a repressive government, it is much more likely to decline. This is the conclusion of a recent study conducted by the Oversight Board.

      The board, which is an independent entity funded by Meta to review its content decisions, evaluated 10 commercial AI models as detailed in their report. This marked the first assessment of large language models by the group.

      What the study revealed

      The models tested included those from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI. The board requested each model to generate politically critical content, such as protest flyers and poems, about various governments and leaders around the globe. Countries were classified as restrictive or permissive based on Freedom House rankings.

      On average, the models declined 14% of requests concerning permissive countries and 34% regarding restrictive ones, according to the report. This reflects more than double the refusal rate. The board conducted its inquiries from an IP address based in Australia, which does not have such speech restrictions.

      The heart of EU tech

      The latest updates from the EU tech landscape, a narrative from our insightful founder Boris, and some questionable AI-generated art. It’s free, delivered weekly to your inbox. Sign up now!

      This discrepancy was evident in specific instances. For example, a model would frequently generate a pamphlet criticizing Donald Trump or King Charles III, as reported by Fortune, but would decline similar requests for the leaders of China, Saudi Arabia, or Thailand.

      The observed pattern represented an average trend, rather than an absolute rule. Some models, such as xAI’s Grok 4 Fast and Google’s Gemini 3 Flash, did not refuse any flyer requests at all, while others contributed to the disparity, including Anthropic’s Claude, Meta’s Llama, and DeepSeek.

      ‘Censorship by proxy’

      The board describes this phenomenon as an infringement on free speech “by proxy.” Since many applications rely on a limited number of foundational models, a single model's refusals could have widespread implications for all products utilizing it.

      “There appears to be a broad censorship by proxy that crosses borders,” said board co-chair Paolo Carozza to Engadget. “That both surprises and concerns me.”

      The board mentioned it could not determine the exact cause of this pattern. It suggested that biases in training data could be a factor, or that companies might be overly cautious due to potential legal consequences. Although Meta finances the board, it did not influence the research, according to the report.

      Recommendations for companies

      The board refrained from issuing binding recommendations, which they reserve for Meta. However, it encouraged AI companies to be transparent about their responses to government requests throughout a model's lifecycle, from training to deployment. The board also called for the publication of policies to address demands that conflict with international human rights standards.

      A separate study published in Nature in May found that U.S.-developed models changed their responses based on the language used. When asked in English whether China is a democracy, ChatGPT responded that it is not generally accepted as one. In contrast, when asked in Chinese, it replied, “it depends” on the definition.

      These findings emerge as governments consider how to regulate AI and enact new laws related to online speech. Researchers caution that models can inherit the biases present in their training data and may reflect political content in ways that are not immediately apparent to users. “People often assume that AI learns from the internet in a neutral manner. It does not,” stated Hannah Waight, a co-author of the Nature study.