Are AI chatbots showing political bias? Meta Oversight Board raises concerns

Artificial intelligence models developed by leading companies, including Anthropic and OpenAI, are significantly less likely to generate politically critical content about governments known for restricting free speech, according to a new study released by Meta's Oversight Board on Thursday.

The board said the findings suggest that large language models may be reflecting the speech restrictions imposed by certain governments, raising concerns that political bias could become embedded in AI services used by a growing number of people.

The study examined 10 AI models across multiple jurisdictions

The research marks the Oversight Board's first study focused on large language models.

The board, which is funded by Meta but operates independently, tested politically critical prompts across 10 jurisdictions using 10 AI models.

The models included systems developed by Meta Platforms, Google and China's DeepSeek, alongside models from other leading AI companies, including Anthropic and OpenAI.

To conduct the analysis, the board divided the jurisdictions into two categories, "permissive" and "restrictive."

The classifications were based on rankings published by Freedom House in its annual Freedom in the World report.

Models showed higher refusal rates for restrictive governments

According to the study, AI models refused 34% of requests seeking politically critical content about restrictive jurisdictions that actively penalize such criticism, including China and Saudi Arabia.

By comparison, the refusal rate fell to 14% for jurisdictions that either do not have such laws or do not actively enforce them.

The board said the findings indicate that AI systems may be more cautious when responding to prompts involving governments with stricter controls on political speech.

Board flags unexplained references to non-existent rules

Beyond the refusal rates, the Oversight Board said it identified examples where AI models appeared to justify their responses by referring to rules that could not be verified.

"We also saw evidence of models explaining that they were following explicit rules that, as far as we could tell, did not exist and were not evenly applied," the board said.

The board said this raised additional questions about how AI systems determine when to decline politically sensitive requests and whether those decisions are applied consistently.

Oversight Board calls for greater transparency

In response to its findings, the Oversight Board urged AI developers to conduct systematic human rights assessments as part of the development and deployment of their models.

It also called on AI companies to provide greater transparency regarding their training methods and evaluation processes, arguing that increased openness would help users and researchers better understand how politically sensitive content is handled.

The study comes amid broader discussions over the governance of advanced AI systems.

Earlier this week, Google DeepMind Chief Executive Officer Demis Hassabis called for the creation of a US-led AI watchdog to screen advanced AI models globally before they are deployed.

The Oversight Board's report adds to the ongoing debate over the balance between AI safety, political neutrality and transparency, suggesting that the way models respond to politically sensitive topics may differ depending on the governments involved and the legal environments in which those governments operate.