An Assessment of How AI Models Protect Free Expression (www.oversightboard.com)

A recent evaluation by the Oversight Board has raised critical concerns about how large language models (LLMs) like those from Anthropic, Google, and OpenAI inadvertently stifle political speech. The study, which tested ten commercial LLMs, revealed that these models are more than twice as likely to refuse to produce politically critical content about repressive regimes compared to permissive ones. This dynamic suggests that LLM users may experience infringements on free expression, reflecting the restrictive laws inherent in certain jurisdictions. The research underscores the necessity for AI developers to incorporate human rights considerations into the training and assessment of LLMs, addressing the opaque influence of national laws on model outputs. The implications of these findings are significant, as LLMs increasingly shape public discourse across various applications. The models' patterns of refusal—often accompanied by vague justifications—may perpetuate censorship-by-proxy, extending the reach of illegitimate restrictions on free speech beyond their original context. This situation calls for urgent attention from the AI community to enhance transparency and accountability, learning from past mistakes of social media companies. By implementing robust human rights due diligence and clear policies on government content restrictions, AI developers can mitigate the risks associated with reinforcing oppressive norms, ensuring models serve as facilitators of free expression rather than barriers.