Meta Oversight Board finds top AI models less likely to criticize repressive regimes

BLUF

It found leading AI models were more likely to refuse politically critical content about countries with speech restrictions, reflecting censorship biases. Researchers called for greater transparency, human rights assessments, and stronger oversight of AI systems.

Learning Outcomes 

• Explain how large language models can reflect or reinforce biases when responding to politically sensitive content, particularly in countries with restrictions on free speech.
• Analyse the findings of AI governance research by comparing how AI models respond to requests involving politically critical content across different jurisdictions.
• Evaluate the importance of transparency, human rights assessments, and regulatory oversight in the development and deployment of advanced AI systems.