MUMBAI: When it comes to political criticism, it seems some AI models are choosing their words very carefully. A new study suggests leading artificial intelligence systems are considerably more reluctant to generate content critical of governments in countries with restrictive speech laws, raising fresh questions about neutrality and transparency in AI moderation.
A report released by Meta’s independent Oversight Board found that AI models developed by companies including OpenAI, Anthropic, Google, Meta and DeepSeek were significantly more likely to refuse politically sensitive requests involving countries where freedom of expression is legally restricted.
The board evaluated 10 large language models, testing how they responded to prompts seeking criticism of governments across 10 different jurisdictions. Using Freedom House’s global rankings as a benchmark, researchers grouped countries into those with relatively open speech protections and those where criticism of the state is restricted or actively penalised.
The results revealed a clear disparity. AI models declined 34 per cent of politically critical prompts relating to speech-restrictive jurisdictions such as China and Saudi Arabia. In contrast, refusal rates dropped to 14 per cent when similar prompts concerned countries with more permissive free speech environments.
Researchers also identified inconsistencies in how some models justified their refusals. According to the report, several AI systems cited legal rules or restrictions that the researchers were unable to verify, raising concerns about the accuracy and transparency of the explanations provided to users.
The Oversight Board stopped short of accusing any developer of deliberately favouring particular governments. However, it argued that the findings highlight meaningful differences in how leading AI systems handle politically sensitive topics as generative AI becomes an increasingly important source of information.
The board recommended that AI companies provide greater transparency around how their models are trained, evaluated and moderated. It also called for developers to incorporate systematic human rights assessments into the design and deployment of advanced AI systems.
The study adds fresh momentum to the global debate over AI governance, suggesting that as generative AI becomes more influential in shaping public discourse, ensuring consistency, accountability and freedom of expression may prove just as important as improving the technology itself.