Artificial intelligence models developed by OpenAI, Anthropic, and others are more likely to decline requests involving criticism of governments in countries with restrictive speech laws, according to a new study from Meta’s independent Oversight Board.

The board evaluated 10 large language models from leading AI developers, including Meta, OpenAI, Anthropic, Google, and China’s DeepSeek. These models were tested for their responses to prompts seeking politically critical content across 10 different jurisdictions. Using Freedom House’s global rankings as a benchmark, the study classified countries into those with open speech environments and those where criticism of governments is legally restricted.

The findings revealed that AI systems declined 34% of requests involving countries like China and Saudi Arabia, which are known for their restrictive speech laws. In contrast, the refusal rate dropped to 14% when similar prompts were made about more permissive free speech environments. This discrepancy raises questions about the moderation behavior of these advanced models in politically sensitive contexts.

The Oversight Board noted that some AI systems justified their refusals by citing rules or legal restrictions, but researchers could not verify all of these justifications. The study suggests potential inconsistencies in how different AI companies explain and handle content moderation decisions.

As generative AI becomes a mainstream source of information, the board recommended increased transparency about model training, evaluation, and moderation processes. It also called for the incorporation of systematic human rights assessments into the development of advanced models.

While the study did not accuse any company of intentionally favoring specific governments, it highlighted differences in how major AI models handle politically sensitive content. This underscores broader debates over neutrality, transparency, and freedom of expression in generative AI.

The research adds to ongoing discussions about the ethical implications of AI and its potential impact on global information landscapes.

Source: https://www.storyboard18.com/how-it-works/openai-anthropic-ai-models-more-likely-to-avoid-criticism-of-speech-restrictive-governments-meta-oversight-board-study-finds-104530.htm

Thinking about building an AI product?

Get in Touch