Meta's Oversight Board finds AI models criticize restrictive governments less
Meta's Oversight Board just found that popular AI models, like those from OpenAI and Anthropic, are much less likely to criticize governments with tough free speech laws.
Their study checked how these AIs responded in 10 different countries, ranked as either "restrictive" or "permissive" by Freedom House.
The results have people worried that these tools might be quietly favoring censorship.
AI rejects criticism in restrictive countries
AI systems turned down 34% of requests for political criticism about "restrictive" places like China and Saudi Arabia, but only 14% in more open countries.
Sometimes, AIs even cited rules that do not exist or are not enforced evenly.
In response, Meta's board is urging AI companies to check for human rights risks and be more transparent about how their models are trained.
There is also talk from Google DeepMind's CEO about creating a U.S.-led watchdog to keep advanced AI in check before it goes live.