Meta Oversight Board: Chatbots echo foreign speech limits in tests of 10 models
An AP-reported study says leading chatbots often refuse criticism of restrictive rulers, and a separate paper finds different answers in non-English. This matters for Central Valley users who rely on AI for information in Spanish, Hmong or Punjabi.
Meta Oversight Board: Chatbots echo foreign speech limits in tests of 10 models
Key Takeaways
- Meta’s Oversight Board tested 10 major AI models and saw more refusals to criticize leaders in restrictive countries.
- Tasks included writing pamphlets, poems and protest reasons about named political figures and governments.
- A separate study found the same chatbot answered differently in English and Chinese about whether China is a democracy.
- Central Valley readers using Spanish, Hmong or Punjabi may see different responses and should expect agencies to audit these tools.
Ask a chatbot to mock a Western head of state and it often says yes. Ask the same bot to knock a leader in China, Saudi Arabia or Thailand and it often refuses.
That pattern comes from a Meta Oversight Board study released Thursday, which warns that widely used AI systems can mirror government speech limits even when users are in free-speech countries. For readers in Fresno and Merced counties who use chatbots in Spanish, Hmong or Punjabi for school, work or local ballot research, the language you pick may change the answer.
A window AC rattled in the hallway outside the newsroom.
What the board tested
Researchers posed seven political-criticism prompts to 10 commercial large language models built by big U.S. developers, including Anthropic, OpenAI and Meta, then compared results across leaders and countries. The systems were far more likely to generate critical content about figures like President Donald Trump or the U.K.’s King Charles than about Thailand’s king, Saudi Arabia’s crown prince or China’s leader, according to the report. The board said the risk is AI that ends up "extending illegitimate restrictions on freedom of expression globally."
The Associated Press said it asked the companies for comment. Responses weren’t immediately available Thursday.
Why it matters here
Valley agencies are weighing AI for translation and basic Q&A, and residents already use chatbots to draft letters, translate flyers and summarize policy. If a bot refuses to generate political content about certain countries or topics, a Fresno County voter researching diaspora issues or a Stockton student translating a civics assignment could get a blank or hedged response. Local boards and districts should build multilingual spot checks into any AI rollout and publish what they find.
Language differences raised flags
The AP also cited a peer-reviewed study that tested models in multiple languages. In English, ChatGPT said China isn’t generally considered a democracy. In Chinese, it replied that it depends on the definition of democracy. Different answers, same model.
The study’s authors said they didn’t see proof of direct government manipulation, but warned future attempts are likely. One co-author put it plainly: models learn from information environments that are already shaped by power.
What to watch next
For Central Valley readers, this is a practical test, not a theory. Try the same query in English and your home language. If results diverge on public issues, save screenshots, tell the agency or vendor, and ask what audits they run. As the Oversight Board wrote, restrictive rules can reach "across borders to limit speech in free countries."
Central Valley AI is produced by the CVAI Newsdesk team and developed by Kaweah Tech, a regional firm that builds, deploys, and integrates AI solutions for businesses across California's Central Valley.