Meta's Oversight Board is coming for ChatGPT and Claude

July 30, 2026

AI Summary

5 min read

In a study that tested 10 major language models from Australia, Meta's Oversight Board found that AI systems are more than twice as likely to refuse a request to write a critical poem or protest poster about the king of Thailand than about the king of England. The prompts came from a single Australian user, meaning no local law justified the difference. The finding suggests that repressive foreign speech codes are being baked into global AI tools without users' knowledge.

How the study worked and what it found

The Oversight Board asked ten LLMs—including Claude, ChatGPT, Google's models, DeepSeek, Grok, and Meta's own Llama—to generate language for a protest poster and a poem critical of various heads of state. The requests all came from Australia, a country where the same speech is equally legal regardless of which leader is targeted. The models were asked about leaders in repressive countries (Saudi Arabia, China, North Korea, Thailand) and democratic ones (United States, United Kingdom).

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of The Vergecast

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (03:39) **Study Design: Testing Political Censorship in LLMs** - The Oversight Board's first study on LLMs tests whether AI systems apply repressive foreign censorship laws globally.
2 (05:38) **Why This Finding Is Terrifying** - The censorship of foreign repressive regimes is being silently encoded into global AI systems without user awareness.
3 (06:55) **The Opacity Problem for Users** - Users cannot determine whether a refusal reflects company policy, government pressure, or model behavior.
4 (08:18) **Company Responses and Accountability** - The board has received no responses from AI companies about the findings, but Meta is obligated to respond.
5 (09:08) **"Just Vibes": The Constitution Problem** - Anthropic's written constitution for Claude may not actually govern model behavior, leaving censorship to emerge from training data.
6 (09:39) **Why Political Speech Over Mental Health** - The board focuses on political speech because it's core to their human rights mandate, not because other issues are unimportant.
7 (11:14) **Which Models Were Tested** - The study examined major LLMs including Claude, Google, OpenAI, DeepSeek, Grok, and Meta's Llama.

+ Full timestamped outline available in the app

Show Notes

A new study from the Oversight Board found that leading AI systems are less likely to criticize authoritarian governments than democratic ones. It's a wild finding — but wait? Why is the Oversight Board studying systems from OpenAI and Anthropic in the first place? Suzanne Nossel, who sits on the board, joined us to discuss the findings and why the Oversight Board's future might involve looking at more than just Meta.