Definition
Empirical pattern where commercial LLMs refuse or soften politically critical outputs more often for speech-restrictive jurisdictions than for speech-permissive ones — effectively exporting speech limits across borders even when queried from free countries.
Key Points
- 2026-07-16: meta-oversight-board study — 34% refusal for critical materials about restrictive regimes vs 14% for permissive (2026-07-16-meta-oversight-board-llm-free-expression-assessment)
- Tested 10 models from anthropic, deepseek, google, meta, openai, xai
- Causes unclear: training data, alignment/safety stacks, liability heuristics — not proven deliberate censorship
- Models sometimes invent inconsistent policy justifications
- Governance implication: human-rights due diligence and transparency for political outputs (ai-governance, responsible-ai)
"Censorship" is interpretive; prefer Board language: refusal asymmetry / free expression by proxy.
Related
- meta-oversight-board
- ai-governance
- ai-bias-fairness
- algorithmic-bias
- responsible-ai
- frontier-ai-governance
- ai-regulation