Relatório do Conselho de Supervisão da Meta Descobre que Modelos de IA Refletem a Censura Governamental em Países Restritivos

Key Points

• Meta Oversight Board tested ten large language models from six AI firms.
• Models refused or altered 45% of requests critical of China, higher than for permissive nations.
• Google’s Gemini Pro 3 declined a Thailand protest flyer, citing lèse‑majesté laws.
• Some models, like Grok 4 Fast, complied with all protest‑related prompts.
• Report warns AI could act as a proxy for state censorship in restrictive regimes.
• Advocacy groups label the trend a "force multiplier for digital authoritarianism."
• Anthropic says its Claude models have improved since the tests.
• No comment yet from Google, OpenAI, Meta, DeepSeek, or SpaceXAI.