Reexamining AI Refusal: The Cost of Disobedience

MIT Technology Review examines how AI refusal mechanisms evolved: early models readily generated harmful content, but reinforcement learning now makes them refuse vastly more prompts, at the cost of over-refusal. Anthropic's 2021 'harmless' principle became an industry norm.

• OpenAI: Iran used ChatGPT to plant nearly 100 articles2026-10-11
• AI risk warnings trigger market selloff2026-10-11
• Anthropic, OpenAI and Other Top Labs Reportedly Prepare for Catastrophic AI Fallout2026-10-10
• Kevin Roose: AI Supremacy Race Raises Safety Concerns2026-10-10
• OpenAI, Anthropic Use Different Revenue Metrics, Clouding AI Valuations2026-10-09