We’re putting too much faith in AI’s ability to say no

What happened: A MIT Technology Review piece argues that AI refusal training has become the load-bearing wall of AI safety, yet the same models retain the harmful capabilities they are trained to decline. Why it matters: The article says refusal lines are drawn secretly by AI companies and soon by governments, so the same mechanism that blocks harm could also stifle legitimate speech. Key numbers: Anthropic's 2021 helpful, honest, harmless framing; OpenAI red-teaming ahead of ChatGPT in 2022.

A MIT Technology Review piece argues that AI refusal training has become the load-bearing wall of AI safety, yet the same models retain the harmful capabilities they are trained to decline.

The article says refusal lines are drawn secretly by AI companies and soon by governments, so the same mechanism that blocks harm could also stifle legitimate speech. Key numbers: Anthropic's 2021 helpful, honest, harmless framing; OpenAI red-teaming ahead of ChatGPT in 2022.