AI can refuse a request. Who sets the rules?
An MIT Technology Review investigation looks at how AI systems learn to refuse some requests. The safeguards are imperfect. The story also asks who decides which questions are off limits, since a refusal can curb valid inquiry. It makes a case about oversight, not a measured failure rate for all AI tools.
What the source reports
Arthur Holland Michel reports for MIT Technology Review on the rules that make AI systems say no to some requests. Training a system to refuse harmful requests is one goal. But the report describes limits to that approach and asks who gets to draw the line. A rule aimed at safety can also block a valid question. That makes the choices behind refusals a matter for public debate, not just a technical setting. The article does not measure a failure rate across all AI systems. Its main point is about how people set and oversee the boundaries.
Original source
- Title
- We’re putting too much faith in AI’s ability to say no
- Author
- Arthur Holland Michel
- Publication
- MIT Technology Review
- Date
- Friday, October 9, 2026