AI safety tests need someone ready to act on the result
Interviews with 37 experts show why AI safety tests matter and where they may fall short. Tests can make a risk easier to see and help set a company commitment. But a score does not force a response. These are expert views, not proof that testing has cut harm.
What the source reports
Lily Stelling and coauthors studied how experts think tests of advanced AI systems might help govern risk. They coded 37 interviews. Experts described tests as a way to make risks visible, tie companies to promises, and build a culture that takes risk seriously. They also doubted how much the tests reveal. Some questioned whether results would lead to action or leave too much choice with developers. Four experts raised a concern about tests becoming a way to grant permission, rather than set limits. These interviews record what experts think; they do not test whether safety checks reduce harm in society. A test is more useful for a decision when someone owns the response before the result arrives.
Original source
- Title
- The Promises and Limitations of Evaluation-Based AI Governance: Evidence from Expert Interviews
- Author
- Lily Stelling, Jacob Charnock, Kevin L. Wei, David Krueger
- Publication
- arXiv cs.CY
- Date
- Tuesday, November 3, 2026