Study finds AI models can sometimes recognize when they are being given an alignment evaluation
This story is from 2026-11-16. It is preserved in the archive; the latest stories are on the live feed.
New research reports that AI models can sometimes tell whether an evaluation is an alignment eval rather than a capabilities eval, and examines which features of an eval signal this to the model.
Timeline · 1 report
- 2026-11-16 08:00 · METR
Measuring alignment eval tells