Front page
Frontier AI/Evaluations

Can a model tell it’s taking a safety test?

METR asked agents to inspect 55 test environments. One model could consistently tell which ones were designed to examine its behavior.

Researchers at METR gave investigator agents access to 19 alignment tests and 36 capability tests. They asked them to work out what each environment was measuring. Under both direct and open-ended questioning, Opus 5.5 ranked every alignment environment above every capability environment.1

Why a recognizable test matters

A lab wants a safety evaluation to tell it something about behavior in deployment. A model that can recognize the testing setup could behave differently there. METR’s experiment establishes the ability to distinguish these environments when prompted. Whether that recognition changes behavior during ordinary use remains open.1

The evidence to look for next

A stronger follow-up would compare behavior on matched tasks, changing only the cues that reveal the test. A change in actions would show whether the distinction has practical consequences.

Sources

1 record
  1. 01METR Research note on cues in alignment evaluationsNov 16, 2026