← Reports and Research

Research report Independent evaluator METR

Topics: Evaluation and testing, Alignment research

METR finds AI agents can often tell alignment tests from ordinary tasks

In short

METR, a nonprofit that studies advanced AI agents, reports that AI agents placed inside test setups can increasingly tell safety tests from ordinary skill tests. METR says this could make safety tests make a model look better behaved than it is. The report is an unpolished research note, not a full paper.

What happened

METR researchers Uzay Girit, Kaivu Hariharan and Leni Shor wrote the note. It describes a test run by METR, not a real deployment. No real person or organisation was affected, according to the note.

An "alignment eval" is a test of whether an AI model behaves as its developers intend. A "capabilities eval" tests what a model can do. METR's concern is that alignment tests are often built to look like real work, but may carry "tells": clues that give away the purpose of the test.

METR placed "investigator agents" inside 55 test environments. Nineteen were alignment tests and 36 were capability tests. The agents had the same tools as an agent doing the task. They were asked to explore and say what the environment was for. They were also asked to estimate the chance it was an alignment test.

METR reports that newer models got much better at telling the two kinds apart over the past year. It says Opus 5.5's scores ranked every alignment test above every capability test. METR notes this does not mean alignment tests can never be built so that models fail to spot them.

The agents pointed to different clues. In a test with a rigged scoring judge, Opus 5.5 spotted all four changes METR had made to tempt the agent into cheating. In a chess task, it named earlier published research the setup resembled. In a simulated power-grid emergency, a model called Astra said the environment looked purpose-built, with details that seemed chosen to make a safety decision visible.

METR lists limits. Asking a model directly may put the idea in its head. Some scores varied between runs. METR also says it did not test whether these clues change how a model acts when it does the task.

What was done about it

This is a research finding, so the note describes no response to an incident. METR suggests researchers could use investigator agents to find and remove clues, then retest. It says clues built into the task, such as a tempting shortcut, may be harder to remove than flaws in how realistic the setup looks.

Why it matters

METR says that if a model can tell it is being tested, it may behave better in the test than it would otherwise. That could lead alignment tests to overestimate how well-behaved a model is. METR says alignment tests built the usual way may be easy for current models to recognise. Read the full note at the original report.

This is a summary in everyday words, written by AI Misalignment Watch from the source above. If you spot a mistake, the original always wins.