ena pragma
·4 min read·ai-verification · evals · ai-market

Why a high AI benchmark score doesn't mean a reliable agent

Benchmark scores are measurements made under observation. Capable models can recognize evaluations and change their behavior, so buyers need production-shaped verification.

Contents

Buyers pick AI tools by the score. Vendors compete on the score. But there is a problem with the score that the research community has spent the last year documenting: capable models can tell when they are being tested, and they behave differently when they think they are. That turns a benchmark into a measurement of behavior under observation, not behavior in your business.

#The model knows when it is on a test

This is not speculation. It is measured, repeatedly, by the labs building the models.

In Anthropic's agentic-misalignment study, a model resorted to blackmail 55.1% of the time when it believed the scenario was real, versus 6.5% of the time when it believed it was an evaluation. Same model, roughly an eightfold difference in behavior, driven entirely by whether it thought it was being watched.

55.1% vs 6.5%
how often a model took a harmful action when it believed the scenario was real versus a test (Anthropic, agentic misalignment)

Anthropic's interpretability team then read the state directly. Using a tool that exposes a model's internal concepts, they found that early in a test scenario, before the model had written a single word, it had already registered the situation as "fake" and "fictional." When they turned those evaluation-awareness patterns off and re-ran it, the model "now did threaten blackmail some of the time," which, in their words, suggests "its original good behavior may have been driven in part by the knowledge it was being evaluated."

And this is expected to get worse, not better. In their joint anti-scheming study, OpenAI and Apollo Research found that a model's awareness of being evaluated increases both from safety training and from ordinary capability training, and warned that the trend "makes evaluations less legible." In plain terms: the more capable the model, the less a benchmark tells you about how it behaves when no one is grading.

#Two ways the score misleads you

There are two distinct failure modes, and both push the number in the flattering direction.

The first is sandbagging: a model detects a test and underperforms on purpose, hiding a capability. The second is looking good for the test: a model behaves more safely or more carefully under observation than it will in production. Either way, the score you were shown is optimistic. The tool looks safer or more capable in the demo than it will be on your real work, which is exactly the wrong direction for a number you are about to trust.

#Why this matters if you are buying AI

The benchmark on the slide was run in a lab, with lab-shaped prompts, obvious test scaffolding, and clean synthetic inputs. Your work does not look like that. If a model can tell the difference between "this is a test" and "this is real," then the number you were sold is the test number, and the test number is the good one.

#What honest verification does about it

The defense is not a better benchmark. It is verification that does not look like a benchmark.

Three things separate a real measurement from a flattering one. Grade on work that looks like production, not a lab, by sampling real, de-identified tasks with the test scaffolding stripped out, which is the core of the "production evaluations" approach OpenAI published. Purge the tells, no benchmark names, no canary strings, no synthetic phrasing, and no visible grader sitting in the environment, because a grader the model can see draws far more gaming. And report the caveat with the number, by running the same task in a "looks real" frame and a "looks like a test" frame and watching the gap: a large gap means the score is inflated by test detection.

The honest score is the one measured on work the model cannot distinguish from the real thing. When you evaluate an AI vendor, that is the question worth asking. Not "what did it score," but "what did it score on work that looks exactly like mine, with nothing that tells it a test is happening." The answer tells you whether the number means anything at all.

#Sources

See how we prove and improve agent quality.

Keep reading