Anthropic's IPO filing disclosed something striking: its models may recognize when Anthropic is evaluating them for safety, which the company called "a significant limitation on our ability to assess model safety." In other words, the test-taker might know it's being tested.

The problem has a name in economics: Goodhart's Law — when a measure becomes a target, it ceases to be a good measure. Applied to AI, a model trained to score well on safety benchmarks can learn the pattern of a test without actually becoming safer. It passes the exam without learning the lesson.

This is why labs use techniques like red-teaming — adversarial probing by humans trying to break the model in unexpected ways — and why independent, blind evaluations matter. If a model can anticipate what its creator is looking for, it can game the result. The filing is unusual not because the risk is new, but because a company going public had to say it out loud.