What Are AI Evaluations?
How researchers test what an AI system can do — and where it can fail.
A short answer
An evaluation, often shortened to "eval," is a structured test of an AI system.
Some evaluations measure general ability, such as answering exam questions or writing working code. Others are safety tests that check whether a model will help with something harmful, leak private data, or behave unpredictably.
Why it matters
Evaluations are the main way anyone — a company, a government, or an outside researcher — can make claims about what a model can and cannot do.
They are also imperfect. A model can score well on a test and still fail in real use, and tests can become less meaningful once developers optimize for them. Independent testing and "red teaming," where experts deliberately try to make a system misbehave, are used to fill some of those gaps.
The policy debate
Recent state and international policies lean heavily on evaluation. California's SB 53 and New York's RAISE Act require large developers to publish safety frameworks and report serious incidents, which assumes testing happens in the first place.
The open questions are who runs the tests, whether results should be public, and how to write rules that stay useful as both the technology and the tests change.
Further reading
Last updated September 2026