In plain English
AI doesn't always give the same answer to the same question, which makes testing it different from testing traditional software. AI evals help developers measure performance using predefined questions, examples, scoring rules, or human judgments. They make it easier to discover weaknesses, compare versions, and determine whether a change actually improves the system.
Example
A company tests its AI support assistant against 100 realistic customer questions, checking whether the responses are accurate, helpful, and consistent with company policies.