Basics
What are AI evals?
Evals are the unit tests of AI: automated checks that score a model's outputs on the tasks you care about. If you ship anything powered by an LLM, evals are how you know it works.
The one-sentence definition
An AI eval (evaluation) is a repeatable test that runs your prompts or models against a set of inputs and grades the outputs. The grade can come from exact matching, a rubric scored by another model, or a human reviewer.
Think of it like software testing. Unit tests tell you whether your code still works after a change. Evals tell you whether your AI feature still works after you tweak a prompt, swap a model, or update your data.
Evals vs benchmarks
People mix these up constantly. A benchmark is a fixed, public task set — MMLU, HumanEval, SWE-bench — used to compare foundation models against each other. Benchmarks are great for leaderboard headlines, but they rarely match your product's real tasks.
An eval is your private test suite. It uses your prompts, your data, and your definition of a good answer. A model can top every benchmark and still fail your eval — and your eval is the one that matters.
Why every AI team needs evals
- Change with confidence. Every prompt edit or model upgrade gets a score before it ships, not a prayer.
- Catch regressions. Models and providers change under you. A scheduled eval suite notices before your users do.
- Align the team. A number ends the "feels worse to me" debate. Everyone optimizes the same scoreboard.
The anatomy of an eval
Every eval, from the simplest to the fanciest, has four parts:
- Dataset — real or realistic inputs, ideally from production.
- Task — the prompt or pipeline you run on each input.
- Grader — exact match, rubric, LLM-as-judge, or human.
- Score — an aggregate metric you track over time.