Concept guide · How AI works
Good-looking is not the same as correct.
Fluent writing is a presentation skill, not proof the answer is right. The habit that catches a confident wrong answer is an eval: a small, fixed set of test cases, each with a known right answer, that you re-run whenever anything changes.
What it actually is
OpenAI's documentation defines it simply: "Evaluations (often called evals) test model outputs to ensure they meet style and content criteria that you specify." A grader decides whether each output is correct, and the run is tallied as passed, failed or errored. It is the same idea as marking against an answer key.
A normal example
You use AI to pull the total from an invoice. You gather ten real past invoices where you already know the ten correct totals. Nine answers look perfect. On invoice seven the total sits in a footer, and the model returns the subtotal instead. Confident, well-formatted, wrong, and caught only because you held the answer key.
The proof: ten cases where confidence fails
A single impressive result proves nothing. A fixed set of ten, scored pass or fail against known answers, tells you how often the tool is actually right on your work, and shows you exactly which case it fails. That is the inspectable object: ten rows, each marked pass or fail, that you keep and re-run.
Try it now
- Pick one repeated task where the answer can be checked.
- Collect ten real past examples where you know the right result, including the two awkward ones that have caught you out before.
- Run all ten and mark each pass or fail against the known answer.
- Keep the set and re-run it after any change of prompt, model or tool.
The decision this gives you
Build the smallest eval that would catch your worst failure, and run it before you trust a change. It is the difference between "this looked fine the once I tried it" and "this passes the ten cases that matter, including the two that used to break."
Judging quality by asking the model "are you sure?" or by how polished the writing is. Both measure confidence, not correctness. A model will restate a wrong answer just as fluently the second time.
The human bit
You define what "correct" means for your work, and you own the answer key. A model cannot grade itself into being right, because it is the thing being tested.
This explainer is anchored to the concept, a small pass-or-fail set anyone can keep, not to any one vendor's dashboard. OpenAI's documentation states it "is deprecating the Evals platform," which becomes read-only for existing users on 31 October 2026 and is "scheduled to shut down on November 30, 2026." The habit outlasts the product. A spreadsheet of ten cases and their right answers is an eval.
Official source
The deprecation dates above were checked against OpenAI's notice on 19 August 2026.