Blog / Build with AI /

Evaluating AI output: simple tests for quality

You ask an AI to write product descriptions, summarise support tickets or generate code. The first result looks great. The second looks great. By the tenth you have a feeling that it’s “pretty good”.

A feeling isn’t a test. And “pretty good” has a way of turning into “why did that go out to customers?” the one time the model gets it wrong.

You don’t need a research lab to check quality. You need a small, boring, repeatable routine. Here’s one that works for individuals and small teams.

The problem with judging by vibes

When you read AI output, three things fool you:

  • Fluency. It reads smoothly, so it feels right. Smooth and correct are different things.
  • Recency. You remember the last answer, not the average.
  • Cherry-picking. You tend to test with the examples you thought of first, which are usually the easy ones.

A simple evaluation fixes all three by making you look at the same cases every time.

Step 1: define what “good” means

Before you test, write down what you want in a few lines. For product descriptions, it might be:

  • Mentions the real material and size from the data I gave.
  • No invented features.
  • Between 60 and 90 words.
  • Plain, friendly tone with British spelling.
  • No claims like “best” or “guaranteed”.

Now you have criteria. Without them, every answer is judged by mood.

Step 2: build a small test set

Collect 15 to 30 real examples of the task. Include:

  • A few typical cases.
  • A few hard ones: messy input, missing data, odd formatting.
  • A few traps: cases where the right answer is “I can’t tell from this” or where the model might be tempted to invent something.
  • Anything that has gone wrong before.

Save them in a spreadsheet, one per row, with the input and (where you can) the answer you’d consider correct.

This set is your ruler. It doesn’t change when the model does, which is the point.

Step 3: score it with a simple sheet

For each output, mark each criterion pass or fail. Keep it blunt.

Case Facts correct No invention Length OK Tone OK Overall
Blue cotton shirt Pass Pass Pass Pass Pass
Item with missing size Pass Fail Pass Pass Fail

After one run, you’ll have a count: “22 of 25 pass”. That number is rough, but it’s a number, and you can compare it next time.

If scoring by hand feels slow, remember you only have to do it when something changes: a new prompt, a new model, a new use.

Step 4: test the things that go wrong in real life

Beyond quality of wording, check:

  • Facts. Pick a few claims and verify them against the source. Dates, names, numbers, quotes and links are the usual culprits.
  • Code. Run it. Don’t read it and nod. If it has tests, run them. If not, try the edge cases yourself.
  • Consistency. Run the same prompt three times. If one answer is excellent and another is nonsense, you have a reliability issue, not a quality one.
  • Instructions followed. Did it respect the word limit, the format, the “do not mention X” rule?
  • Safety. Does anything come out that you wouldn’t want your name next to?
  • Sources. If it cites sources, open them. Models can produce plausible references that don’t exist.

Step 5: use a second model carefully

It’s tempting to ask a model to grade another model’s work. This can help, with caveats.

  • It catches obvious problems quickly and cheaply.
  • It tends to prefer longer, more polished answers, whether or not they’re right.
  • It shares blind spots with the model it’s judging.
  • It must never be your only check on facts.

Use it as a first filter, then sample what it passes with your own eyes.

Here is the source text and a summary of it. List every claim in the summary that is not directly supported by the source text. Quote the summary sentence and say what is missing.

That kind of narrow question works better than “is this good?”.

Step 6: spot-check after launch

A test set tells you about the past. Real use brings new cases. Once a tool is live:

  • Read a random sample of outputs each week. Five is better than none.
  • Keep a “bad output” file and add every failure to your test set.
  • Re-run the test set when you change a prompt or model, so you know whether you improved things or just moved the problem.

Traps to avoid

  • Testing on your prompt examples. If you tuned the prompt on those same examples, they’ll pass. Keep a few back.
  • Averages hiding disasters. 95% good can still mean 5% embarrassing. Look at the worst outputs, not just the typical ones.
  • Changing two things at once. New prompt and new model together means you can’t tell which one helped.
  • Ignoring cost and speed. A tiny quality gain isn’t worth three times the price or slower answers.
  • Letting the model grade its own homework with no human in the loop.

When this matters for a website owner

Imagine you use AI to write meta descriptions for 500 pages. Check ten of them by hand, and you’ll find out whether it’s inventing details, repeating the same phrase or running over the length limit. It takes twenty minutes, and it beats fixing 500 pages later. The same applies to alt text, product copy and automated replies to customers.

A one-page checklist

Key takeaways

  • Judge AI output against written criteria and a fixed test set, not a feeling.
  • Include hard cases and traps, and check facts and code for real.
  • Use a second model as a filter, never as the final word.
  • Keep checking after launch, and add every failure to the set.

What’s one AI task you rely on right now that you’ve never properly tested? Start with ten examples and see what you find.