DEV Community

Cover image for Same Answer, Two Verdicts: Why QA Teams Disagree About AI Output
Secondverdict
Secondverdict

Posted on

Same Answer, Two Verdicts: Why QA Teams Disagree About AI Output

A user takes a photo of chicken curry with rice. The model returns "curry". Is that a pass or a fail?

One tester says pass: the dish is a curry. Another says fail: the protein is missing, the rice is missing and the calorie estimate will be wrong. Both are careful people looking at the same output. They disagree because they are answering different questions.

Is curry correct? Right at the dish name level, wrong at ingredients and nutrition

This is the most underrated problem in testing AI features. We spend a lot of effort on datasets and metrics, and very little on the rules that turn an output into a verdict. When two teams grade the same model, the gap between their pass rates often says more about their rules than about the model.

Where disagreements come from

Three sources of grading disagreement

1. The unit of correctness is not defined. Is the model graded on the dish name, on each ingredient or on the nutrition numbers the user finally sees? "Curry" can be right at one level and wrong at another.

2. Partial answers have no agreed status. Classic testing is binary. AI output is not. If the rules do not say what to do with "almost right", every tester invents a private threshold.

3. The expected result is an opinion. For a login form, the expected result comes from a spec. For a food photo, it comes from a person looking at the picture. If two graders label the ground truth differently, the model fails one of them by definition.

A case from my own work

I test a food recognition feature, and one disagreement taught me more than any metric: duplicates.

Imagine a photo with two different sandwiches. The model returns two items, both named "sandwich". To one grader this is an obvious defect: the same item appears twice. To me it is not. If the two items have different ingredient lists, they are two different items that happen to share a name. The model saw two things on the plate and described both.

Neither reading is careless. The word "duplicate" was simply never defined. One side compared names, the other compared content. Until that definition is written down, the same output will keep producing two verdicts, and no amount of re-testing will close the gap.

Two sandwiches: one grader sees a duplicate, the other sees two items

What actually helps

Four steps to calibrate grading before a test round

  • Write the grading rules down before the test round, not after.
  • Include ten to twenty borderline examples with the agreed verdict and the reason.
  • Have both teams grade the same small sample independently, then compare only the cases where they differ. Those cases are your real specification.
  • Define the words everyone assumes are obvious: duplicate, partial match, wrong item.

One more rule I insist on: a disputed case is never settled by seniority. It is settled by asking what the user would experience. If the user logs a meal and gets a number that is far off, the elegant argument for "technically correct" does not matter.

Take this with you

Checklist: answer these before the test round

  • What exactly is graded: the item name, each ingredient or the numbers the user sees?
  • What counts as a partial answer, and is it a pass, a fail or a separate category?
  • What counts as a duplicate, and is it compared by name or by content?
  • Who labels the expected result, and what happens when two labelers disagree?
  • Which user impact makes a case a fail even when the output is technically correct?
  • Who settles a disputed case, and by which rule?

If your team cannot answer one of these in a sentence, that is where your pass rates will drift apart.

Template: one borderline case

Copy this for every disputed case and keep the filled cards next to your grading rules.

Case:            two items on the plate, both returned as "sandwich"
Model output:    sandwich (bread, ham, cheese), sandwich (bread, egg, lettuce)
Disputed point:  is this a duplicate?
Verdict:         pass, two different items
Reason:          items with the same name and different ingredients are different items
Rule it creates: duplicates are compared by content, not by name
Enter fullscreen mode Exit fullscreen mode

The last line matters most. A case that does not produce a rule will be argued about again next round.

A pass rate without shared grading rules is not a quality metric. It is a measure of how strict one particular group of people felt that week.

Top comments (0)