An AI output checker is judged by what it refuses to claim.

It is not a detector. A checker that scores text for being AI-written is answering an authorship question, and answering it badly, because the signals it relies on are style, not provenance.

It is not a plagiarism check either. Similarity search compares your text to a corpus; it says nothing about whether the claim in paragraph two is true, whether the code executes, or whether the file kept its metadata.

It is also not a grader of quality. Quality includes fitness for a purpose, and purpose lives outside the document. A checker can only report defects against rules it was given.

Output checks fall into groups that fail in different ways, and a tool that only covers one group will feel useful while missing whole categories.

Before you trust a checker, build a file containing defects you planted on purpose: one plausible key, one eval, one empty catch, one invented statistic, one unfinished placeholder, one heading that skips a level. Run it and count what comes back.

Then run it twice. A deterministic checker returns the same findings and the same score, which is the property that makes a report citable. A model-judged one drifts, and the drift is the earliest sign that the result is an opinion.

Finally, feed it a clean, deliberately plain paragraph and see whether it invents findings. A tool with no false-positive discipline will always look thorough, because it always finds something.

A score with no coverage statement is a guess wearing a suit. The report should say how many checks ran, what types of input they apply to, and where truncation occurred when a document exceeded the input limit.

That sentence is what keeps a green result honest. Zero findings across fourteen checks on a fragment is not the same claim as zero findings across forty checks on a full file, and the reader cannot tell the difference unless the tool tells them.

The same logic applies to images. A checker can report that provenance metadata is absent, present or damaged. It cannot validate a signature chain against a trust list, and it should say so instead of implying verification.

Severity weighting exists so a single credential finding cannot be averaged away by nine cosmetic notes. A checker without weighting produces scores that look stable and mean nothing.

The other side of the ledger is false positives, which are taxed in human attention rather than money. One rule that fires on legitimate code teaches people to skim, and a skimmed report is worse than no report because it creates the appearance of review.

Sooner or later you will disagree with a rule. A good checker makes that productive: the finding names the pattern, you decide it is acceptable in a fixture, and you note it. The tool has then done its job by starting an argument with a location.

No output checker can tell you whether the argument holds, whether the audience will read a claim as a promise, or whether a senior engineer would merge the change. Those are judgement, and judgement needs an owner with a name.

The handover should be cheap. A report that travels with the review request means the specialist starts from findings rather than from a blank page, and the two artefacts then describe different things: absence of known defects, and an attributed decision about fitness.

That pairing is the only version of automated QA that holds up under scrutiny, because each half states exactly what it proves.