A false positive is not an edge case. It is the predictable result of the base rate.

A false positive is a piece of human writing scored as machine-generated. It is not a near miss or a low-confidence reading; it is the tool asserting something about authorship that is untrue.

False negatives get discussed less and matter just as much: generated text scored as human, which is now common enough that a detector's clean result on a suspicious document tells you very little.

Both errors come from the same place. Detectors estimate whether text looks statistically typical of model output. That is a guess about style, and style overlaps between people and models for entirely innocent reasons.

The consequence is asymmetric. A false negative embarrasses a detector. A false positive can fail a student, damage a professional reputation or end a contract, which is why the standard of evidence should be much higher than a percentage.

Most tools measure properties such as how predictable each word is and how much sentence length varies. Model output tends to be predictable and even, so predictability and evenness are treated as evidence.

Every one of those properties has innocent causes. A writer who has been edited heavily becomes smoother. A writer working in a second language becomes more formal and more predictable. A genre with conventional phrasing, such as a lab report or a legal memo, becomes statistically uniform.

The models themselves keep changing, and a detector tuned on one generation's output drifts against the next. Tools also disagree with each other on the same text, which is hard to explain if any of them is measuring authorship directly.

So the underlying claim, which is that a style score reveals who or what produced a document, is not one the technology can support. It can support a suspicion, and a suspicion is not a finding.

The pattern is documented and consistent, and it is not random. Certain groups produce the prose the tools score as machine-like.

Suppose a detector is 95% accurate at identifying human text and 5% of the documents submitted are actually generated. Take 1,000 documents: 50 generated, 950 human.

It catches about 47 of the 50 generated ones. It also misclassifies about 48 of the 950 human ones, a rate of one in twenty.

So of the 95 flagged documents, roughly 48 are human. Half the accusations are wrong, and the tool is described as 95% accurate.

This is not a flaw that a better model fixes; it is arithmetic. When the thing you are detecting is rare, most positive results are false, and the only fix is to require corroborating evidence before acting.

A percentage invites a decision because it looks like a measurement. A probability about authorship is not a measurement, and presenting it as one transfers responsibility from the tool to whoever types the number into a report.

Consequences should scale with certainty, and certainty here is low by construction. Flagging a document for a conversation is proportionate. Failing a student, retracting a paper or terminating a contract on a score alone is not.

Process matters more than accuracy. If there is an appeal route, a person who reads the work and a requirement for more than one kind of evidence, the errors get caught. Without those, the tool is the decision.

Institutions that adopted detectors quickly are now withdrawing them for exactly this reason: the error rate is concentrated among the people least able to contest it.

Specificity, first. Which tool, which version, which threshold, on what date. A detector's output changes between versions, and an accusation that cannot name its instrument cannot be checked.

Corroboration, second. Edit history, drafts, notes, reading, a supervisor's recollection, the ability to explain the argument without the document. Any of these outweighs a style score.

Consistency, third. If the same writer's earlier work scores the same way, that is evidence about the writer's style, not about this document.

And proportionality, fourth. Ask what the person loses if the accusation is wrong, and whether the evidence would justify that loss in any other domain.

Keep drafts. Version history in a document editor or a writing app is the single most useful artefact, because it shows the work emerging rather than appearing complete.

Keep notes and sources. A page of links, a reading list, a folder of screenshots: anything showing the argument was built from material you read.

Keep the outline. A structure you can explain in conversation is harder to fake than a paragraph you cannot.

If you write formally, or in a second language, say so before anyone asks. A sentence in the author note about your background prevents a whole category of suspicion, and it is nobody's business until it is.

Start with a conversation and the document in front of you. Ask the author to explain a specific paragraph, extend an argument, or defend a choice. Understanding is not proof of authorship but its absence is informative.

Require more than one category of evidence. Style, history, conversation and context should agree before a serious consequence is applied, and disagreement should resolve in the author's favour.

Publish the policy before using any tool at all. If the rule is that detector scores are not used, staff and students know where they stand and the accusation industry loses its market.

Record outcomes, including the cases where the tool was wrong. Institutions that track false positives change their policies, and institutions that do not keep arguing from a single number.

The productivity question is whether the work is any good, and that is answerable without knowing who wrote it. Accuracy, sourcing, originality and whether the conclusion follows are all checkable.

Verification also scales better. Reading a claim and finding the source is reliable in a way that style scoring is not, and it catches the failure that actually matters: a confident sentence with nothing behind it.

This is the argument for a QA posture rather than a detection posture. You check the output against the world, not against a guess about its author, and the result is a decision you can defend when challenged.

Where provenance genuinely matters, use credentials rather than style: signed content, watermarks, camera or editing records, published workflow notes. Those are evidence about production, which is what the question was.

Ask for the specifics in writing: which tool, which version, which output, which date. Requests this simple eliminate a large share of informal accusations, because many are based on a feeling nobody has documented.

Then produce your artefacts and offer to discuss the work. Drafts, sources and an explanation of your reasoning are stronger than any counter-score.

Do not argue that the tool is inaccurate in the abstract. Argue about your document, and about what the consequence would be if the accusation is wrong. That is the argument a reasonable decision-maker can act on.

Finally, put the same standard in writing for next time. A file of your own process notes and version history makes the next challenge a five-minute exchange rather than a dispute.