What AI Detectors Can And Cannot Prove

AI detectors are now used to grade students, screen job applicants, and question freelancers about their invoices. The scores they produce look like evidence. They are not, and the gap between what they measure and what people do with the result is causing real harm.

This is not an argument that AI writing should go unexamined. It is an argument for understanding what a percentage from a detector actually represents.

They Measure Predictability, Not Authorship

A detector does not recognise text a model produced. It has no record of what was generated and no signature to look up.

What it measures is statistical: how predictable the word choices are, and how much variation there is in sentence structure and length. Language models tend to select high-probability words and produce even, regular prose, so text with those properties scores as machine-like.

The immediate consequence is that the detector is measuring a property of the writing, not a fact about its origin. Predictable, evenly structured human writing scores as AI. Deliberately varied AI output scores as human.

That is not a bug to be fixed in a future version. It is what the method is.

The Misfires Land On Specific People

Because the signal is regularity, the errors are not evenly distributed. They concentrate on writers whose prose is naturally more uniform.

Non-native English speakers are the most documented case. Writing carefully in a second language tends to produce more conventional constructions and less idiosyncratic phrasing, which is exactly what the detector reads as machine-generated.

Technical and legal writing has the same property by design. A well-drafted clause or a bug report is supposed to be predictable; ambiguity is the failure mode those genres exist to avoid.

And anyone writing to a house style, a template, or an accessibility standard is producing regular text on purpose.

So a detector used as a screening tool systematically produces more false accusations against people writing in a second language and people writing in constrained professional registers. That is a fairness problem before it is a technical one, and it is the kind of disparate impact that frameworks like the NIST AI Risk Management Framework exist to make people look for. Writing professional English as a non-native speaker covers the underlying dynamic.

The Individual Tells Are Worse Than The Statistics

If aggregate detection is unreliable, single-feature detection is worse, and it is what most informal accusations rest on.

The em dash is the current favourite. It is a normal punctuation mark used by careful writers for a century, and its presence proves nothing at all. The em dash is not proof of AI covers why this particular claim will not die.

Others in the same family: the word “delve”, tricolon lists, the phrase “it’s important to note”, perfectly balanced paragraph lengths. Each is a genuine tendency of model output and none is remotely diagnostic in a single document.

What these tells are actually useful for is editing your own work. If you want AI-assisted text not to read as machine-written, they are a decent checklist. As evidence about someone else, they are worthless. Small tells that make writing look AI generated treats them as an editing aid, which is the right use.

What A Score Can Legitimately Support

Detectors are not useless. They are useful as a signal that prompts a human process, and useless as a verdict.

Reasonable: flagging a submission for a conversation, prioritising which of two hundred applications a person reviews, noticing that a supplier’s output changed character abruptly.

Not reasonable: failing someone, terminating a contract, or making an allegation on the strength of a percentage. The false positive rate on the affected groups is too high, and the person accused has no way to prove a negative.

The reliable evidence is process, not text. Drafts, version history, the ability to discuss the reasoning, familiarity with the sources. Those establish authorship in a way no classifier can.

If You Are Accused, Respond With Process

The instinct is to argue about the tool. That rarely works, because the person holding the score usually cannot evaluate the criticism.

Before:

That detector is garbage, it flags everything, there are studies showing it fails on non-native speakers. I did not use AI.

After:

I wrote this. My drafts are in version history from the 3rd onward and show how the argument developed, and I am happy to walk through the sources and the reasoning. I did use a tool to tighten the phrasing in the final pass, which I am glad to describe. If the detector score is the concern, I would rather address it with the drafts than with a debate about the tool.

The second is stronger because it offers verifiable evidence and concedes the true part without conceding the accusation. Denying tool use you did make is the one move that turns a recoverable situation into a credibility problem.

A Wrivio Context for responding to an accusation could say:

Rewrite this as a calm, factual response. Keep the offer of evidence and every date and detail exactly as written. Do not add apology, do not add defensiveness, and do not concede anything the text does not already concede.

Press Ctrl+Shift+Space, paste your draft, and check the diff. The two failure modes here are a model making you sound either apologetic or aggressive, and both are worth catching before you send.

This describes how these tools work rather than what your rights are. If an accusation carries academic or employment consequences, get advice from someone qualified in that setting.

Common Questions

Can AI detectors reliably tell if text was written by AI?

No. They measure how predictable and uniform the writing is, which correlates with model output but also with careful human writing, so both false positives and false negatives are common.

Why do detectors flag non-native English speakers so often?

Because writing carefully in a second language tends to produce conventional, regular constructions, which is exactly the statistical property detectors read as machine-generated.

Does using an em dash mean text was AI-written?

No. It is standard punctuation used by human writers long before language models existed, and no single stylistic feature is diagnostic in one document.

What actually proves I wrote something?

Process evidence: drafts, version history, the ability to discuss the reasoning and sources. No classifier can establish authorship, but a document’s development history usually can.

Download Wrivio for Windows to rewrite your own drafts on your own machine, keeping the version history that actually demonstrates how the work developed.