Do AI detectors make mistakes in both directions?

Yes. They produce false positives on writing by people and false negatives on AI text that has been edited or style-imitated. No single threshold removes both errors at once, because both are driven by the same measures.

Which error is worse, a false positive or a false negative?

A false positive is usually worse. It accuses a real person of something they did not do. A false negative only means a score understated the problem, which matters far less in most settings.

What should you do with an AI detection score?

Treat it as one signal among several. Confirm with evidence about the writing process, avoid acting on a score alone, and treat every assessment as an estimate rather than a verdict.

Blog · 28 September 2026

AI detectors fail in both directions: and that tells you how to use them

Two stories ran in the same week. In one, the most trusted detector in the field was described as having a "human problem." In the other, a journalist beat the "world's best" detector in about half a minute. One is a false positive. One is a false negative. They are the same story.

Two failures, one cause

This week Bloomberg published a piece titled The World's Most Reliable AI Detector Has a Human Problem. In the same period, PCMag published I Outsmarted the 'World's Best' AI Detector in 30 Seconds. Elsewhere, a university provost's published op-eds and articles were flagged as AI-written by a detector marketed on a "near-zero error" claim.

Read those together and a pattern appears that the industry mostly avoids discussing:

Direction one, the false positive. A person is flagged. This is the failure that ruins reputations, ends assignments, and produces the campus backlashes now being written about.
Direction two, the false negative. A machine passes. This is the failure that makes the whole exercise feel pointless to anyone paying for it.

They are not separate bugs. They are two sides of one dial.

Why no setting makes both go away

Every statistical detector works the same way: it measures properties of the text, computes how unusual those properties would be under human writing, and compares the result against a threshold. Above the line: flagged. Below: not flagged.

Move the threshold down, and you catch more machine text, and flag more humans. Move it up, and you protect more humans, and let more machines through. There is no position on that dial where both error rates are zero. That is not a limitation of a particular vendor. It is what a threshold is.

A product that advertises "near-zero error" is describing one side of the dial and not mentioning the other. The dial did not disappear; you just cannot see where it was set.

This is why we publish both numbers instead of one. Our current model, at the strict threshold: 0.9% false positives against held-out human writing, catching 91.3% of AI text. Push it to our balanced setting and false positives rise to 5.0% for a higher catch rate. Neither number is "the accuracy." They are the same model at two different points on the dial.

Direction one is worse than a bad score

We have written before about how badly false positives land, pre-ChatGPT books scoring as AI, the Declaration of Independence scoring 99.99%. The mechanism is simple: detectors penalise writing that is uniform, plain, and low-variance. Legal writing, news writing, technical writing, and translated or non-native English all look like that. A clear writer is, structurally, the closest thing to a machine.

What is new is the scale of the consequence. When a tool carries a reputation for accuracy, a flag from it is treated as a finding rather than a question, and the burden lands on the person flagged, who has to prove a negative about their own authorship. That is an unwinnable position, and it is why we treat false-positive rate as the primary number to publish, not the flattering one.

Direction two: the false negative

No detector of this kind is unbeatable, and we do not pretend otherwise. Detection measures properties of a document, never the identity of its author, so any such tool can miss AI-written text. That is a statement about how the method works, not a how-to: we do not build, publish or advise on ways to defeat detection.

A detector is a smoke alarm, not a witness. Treat a verdict as a question to be answered with evidence , drafts, revisions, provenance, never as a finding on its own.

So what do you actually do with it?

Decide which error you can live with, before you run anything. Publishing a flagged human and publishing machine text are different harms. Pick your threshold deliberately: strict when the cost of accusing someone is high, balanced when you are triaging volume.
Never treat a score as a finding. A flag is a question. Answer it with evidence about the process, drafts, revisions, notes, provenance, not with the same tool set to a different sensitivity.
Look at the features, not the percentage. Every verdict from our tool comes with the individual signals behind it: paragraph uniformity, sentence-length regularity, opener repetition, simile density. A high score with no structural markers is a much weaker signal than a moderate score with four of them.
Keep the person in the loop. Any decision with legal or similarly significant effects needs a human who can explain it. That is in our licence terms, not just our advice.

What we will not claim

We will not claim zero false positives, and we will not claim we cannot be beaten. Our model is trained on human writing that deliberately includes non-native English and simple, uniform prose, because those are the registers other detectors get wrong most often, and we score 0.9% false positives at the strict threshold as a result.

What we offer is a detector with its dial visible: two named thresholds, both error rates published, the features exposed, and an engine that runs entirely on your own machine so the text you are checking never leaves your infrastructure. That is the honest version of this product. Everything else in this field is marketing.

Try the dial yourself

The free checker runs on the page. If you want it in your own workflow, no uploads, no API key, no word limits, the offline pack runs locally.

Tier	Price	What it covers
Personal	$19 one-time	One person, unlimited runs, offline
Commercial	$99 one-time	Embed it in one of your own products or services, up to 5M words/year
OEM	from $999/year	Multiple products, white-labelling, no volume cap

See the product & pricing

Or email support@prosechecker.com if you want help choosing which error you should be optimising against.