How Accurate Are AI Detectors? And What If You're Wrongly Flagged

What detectors actually measure

Text detectors generally work from a simple statistical idea. A language model, at each point in a sentence, has a strong sense of which word is most likely to come next. Text produced by such a model tends to sit close to those likely choices, consistently.

Human writing wanders more. People pick unexpected words, vary sentence length unpredictably, follow tangents, and occasionally write something slightly odd. Detectors quantify that: how predictable is this text, and how much does its variability change from sentence to sentence? Very predictable and very even reads as machine-like. Uneven and surprising reads as human.

That's the whole mechanism, and its limitation is visible from the description: predictability is a property of the writing, not of who produced it. Nothing in the text carries a signature.

Why false positives happen

Once you know the mechanism, the pattern of errors is predictable:

• Non-native English writers are disproportionately affected. Someone writing carefully in a second language tends to use common vocabulary and straightforward sentence structures — deliberately, because that's how you write clearly in a language you're still mastering. That reads as low-variability text, which is what the detector penalises. This is the most serious fairness problem with these tools, and it falls on exactly the people least able to argue back.
• Formal and technical writing scores high. Lab reports, legal summaries, instructions, and documentation are supposed to be predictable. Ambiguity is a defect in that kind of writing.
• Heavily edited writing scores higher than a rough draft. Editing removes the odd phrasing that made it look human. A polished final version can be flagged when the messy draft wouldn't have been.
• Grammar and style tools push text toward the flagged zone. Accepting suggestions that make writing more conventional makes it more predictable by definition.
• Short pieces are unreliable. There isn't enough text to measure variability. Most detectors say so in their own documentation, and most users don't read it.
• Quoted material and common phrasing — anything standard or widely repeated — inflates the score.

Meanwhile, in the other direction: rewriting machine text by hand, or running it through a tool designed to add variability, lowers the score. So the tool is easiest to fool by the people deliberately trying, and hardest on the people writing honestly.

What "99% accurate" means, and doesn't

Vendors advertise very high accuracy figures. Treat them carefully, for two reasons.

First, accuracy on whose test? A number produced by a vendor on their own dataset tells you how the tool performed on the examples they chose. Independent evaluation on messy real-world writing generally produces lower numbers than vendor claims. Look for who ran the test.

Second, and more important: even a genuinely excellent false-positive rate produces a lot of wrongly accused people at scale. This is the part that gets skipped. Suppose a detector wrongly flags 1% of human writing. In a course with 1,000 submitted essays, that's roughly 10 students accused of cheating who didn't. If the consequence of a flag is a misconduct hearing, "1% error" is not a small number — it's ten people.

That arithmetic is why a growing number of institutions have stopped treating detector output as evidence, and why the sensible policy is to treat a high score as a reason to ask a question, never as a finding.

If you're accused — what to do

Order matters here. This is stressful and the instinct is to argue about the tool, which is the weakest move available to you.

1. Don't panic and don't confess to something you didn't do to make the meeting end. People do this under pressure and it's very hard to undo.
2. Ask what the actual evidence is. Specifically: which tool, what score, and whether anything other than the score forms part of the accusation. A score alone is a weak basis and asking politely for specifics establishes that.
3. Don't attack the tool first. "Detectors are unreliable" sounds like what a guilty person would say. Lead with your own evidence, and let the tool's limitations be the supporting argument rather than the main one.
4. Produce your process, not your conclusions. See the next section — this is what actually works.
5. Offer to discuss the content. Someone who wrote a piece can explain why they structured it that way, what they cut, and which source they found least convincing. Someone who didn't, usually can't. Volunteering this is persuasive precisely because a person passing off machine text would avoid it.
6. Put it in writing. Follow up any conversation with a short email summarising what was said. It creates a record and it tends to make everyone more careful.
7. Ask about the appeals process early, before you need it.

The evidence that actually works

Detectors can't prove authorship. Neither can you, directly — but you can show the process, and a process is very hard to fake after the fact.

• Version history is the strongest thing you have. Google Docs, Microsoft Word with autosave, and most cloud editors keep a detailed revision timeline. It shows text appearing gradually, being rewritten, sentences moved, paragraphs deleted at 11pm. Machine-generated text arrives as one large paste. This distinction is visible at a glance and it is the single most convincing thing you can produce.
• So write in a tool that keeps history, and don't compose in a plain text editor if this is a risk for you. Our file recovery guide covers where to find version history in each service.
• Keep your notes, outlines, and sources. Photographs of handwritten notes with dates, browser history showing the research, downloaded PDFs with your highlights.
• Keep the earlier drafts, including the bad ones. A messy first draft is excellent evidence.
• Don't fabricate any of this afterwards. Timestamps are checkable, and being caught constructing evidence is far worse than the original accusation.

If you did use AI as an assistant — for an outline, for editing, to explain a concept — say so plainly and describe exactly how. Most institutions have a policy that permits some uses. Being straightforward about a permitted use is a much better position than being caught minimising it.

If you're the one checking

A few things worth knowing before acting on a score:

• Treat a high score as a prompt to look, not a conclusion. It is a smoke alarm, not a verdict.
• Ask for the version history before anything else. It resolves most cases in a minute, in either direction, and it doesn't require you to adjudicate a statistical claim.
• Be aware of who the errors fall on. If your flags cluster on international students, that is a property of the tool, not of those students.
• Never run the same text through several detectors and treat agreement as confirmation. They use similar methods, so they fail on the same text for the same reasons. Agreement between them is not independent corroboration.
• Have a written policy that says what is permitted and what evidence you rely on. Ambiguity is what makes these situations painful for everyone.

Detecting AI images is a different problem

Image detectors work differently and are worth separating out. They look for artefacts of the generation process rather than statistical predictability.

They also get less reliable over time, because each generation of image model removes the tells the previous detectors were trained on. Advice built on visual giveaways — hands, text in the background, garbled reflections — dates quickly, and much of what circulates is already out of date.

The more durable approach with images is provenance rather than forensics: where did this file come from, who published it first, and does a reverse image search show it appearing somewhere earlier with different context. That reasoning doesn't expire when the models improve. Our guide to spotting a fake website covers reverse image search, which is the same technique applied to a different question.