How Do AI Detectors Work, and Should You Trust the Score?

AI detectors have become standard equipment in classrooms, editorial teams, and hiring processes, largely because AI writing stopped being obvious. The output no longer reads as robotic; it reads as competent and slightly bland, which describes a great deal of human writing too.

Understanding what these tools actually measure explains both why they work at all and why a score should never be treated as a verdict.

The short version

• Detectors measure how statistically predictable text is, not who wrote it.
• Scores vary substantially between tools on the same passage.
• False positives cluster on non-native speakers, technical writing, and anyone with a plain, consistent style.
• No detector output is evidence of anything on its own.

What a detector is looking at

Perplexity: how surprising is the next word

Language models are next-word predictors. Text produced by one tends to consist of high-probability continuations — the word that a model would most likely have chosen. Human writing wanders more: an unusual verb, an odd construction, a word choice driven by a private association.

Detectors estimate this predictability using their own language model. Consistently low surprise reads as machine-generated.

The obvious problem: plenty of humans write predictably. Technical documentation, legal drafting, formulaic business prose, and the careful English of someone writing in a second language are all low-perplexity by nature.

Burstiness: does the rhythm vary

Human writing is uneven. A long, winding sentence with several clauses, then a short one. Paragraphs of wildly different lengths. Generated text tends toward the mean — sentences of similar length, paragraphs of similar weight, a rhythm that never surprises.

Detectors measure this variance. Low variance suggests generation.

Again, the same caveat: heavily edited writing, house-style-conformant writing, and writing by people who were taught to write uniformly all show low burstiness.

Trained classifiers

The third approach skips hand-picked signals and trains a model on large corpora of known-human and known-machine text, letting it find whatever patterns it finds.

This works well on text resembling the training data and degrades on anything else: a new model’s output, a different domain, a language the classifier saw little of, or text that has been edited by a human after generation. That last case is now the common one, and it is the case detectors handle worst.

How reliable are they, honestly

Published accuracy figures are usually measured on clean cases: fully generated text versus fully human text, in English, in a familiar domain. Real conditions are messier, and performance drops.

Three findings recur across independent evaluations:

• Non-native English writers are flagged disproportionately. This is the most serious and best-documented failure mode, and it makes unexamined detector use in education an equity problem.
• Light editing defeats detection. A human pass over generated text — reordering, substituting, adding a real detail — moves the statistical signature substantially.
• Tools disagree with each other. The same passage routinely scores 90% AI on one service and 15% on another. That disagreement alone tells you the score is not measuring a property of the text so much as a property of the detector.

Using them without causing harm

Treat a score as a prompt to look, not a finding. High score means read the work carefully, not accuse.

Never act on a score alone. If it matters, use process evidence: draft history, version control, the ability to explain the work, a conversation.

Tell people the policy in advance. Ambiguous rules plus automated enforcement is how good-faith people get penalized.

Check the failure mode before the tool. If a false accusation is expensive — a student’s record, a contractor’s reputation — the tool’s error rate on people like them is the number that matters, not its headline accuracy.

Remember what you actually care about. In most contexts the real question is not “was AI involved” but “is this accurate, original in substance, and does the person understand it.” Those are answerable directly.

If your own writing gets flagged

It happens, and it happens most to people who write clearly and consistently. What helps:

• Keep drafts. Version history is the most persuasive evidence there is.
• Be able to talk about it. Anyone who wrote something can discuss the choices they made and why.
• Do not rewrite to defeat a detector. Deliberately adding irregularity makes writing worse and proves nothing.

Frequently asked questions

Can human writing be flagged as AI-generated?

Yes, routinely. Plain, consistent, well-edited prose has exactly the statistical profile detectors associate with generation. Non-native English speakers and technical writers are flagged at notably higher rates.

Does editing AI text make it undetectable?

Substantial human editing reliably moves detection scores, which is one reason scores are not meaningful evidence. This is a description of how the tools behave, not a workaround worth building a workflow around.

What is the difference between an AI detector and a plagiarism checker?

A plagiarism checker matches your text against a corpus of existing documents and reports overlaps — a factual comparison. A detector makes a probabilistic guess about how text was produced, with no source to point to. The first can show its working; the second cannot.

Should schools use AI detectors?

Cautiously, and never as the sole basis for a decision. The false-positive distribution makes unexamined use actively unfair. Assessment redesign — in-class work, oral defense, process portfolios — addresses the underlying issue better than detection ever will.

Measuring texture, not authorship

An AI detector answers “does this text have the statistical texture of generated writing.” That is a genuinely different question from “who wrote this,” and the gap between them is where every false accusation lives. Use the score as a reason to look closer. Never use it as the answer.