Skip to content

Glossary

AI detectors: what the score actually measures, and what it cannot show

An AI detector estimates, from statistical properties of a text such as how predictable each word is, the probability that a passage was machine-generated. The output is a likelihood, not evidence. Detectors produce false positives, and they flag writers whose first language is not English disproportionately often.

Why it matters when you are writing

These tools work by measuring regularity. Machine-generated text tends to choose high-probability words and to vary sentence length less than human writing does, so a classifier trained on both can separate them better than chance. Better than chance is not the same as reliable at the level of one document, and a percentage presented to two decimal places gives an impression of precision the underlying method does not support.

The false-positive pattern is well documented and it is not random. Writing that is careful, formulaic, or lexically conservative scores as more machine-like — which describes second-language academic prose, heavily edited technical writing, and the standard phrasing of a methods section. Students who write in a plain, controlled register because they are working in their second or third language are flagged more often than fluent native speakers producing looser prose. That is a property of the measurement, not a finding about the writer.

If you are flagged, the useful response is evidentiary. Detector output is not proof of anything, and most institutions say so in their own policy; ask what the policy actually permits it to be used for. Then show your process — drafts, outlines, notes, revision history, whatever record exists of the work developing over time. A record of how the writing came about is the kind of evidence a panel can weigh, which a probability score is not.

IN PRACTICE

Reading a detector result honestly

A score of “92% likely AI-generated” means the classifier’s statistical profile for this passage resembles machine-generated text. It does not identify a model, a prompt, or an author, and it cannot distinguish careful second-language prose from generated prose.

OFTEN CONFUSED WITH

What AI detector is not

Plagiarism checker
A plagiarism checker compares your text against a database of existing documents; an AI detector compares statistical patterns and matches nothing.
Similarity score
A similarity score reports how much of your text matches known sources; an AI-detection score is a probability estimate about how the text was produced.
Proof of misconduct
A detector output is a probabilistic signal that can be wrong about any individual document; misconduct is a finding a person has to make on evidence.

Questions about ai detector

How accurate are AI detectors?

Accuracy varies by tool and by text, and no detector is reliable enough to settle a case about one document on its own. Published evaluations consistently find both false positives and false negatives.

Why do detectors flag non-native English writers more often?

Because the signal they measure is predictability. Prose written in a careful, conventional register looks statistically more regular, and second-language academic writing often is more regular by design.

What should I do if my honest work is flagged?

Ask what your institution’s policy allows the score to be used for, then offer evidence of your process — drafts, notes, outlines, revision history. Argue from the record of how the work developed, not against the number.

Do detectors say which model wrote the text?

No. They return a likelihood that text is machine-generated. They do not identify a model, a version, or a prompt, and claims otherwise should be treated skeptically.

Write it once.
Prove you wrote it.

Copyeditor is an academic writing workspace — structured drafting, citations that build their own bibliography, and a signed record of how your document came together.

Apply for beta

*Beta opens this season — starting with students and researchers.

The Copyeditor writing workspace