Can an AI detector
prove who wrote it?

A score describes one method applied to one sample.

AI text detectors analyze a passage and classify or rank it against patterns their method was built to recognize. Tools can use different models, features, thresholds, and reference data. Watermark detection is a distinct approach that depends on a compatible mark being present.

The result tells you what the detector estimated for the submitted text. It does not show who typed, edited, or approved it, and it cannot reconstruct the path from a first draft to a final version.

Research also examines watermarking: a generator deliberately changes token selection to embed a detectable signal. That method needs a compatible generator and can be affected by how the text is edited. It is different from a classifier trained to recognize text patterns.

A text detector and a plagiarism checker also answer different questions. Similarity software looks for matching passages in a comparison collection; a detector estimates whether text resembles its tested examples. Neither result alone establishes how a piece of writing was produced.

A benchmark can show capability without proving universal reliability.

NIST’s 2025 text-to-text pilot evaluated human and generated summaries against source articles across several rounds. Results varied substantially between generator and detector systems: some generators confused most tested detectors, while some detectors distinguished outputs from nearly all tested generators. The systems also improved over the rounds.

That is useful evidence about the measured task. It does not establish the accuracy of every detector on essays, short messages, other languages, edited passages, or future models. The test conditions are part of the result.

Earlier evaluations also found uneven performance and errors when human writing was translated or generated text was edited. Read the peer-reviewed evaluation by Weber-Wulff and colleagues ↗.

A detector can misread a person’s writing as generated.

A 2023 study tested several GPT detectors on English writing by native and non-native speakers. In the study’s sample, non-native English writing was more often misclassified as AI-generated. The researchers also found that simple changes to prompts could affect detection. These findings describe the detectors and texts they tested; they are a reason to test a system for bias, not a universal performance estimate for every current tool.

When a result may affect a student, applicant, employee, or creator, a false positive carries a real cost. The person should be able to see how the conclusion was reached, provide context, and challenge an automated flag.

Read the study on detector bias ↗

The same detector can mean different things in different settings.

A score is often mistaken for “the chance this person used AI.” But the chance that a flagged passage is generated also depends on how common generated text is in the group being checked, and on how often the detector flags human writing. A test score does not supply those facts by itself.

The lesson is to ask how the detector was tested, what population its scores describe, and how its threshold trades missed cases against false alarms. A single percentage without that context is easy to overread.

For authorship, look for process and provenance.

1. Define the question. Distinguish “does this sample resemble the detector’s examples?” from “who wrote this?” or “was a rule broken?”
2. Check the work’s context. Review relevant drafts, source material, edit history, and the rules the writer was given, where those records are available.
3. Ask the writer. Give the person a fair chance to explain their process and respond to concerns.
4. Keep detector output in proportion. If a tool is used, treat it as a limited screening signal inside a review process—not as a verdict.
5. State uncertainty. If the available evidence cannot establish what happened, say so.

For teams evaluating a detector, test it on the languages, genres, lengths, and editing patterns they actually encounter. Measure false positives and false negatives, check subgroup performance, and repeat the evaluation as models and workflows change.

What detector scores can—and cannot—settle.

Follow the evidence into the next question.

Read why a fluent AI answer can still be wrong ↗, how to read an AI benchmark ↗, and how to check the source and context of an image ↗.

Explore the AI field guide ↗ for concise explanations of evaluation, provenance, and other system concepts.