What an AI text detector estimates
AI text detectors use different methods. Some train classifiers on human-written and machine-generated examples; others use zero-shot scoring approaches. Each result reflects that system’s method, data, and settings. It is an estimate, not a record of how a passage was written, and services can disagree about the same text.
A percentage shown by a service should not be read as the probability that a named person used AI unless the provider defines and validates that interpretation for the exact version and use case. A confident-looking number is still a model output.
Why a score can be wrong
Research has tested detectors on particular tools, versions, languages, and text collections—not every kind of writing. A 2025 NAACL study found that strong overall classification metrics did not necessarily describe performance at a low false-positive threshold in its evaluation. A 2023 study reported false positives and false negatives in its test set and found that paraphrasing and translation affected detection. Those studies show limits in the tested systems; they do not give a universal error rate for today’s products.Read the 2025 evaluation (opens in a new tab) and the 2023 detector study (opens in a new tab) for their methods and limitations.
A separate study of seven detectors found that its non-native English essay sample was often classified as AI-generated. The authors describe a specific dataset and detector set, so treat the result as evidence of a risk—not a rate that applies to every writer or current service. The paper and its data description (opens in a new tab) explain the scope.
Use a result as a prompt to review, not as proof
- Check the tool’s current documentation: what text it accepts, how its score is defined, and which version was tested.
- Review the work itself. Look at sources, drafts, revision history, and the author’s explanation when those are available.
- Give the writer a chance to respond before drawing conclusions. Keep detector output separate from verified evidence.
- Avoid submitting confidential or personal material to a third-party detector without first checking its data policy.
Sources and scope
The linked research evaluates named detectors and samples from 2023 and 2025. AI models and detector services change, and those studies are not current product audits of ZeroGPT. This article makes no accuracy claim about ZeroGPT or any other detector.