How to Detect AI-Generated Text: Evidence and Limits

plagiarism-checker-online.net Editorial Team  |  Updated October 4, 2026

How to detect AI-generated text is a question about evidence, not a reliable visual checklist. Reading style alone cannot establish authorship, and a detector score is not automatically a probability that a writer used AI. Review the sources, context, drafts, and relevant rules. Treat a concern as a reason to investigate fairly, not as a verdict.

Understand what a detector score means

An AI detector classifies text according to its model and reporting system. Products may display scores, labels, or highlighted passages, but these outputs do not all mean the same thing. Check the vendor's definition before interpreting a number.

Turnitin's AI Writing Report guidance defines its percentage in terms of qualifying prose flagged by the model. It is not the probability that the whole document was written by AI, and it is not the probability that the student cheated. Its denominator is qualifying text, not necessarily every component of the submission.

A result from our AI checker also cannot certify authorship or predict another system's response. More highlighted passages may warrant review, but they do not turn the score into independent proof.

Check the tool's validation and supported input

Ask whether the product supports the language, genre, and document format being analyzed. Check minimum input requirements in its current documentation rather than assuming a universal length threshold. A method tested on unedited English prose cannot be assumed to work on translated text, proofs, or code.

Weber-Wulff and colleagues' study tested detection tools under different text conditions and identified reliability problems. Those findings do not provide a fixed error rate for every current product, but they show why a benchmark's conditions matter. A vendor claim should not be treated as independent validation of a disciplinary use.

Read for quality, not a machine signature

Repetitive transitions, generic examples, polished structure, and words such as "delve" can be reasons to edit or ask for detail. Human writers use them too. Finding a cluster of such features does not establish a quantified probability of AI use.

Clark and colleagues' human-evaluation research shows why readers' judgments about generated text require caution. It does not supply a guaranteed manual test for present-day models. An unusual voice is not proof of human authorship, and conventional prose is not proof of generation.

Check whether examples are verifiable and whether the reasoning actually supports the conclusion. These are useful editorial questions even when nobody suspects AI. A vague sentence is a writing problem first.

Do not turn perplexity or sentence length into a verdict

Perplexity is a language-model measure of text predictability. A reader cannot calculate it from a feeling that the wording is unsurprising. Predictable human writing exists, including formal and constrained writing.

Sentence-length variation also describes style rather than establishing origin. Counting words in a short run of sentences cannot reliably separate human writing from AI-assisted writing. See our linguistic fingerprinting guide for the difference between statistical features and authorship evidence.

Liang, Yuksekgonul, Mao, Wu, and Zou (2023), Patterns 4(7), 100779, doi:10.1016/j.patter.2023.100779, documented misclassification of non-native English writing in the tested samples. Do not treat formal vocabulary or a writer's language background as grounds for an accusation.

Verify the sources and specific claims

Locate the cited publications, check that they exist, and compare the cited passages with the argument. A fabricated reference or an unsupported claim is a substantive problem, but neither is unique to AI writing. Identify the error before deciding what it says about the process.

Similarly, text similarity can reveal borrowed wording that needs attribution. It does not determine whether a chatbot produced the draft. Source matching and AI classification are separate methods, as our plagiarism checking page explains.

Discuss the work and review genuine drafts

Where the academic procedure permits it, ask the writer how the argument developed, why a source was selected, or how an example supports the claim. Request available drafts or notes without assuming everyone has identical records.

A fluent explanation does not prove independent authorship, and difficulty answering does not prove AI use. Anxiety, disability, language background, and the time elapsed since writing may affect the conversation. Give the writer a fair opportunity to explain and follow any required adjustments.

Compare earlier work only when the samples are relevant. Changes in topic, genre, feedback, or language support can explain a style difference. Improvement is not inherently suspicious. A comparison can suggest a question, not answer it.

Keep policy and evidence separate

AI assistance may be permitted, required, limited, or prohibited for the task. Evidence that a tool was used does not by itself establish a breach. Review the actual instructions and the assistance involved. Our university policy guide explains how to find those rules.

Combining several weak, correlated style impressions does not necessarily make strong evidence. Avoid counting predictable wording, regular structure, and low sentence variation as though they were independent tests. Document what each piece of evidence establishes and what alternative explanations remain.

If your own work is flagged

Ask for the highlighted passages, the reason for the concern, and the institution's review procedure. Preserve genuine drafts, source notes, and records of authorized assistance. Do not rewrite honest prose solely to chase a lower detector score or submit a false account of your process.

Read the AI Detection Guide for the wider context. A private precheck cannot prevent an allegation, reproduce an institutional system, or guarantee an outcome. It can only support a review within its stated limits.

Frequently Asked Questions

Can you tell if text is AI-written just by reading it?

Not reliably. In a 2021 study by Clark et al., untrained readers told GPT-3 text from human text at about chance level. Reading for personal voice and specific detail helps, but treat what you notice as a reason to ask questions, not as a conclusion.

Does a word like "delve" prove a text was written by AI?

No. Researchers have found that words such as "delves" and "underscores" became much more frequent in PubMed abstracts after 2022, which points to growing use of language models. Human writers use those words too, especially careful non-native writers, so a few of them prove nothing on their own.

What should a teacher do after suspecting AI use?

Start with a conversation, not an accusation. Ask the student to walk through the argument and the sources, request drafts or version history, and compare the paper with earlier work. Then follow your institution's procedure, and don't base the decision on a detector score alone.

Do AI detectors work for languages other than English?

Less reliably, and often not at all. Most detectors were built and tested mainly on English, and Turnitin's AI writing detection only supports a short list of languages. Weber-Wulff et al. (2023) also found lower accuracy on machine-translated text. Check the vendor's list of supported languages before you rely on a score for German or any other language.

This article is part of our AI Detection Guide.