Quick answer: An AI writing detector is a classifier that estimates how likely it is that a passage was produced by a large language model, usually by measuring how predictable the text is to another language model.

Detectors can flag long passages of unedited chatbot output, but independent research shows they are neither accurate nor reliable enough to prove authorship: OpenAI withdrew its own classifier in July 2023 because of its low accuracy, a Stanford-led study found seven detectors misclassified more than half of non-native English essays as AI-generated, and a 2023 evaluation of 14 tools concluded that none was accurate or reliable. Treat any detector score as a prompt for a conversation, never as evidence on its own.

If you are a student, the question behind “AI writing detector” is usually one of two things: will my work be flagged, or which detector should I trust? This guide answers both fairly. It explains how detectors work, what the published evidence says about their accuracy, how the main types of tool differ, and what to do if your own writing is flagged. It does not recommend ways to disguise AI-generated text, because submitting machine-written work as your own is the problem detectors exist to address.

How AI writing detectors work

Most detectors rely on some combination of three ideas.

• Perplexity. Perplexity measures how “surprised” a language model is by each next word. The researchers behind the Stanford-led study published in Patterns describe it simply: if a model can predict the next word easily, perplexity is low; if the next word is hard to predict, perplexity is high. Text generated by a model tends to be low-perplexity, so many detectors treat very predictable prose as a sign of AI.
• Variation across sentences. Human writing tends to vary more from sentence to sentence in length and structure, while unedited model output is often more uniform. Some tools score this variation alongside perplexity.
• Trained classifiers. Commercial detectors also train machine-learning models on large sets of human and AI text and learn features that separate them. These models output a probability or a percentage of text predicted to be AI-written.

All three approaches share a weakness: they measure style, not origin. A careful human writer who uses plain, predictable sentences can look “AI-like”, and AI output that has been edited or paraphrased can look human. That is why accuracy depends so heavily on who wrote the text and what happened to it afterwards.

What the evidence says about accuracy

Several well-documented results frame what detectors can and cannot do.

OpenAI’s own classifier was withdrawn

When OpenAI launched its AI text classifier in January 2023, it reported that on a “challenge set” of English texts the classifier correctly identified 26% of AI-written text as “likely AI-written” while incorrectly labelling human-written text as AI-written 9% of the time. On July 20, 2023, OpenAI made the classifier unavailable because of its low rate of accuracy. The company that built the most widely used chatbot could not build a detector it was willing to keep online.

Detectors are biased against non-native English writers

Liang, Yuksekgonul, Mao, Wu and Zou (2023, Patterns, 4(7)) tested seven widely used GPT detectors on 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth-grade students. The detectors classified the US essays accurately but labelled more than half of the TOEFL essays as AI-generated, with an average false-positive rate of 61.3%. All seven detectors agreed that 19.8% of the human-written TOEFL essays were AI-authored, and at least one detector flagged 97.8% of them. The flagged essays had significantly lower perplexity, which the authors link to the more limited range of expressions that non-native writers often use.

An independent test of 14 tools

Weber-Wulff and colleagues (2023, International Journal for Educational Integrity, 19, article 26) evaluated 12 publicly available detection tools and two commercial systems widely used in academia, Turnitin and PlagiarismCheck. They concluded that the available tools are neither accurate nor reliable, that they have a main bias towards classifying output as human-written rather than detecting AI-generated text, and that obfuscation techniques significantly worsen their performance.

What vendors claim

Vendors publish their own figures, and they are worth reading carefully. In a June 2023 post, Turnitin’s chief product officer stated that its document-level false-positive rate is less than 1% for documents with 20% or more AI writing, and that its sentence-level false-positive rate is around 4%. Those claims refer to specific conditions and thresholds, so they do not mean that 99% of individual flags are correct.

Why a “1% false-positive rate” still matters

Even a low error rate becomes large at scale. Vanderbilt University made this point when it disabled Turnitin’s AI detector in August 2023. It noted that it had submitted 75,000 papers to Turnitin in 2022, and that a 1% false-positive rate would have meant around 750 student papers incorrectly labelled as partly AI-written. The university also cited concerns about how the tool worked and the evidence that detectors are more likely to flag non-native English writers.

The arithmetic is worth doing for your own context. If a department marks 2,000 dissertations a year, a 1% false-positive rate would wrongly flag about 20 of them, and each of those students would face an anxious process to show that their work is their own. The research above suggests the real rate for some groups of writers is much higher than 1%.

The main types of AI writing detector compared

Rather than rank individual products, whose models and settings change frequently, it is more useful to compare the categories of tool you are likely to meet.

If your university uses Turnitin, the explainer on whether Turnitin can detect ChatGPT covers how its indicator is shown to instructors. The key point for students is that no consumer tool reproduces your university’s settings, so a low score on a website is not a guarantee and a high score is not a verdict.

How universities are using detector scores

Practice varies widely. Some institutions have switched AI indicators off; others leave them on but instruct staff not to use a score as the sole basis for a misconduct allegation; many now emphasise assessment design, oral examination and process evidence instead. Two consequences follow for students. First, your institution’s written policy matters more than any vendor’s marketing, so read the assessment rules and the academic integrity policy for your programme. Second, the evidence that protects you is the same evidence an investigator would ask for: drafts, notes, version history and the sources you read.

Detectors also miss a great deal. In a blind study at a UK university, Scarfe and colleagues (2024, PLOS ONE) found that 94% of fully AI-written exam submissions went undetected by markers. Detection is therefore a weak tool in both directions: it can accuse honest writers and miss dishonest ones.

What to do if your own writing is flagged

1. Do not panic or rewrite in a hurry. A score is not a finding of misconduct. Ask which tool produced it, what score was recorded and which passages were highlighted.
2. Gather your process evidence. Export the version history of your document, collect your outlines and notes, and list the sources you read with the dates you accessed them.
3. Explain your writing process. Be ready to talk through how you developed an argument, why you chose a source and how a paragraph changed between drafts. Authors can do this easily for their own work.
4. Raise relevant context. If English is not your first language, or if a passage is formulaic by convention, such as a methods description, say so; the research above shows these are known causes of false positives.
5. Use the formal process. Follow your institution’s procedure, and ask your students’ union or advice service for support if a formal allegation is made.

The site’s guide on reducing AI detection false positives on your thesis covers the process evidence in more detail, including what to keep from the start of a project.

Should students run their own work through a detector?

It can be reassuring, but it is a limited check. Different tools disagree, none reproduces your university’s configuration, and a high score on honest work can cause needless worry or prompt you to make your writing worse in an attempt to “sound human”. A better use of the time is to make your process visible: write in a document with version history, keep notes as you read, and declare any permitted AI use clearly. If your programme requires an AI statement, the guide on writing an AI-use declaration for your thesis shows a format that covers tools, purposes and the parts you wrote yourself.

Why detection is getting harder, not easier

Each new generation of language models produces text that is more varied and less predictable than the last, which erodes the low-perplexity signal many detectors depend on. At the same time, students increasingly use AI for small, mixed tasks, such as rephrasing a sentence, suggesting a transition or tidying grammar, which leaves documents that are mostly human with scattered machine-edited phrases. That mixed authorship is exactly where sentence-level false positives are most common, according to the vendor figures above. The combined effect is that a single percentage says less and less about who wrote a document. This is one reason many universities are moving away from detection as the main safeguard and towards assessment designs that ask students to show their reasoning, such as oral defences, staged submissions and reflective commentaries on the writing process.

How to cite detector research in your own work

If you are writing about AI detection in an essay or thesis, cite the primary studies (Liang et al., 2023; Weber-Wulff et al., 2023; Scarfe et al., 2024) rather than blog summaries, give the conditions behind any vendor figure, and avoid presenting a vendor’s claimed false-positive rate as an independent finding. The site’s summary of AI detection accuracy statistics collects further figures with their sources.

For a related guide, see what AI essay writers do and what universities allow.

Where Tesify fits

The best protection against a false flag is a visible writing process. Start structuring your thesis in Tesify. It helps you structure and organize your thesis chapter by chapter, and 9,000+ students have written 15,000+ chapters with Tesify. Every word stays 100% written by you.

Frequently asked questions (FAQ)

How accurate are AI writing detectors?

Not accurate enough to prove authorship. An independent evaluation of 14 tools in 2023 concluded they are neither accurate nor reliable, and OpenAI withdrew its own classifier in July 2023 because of low accuracy.

Can an AI writing detector flag human writing?

Yes. In a 2023 study published in Patterns, seven detectors labelled more than half of essays written by non-native English speakers as AI-generated, with an average false-positive rate of 61.3%.

How do AI writing detectors decide a text is AI-generated?

Most estimate how predictable the text is to a language model, a measure called perplexity, and many combine this with a trained classifier. Very predictable, uniform prose scores as more AI-like.

Does a high AI score mean I will be accused of misconduct?

Not automatically. Many institutions instruct staff not to rely on a detector score alone, and some have disabled AI indicators. Your institution’s policy and the evidence of your writing process matter more.

Which AI writing detector do universities use?

Many use the AI writing indicator integrated into their similarity-checking platform, such as Turnitin, but settings and visibility are chosen by each institution, and some have turned the feature off.

Will editing AI text stop a detector from flagging it?

Research shows edited and paraphrased AI text is often missed, but disguising machine-written text and submitting it as your own is still misconduct under most university policies, whatever the detector says.

What should I keep to prove I wrote my thesis?

Keep your document’s version history, outlines, notes, annotated sources and correspondence with your supervisor. These show how your work developed and are the strongest response to a false flag.

Is it worth checking my own work with an online detector?

It has limited value, because different tools disagree and none matches your university’s configuration. Making your writing process visible and declaring any permitted AI use is a more reliable protection.