Can AI detectors tell if an essay was written by ChatGPT?

If an AI detector says that an essay was written by ChatGPT, is it right?

Sometimes. But that is not quite the same as saying the detector can tell who wrote the essay.

AI detection has improved significantly since the first wave of ChatGPT detectors appeared. Some modern systems are remarkably good at distinguishing untouched AI-generated text from ordinary human writing under controlled conditions. At the same time, research continues to find substantial differences between detectors, between different AI models, and between untouched AI text and work that has subsequently been edited.

The important distinction is this: an AI detector analyses the characteristics of the writing in front of it. It does not establish its authorship.

That matters enormously in education. A detector may provide useful evidence that a piece of writing deserves closer examination. A percentage on a screen, however, is not proof that a student opened ChatGPT and asked it to write an essay.

What does an AI detector actually detect?

There is no hidden ChatGPT watermark in an ordinary essay for a detector to uncover. Nor can an AI checker look inside a student’s ChatGPT account, discover their prompt history or trace a paragraph back to a particular conversation.

Instead, AI detectors work as classifiers.

They examine text and estimate how closely its characteristics resemble writing produced by large language models. Different detectors use different methods, and the details of commercial systems are not always public, but they may consider patterns across words, phrases and sentences, the predictability of language, relationships between tokens, sentence structure and other statistical or linguistic features.

Two terms frequently encountered in explanations of AI detection are perplexity and burstiness.

Perplexity is, very roughly, a measure of how predictable a sequence of language is. Large language models generate text by repeatedly estimating what is likely to come next. Their output can therefore have statistical patterns that differ from human writing.

Burstiness describes variation. Human writing often moves unevenly between short and long sentences, simple and complicated constructions, common and unusual words. Machine-generated prose can sometimes be more consistent.

These ideas are useful for understanding AI detection, but they shouldn’t be mistaken for a recipe used by every detector. Current systems may rely on much more complicated machine-learning models trained on large collections of known human and AI-generated texts.

What the detector ultimately produces is a classification or probability based on patterns.

It isn’t identifying an electronic fingerprint left by ChatGPT.

So can AI detectors detect ChatGPT essays?

Yes — particularly when the text has been generated by ChatGPT and submitted with little or no editing.

But the accuracy depends heavily on what is being tested.

In a 2023 comparison of 16 AI-text detectors, William Walters found considerable differences between products. Several performed well, while others struggled, particularly with text produced by more advanced language models.

Another major study, by Weber-Wulff and colleagues, tested 14 detection systems and reached a much more cautious conclusion. Performance varied widely and deteriorated when AI-generated text had been edited, paraphrased or otherwise altered.

Research has continued to produce similarly mixed results.

A 2025 study by Tufts, Zhao and Li tested detectors against text from models and subject areas that the systems had not necessarily encountered during development. They found that performance could decline dramatically in unfamiliar conditions and that relatively modest attempts to alter generated writing could make detection more difficult.

More recent research also illustrates how quickly the problem changes. A 2026 higher-education study tested several well-known detectors using human essays, completely AI-generated papers, mixed human/AI papers and “humanised” AI writing. Some systems performed very well in parts of the experiment while others substantially underestimated the amount of AI-generated material.

That doesn’t mean AI detection is useless. Far from it.

It means that statements such as “AI detectors are 99% accurate” need context. Accurate on what? Which AI model? Which language? What length of document? Was the AI output untouched? Had a human rewritten it? What threshold was used? And was the detector being tested on material similar to the data on which it was developed?

Those details can completely change the result.

Why two AI detectors can give completely different answers

Put the same essay through several AI checkers and you may receive surprisingly different scores.

That isn’t necessarily evidence that one of them is broken.

Detectors can differ in their training data, classification model, treatment of individual sentences, minimum document length and, crucially, the threshold at which something is classified as AI-generated.

There is an unavoidable trade-off here.

Make a detector very cautious and it is less likely to accuse a human writer incorrectly. The price is that it may miss more AI-generated text.

Make it highly sensitive and it may catch more heavily edited or difficult-to-detect AI writing. The price is an increased risk of human writing being flagged too.

Some services make that decision on the user’s behalf. Others expose it. For example, the PlagPointer AI detector allows the user to adjust AI detection sensitivity, so a cautious check intended to minimise false positives does not have to use the same threshold as a check intended to look more aggressively for heavily edited AI text.

That is a more useful way to think about detector settings than assuming there must be one universally “correct” threshold. There isn’t.

False positives: when human writing looks like AI

A false positive happens when genuinely human-written material is classified as AI-generated.

This is the error that understandably worries students most.

There are several reasons it can happen. Some forms of human writing are naturally predictable and highly structured. Academic prose is a good example. A literature review may contain conventional introductions to sources, repeated citation structures and careful transitions between studies. Technical writing can be similarly constrained.

A student who writes clearly and formally isn’t therefore immune from an AI flag simply because the work is entirely their own.

Earlier research also raised concerns about non-native English writers. Some detection methods appeared more likely to flag writing with relatively predictable vocabulary and sentence structures. Subsequent work has shown that this problem is not inevitable: detectors trained on appropriately sampled data can substantially reduce or avoid that bias. It is nevertheless another reason not to generalise from the performance of one detector to every AI detection system.

Grammar and rewriting software complicates the picture further. A student may write an essay themselves and then use a tool to correct grammar, smooth sentences or improve expression. At some point, the final document may contain linguistic characteristics that differ quite substantially from the student’s original draft.

Who “wrote” that text is no longer a particularly simple question.

False negatives: when ChatGPT writing looks human

The opposite mistake is a false negative: AI-generated text is classified as human.

This is equally important when evaluating claims about detector accuracy.

An AI detector may be very good at recognising a straightforward ChatGPT response produced from a simple essay prompt. That doesn’t mean it will perform equally well on text produced with a detailed prompt specifying style, vocabulary, sentence length and intended audience.

The underlying AI models are changing too. Research carried out on GPT-3.5 cannot simply be assumed to describe the detection of GPT-4, GPT-4o or later generations of models.

That creates an unusual technological contest. Generative models become more capable of imitating varied human writing, while detector developers retrain their systems to recognise newer output.

The performance of an AI detector is therefore something that has to be tested repeatedly. It isn’t a fixed property of the product.

Why an AI score is not proof of academic misconduct

Suppose a university’s detector reports that an essay is “80% AI”.

It is tempting to read that as: there is an 80% chance the student cheated.

That may not be what the number means at all.

AI detection percentages can represent different things depending on the system. They may refer to the proportion of text classified as AI, an aggregated confidence measure, or some other model-specific calculation.

More importantly, even a very accurate classification system cannot answer questions it has never observed.

It doesn’t know whether the student used ChatGPT.

It doesn’t know whether AI use was permitted.

It doesn’t know whether the student used AI to brainstorm, translate, proofread or generate entire paragraphs.

It doesn’t know whether somebody else edited the essay.

It doesn’t know whether a flagged passage came from the student’s own previous work.

And it doesn’t know how the essay developed over the previous three weeks.

There is also a statistical problem that is often overlooked: false-positive rates have to be considered alongside how common the thing being detected actually is.

Imagine a hypothetical detector that catches 80% of AI-written essays and falsely flags 1% of human essays.

If 100 of 1,000 essays were AI-written, it might flag around 80 genuine cases and nine human essays. Most flagged essays would therefore genuinely contain AI writing.

But suppose only ten of those 1,000 essays were AI-written. The same detector might identify eight of them while falsely flagging around ten human essays.

Suddenly, fewer than half of the flagged essays would actually be AI-written.

Nothing about the detector changed. The context did.

This is one reason a detector result should be treated as evidence requiring interpretation, not an automatic verdict on misconduct.

What happens when AI writing is edited or paraphrased?

Editing is one of the hardest problems for AI detection.

Research has repeatedly shown that rewriting AI-generated text can alter the features on which detectors rely. Studies involving paraphrasing tools such as QuillBot and Wordtune have found that detection rates can fall after generated material is rephrased.

Human editing can have the same effect.

That makes sense. If a student substantially rewrites a paragraph — changing its vocabulary, syntax, argument and rhythm — the resulting text has fewer of the statistical characteristics of the original generated version.

But this subject is sometimes oversimplified online into “paraphrasing beats AI detectors”. The evidence doesn’t support such a universal claim.

Some detectors remain effective against particular types of rewritten text. Others don’t. Results vary according to the detector, AI model, rewriting method and amount of alteration.

There is another complication: not all editing is an attempt to evade detection.

Consider these three situations:

1. A student asks ChatGPT to produce an entire essay and changes a handful of words.
2. A student writes an essay and uses an AI tool to improve awkward sentences.
3. A student discusses ideas with ChatGPT, researches the subject independently and writes the final essay themselves.

Trying to force all three into a simple “AI” or “human” category loses a great deal of useful information.

Universities increasingly have to decide not merely whether AI was involved but what role it played in producing the submitted work.

A detector cannot answer that question on its own.

AI detection is not the same as plagiarism checking

AI detection and plagiarism detection are often bundled together, but technically they are solving different problems.

An essay written entirely by ChatGPT may contain no plagiarism whatsoever. A language model can generate a fresh combination of words which doesn’t match an existing source closely enough to trigger a plagiarism checker.

Conversely, somebody can manually copy material from a journal article. An AI detector may correctly conclude that the passage looks human-written, while a plagiarism checker finds the source almost immediately.

An essay can therefore be:

• human-written and original;
• human-written but plagiarised;
• AI-generated and original;
• AI-generated and also contain copied material; or
• some mixture of all of these.

That is why universities shouldn’t treat an AI score as a replacement for conventional source checking.

Can a lecturer tell that an essay was written by ChatGPT?

Humans aren’t infallible detectors either.

Several studies have asked teachers, students or expert reviewers to distinguish AI-generated writing from human work. Results have often been surprisingly poor, and researchers have found that confidence in the judgement does not necessarily correspond with accuracy.

People tend to look for supposed “AI words”, overly polished prose, repetitive conclusions, formulaic headings and suspiciously balanced paragraphs.

Those clues can sometimes be useful.

They can also describe perfectly ordinary academic writing.

And as millions of people become familiar with AI-generated prose, the picture is changing again. A 2025 study found that frequent users of ChatGPT were considerably better than inexperienced readers at identifying AI-generated writing in the particular texts tested.

So human judgement has a role — but “it sounds like ChatGPT” shouldn’t be the end of an academic misconduct investigation either.

What should students do about AI detectors?

For students, the most useful strategy isn’t trying to engineer an essay that receives a particular AI score. It is making sure there is good evidence of how the work was actually produced.

Check the rules first. Universities, departments and even individual assignments can have different policies. Using AI to plan an essay may be permitted where using it to draft paragraphs is not.

Keep your working material. Notes, research, reading lists, essay plans, earlier drafts and document version history can be far more informative than an AI percentage if somebody later questions authorship.

Keep a record of permitted AI use. If your institution allows tools to be used for brainstorming, proofreading or another limited purpose, keep a note of what you used and how.

Do not panic over a detector result. A flag is not proof that you have done anything wrong. Look at the sections identified and consider whether there is a plausible reason for them being classified that way.

Do not endlessly rewrite genuine work simply to make a detector say “human”. Apart from potentially making the essay worse, changing authentic work in response to an imperfect classifier rather defeats the purpose of proving that it is yours.

If you are challenged, explain your writing process. Drafts, research notes, editing history and the ability to discuss your argument can provide context a detector simply does not possess.

Students who have actually used AI should be equally cautious about drawing the opposite conclusion. A zero or low AI score does not establish that the work complies with university rules. A detector failing to recognise AI-generated text doesn’t change how that text was produced.

What does a sensible use of AI detection look like?

The strongest case for AI detection is as a screening and investigative tool.

A detector can highlight passages worth looking at more closely. A lecturer can then consider other information: differences in writing style, unusual references, drafts, version history, previous work, the student’s explanation and the institution’s rules governing generative AI.

That is quite different from automatically accusing somebody because an algorithm produced a high percentage.

The distinction also protects the usefulness of AI detection itself. Expecting a detector to prove authorship sets an impossible standard. Using it to identify patterns that may justify further examination is a much more realistic role.

The same principle applies to students checking their own work. An AI detector can show how a document is likely to be classified by automated software. That information may be useful, especially where legitimate editing or writing tools have been used.

It still isn’t an oracle.

So, can AI detectors tell if an essay was written by ChatGPT?

They can often detect evidence consistent with ChatGPT-generated writing, and the better detectors are considerably more sophisticated than the simple tools that appeared when ChatGPT first became popular.

But they cannot look backwards and establish who actually wrote an essay.

False positives remain possible. False negatives remain possible. Editing can change results. Different detectors can disagree. New language models can temporarily outpace systems trained to recognise older ones.

Most importantly, an AI detection score answers a narrower question than many people assume.

It asks, in effect:

“How much does this writing resemble the kinds of AI-generated text this detector has learned to recognise?”

That can be a useful question.

It just isn’t the same question as:

“Did this student use ChatGPT to cheat?”

No percentage, however impressive it looks, should erase that distinction.

References

Bernabei, M., Colabianchi, S., Falegnami, A. and Costantino, F. (2023). Students’ use of large language models in engineering education: A case study on technology acceptance, perceptions, efficacy, and detection chances. Computers and Education: Artificial Intelligence, 5, 100172.

Fleckenstein, J., Meyer, J., Jansen, T., Keller, S., Köller, O. and Möller, J. (2024). Do teachers spot AI? Evaluating the detectability of AI-generated texts among student essays. Computers and Education: Artificial Intelligence, 6, 100209.

Jiang, Y., Hao, J., Fauss, M. and Li, C. (2024). Detecting ChatGPT-generated essays in a large-scale writing assessment: Is there a bias against non-native English speakers? Computers & Education, 217, 105070.

Liu, J.Q. et al. (2024). The great detectives: humans versus AI detectors in catching large language model-generated medical writing. International Journal for Educational Integrity, 20.

Tufts, B., Zhao, X. and Li, L. (2025). A practical examination of AI-generated text detectors for large language models. Findings of the Association for Computational Linguistics: NAACL 2025.

Walters, W.H. (2023). The effectiveness of software designed to detect AI-generated writing: A comparison of 16 AI text detectors. Open Information Science, 7.

Weber-Wulff, D. et al. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26.