You wrote the essay yourself. Then a message arrives: a detector has scored it as likely AI-written, and someone wants to talk to you about it.
It is a frightening message, and it helps to know two things. A detector's score is an estimate, and the companies that make detectors say so themselves. And there is a fair way through, which rests on something a detector cannot see: the record of how you did the work.
One part of that record you can check yourself, tonight: your sources. Whether a reference exists is settled by the publisher's record, not by a probability, which makes it the one point in a case like this that can be looked up rather than estimated, and EdCitation's free Verify references does the looking up. The rest of this guide covers the score, the process and the evidence.
What does an AI detector actually measure?
An AI detector measures how closely a text resembles the writing a language model tends to produce, and returns a probability or a percentage. It does not observe how the text was made. It has no access to your drafts, your notes or your keyboard.
Many detectors lean on how predictable the wording is, a measure researchers call perplexity. Language models choose likely words, so very predictable text looks machine-made. The trouble is that plenty of human writing is predictable too: formal academic prose, writing built from set phrases, and writing by people working in a second language with a smaller vocabulary.
What a score is not
A score is not a measurement of who wrote something, and three things follow from that.
It is not stable under small changes in the text. Perkins et al. (2024) tested six detectors on 805 pieces of machine-generated writing and found an accuracy of 39.5% on unmodified AI text, falling to 22.2% once simple evasion techniques were applied, a drop of 17.4 points. A number that moves that far when the words are shuffled is not measuring authorship.
It is not symmetrical. Weber-Wulff et al. (2023) put 12 freely available tools and two commercial systems, Turnitin and PlagiarismCheck, through the same documents and concluded that the available tools are "neither accurate nor reliable", leaning mainly towards calling text human-written. Missed AI text was the commoner error, which is the mirror image of the error students fear.
It is not evidence of intent. Even a correct flag says nothing about whether a student pasted an answer in, used a grammar tool their course allows, or wrote in a plain style.
How often do AI detectors flag human writing?
It depends on the detector and on the writer, and for some groups of writers the published error rates are high. These are the main public findings, each from its own source.
| Source | What it reported |
|---|---|
| Liang et al. (2023), Patterns | Seven public detectors were run on 91 human-written essays from the TOEFL English exam and on 88 essays by US eighth-grade students. They classified the US essays almost without error, but their average false positive rate on the TOEFL essays was 61.3%. All seven wrongly flagged 19.8% of them, and 97.8% were flagged by at least one. |
| Chechitelli (2023a, 2023b), Turnitin | Turnitin's Chief Product Officer gave a document-level false positive rate of less than 1% for documents with 20% or more AI writing, and a sentence-level rate of around 4%. The company reports that false positives are commoner in mixed documents and where the AI writing score falls between 0 and 19. |
| OpenAI (2023) | OpenAI's own classifier, launched on 31 January 2023, labelled human-written text as AI-written 9% of the time in the company's tests and caught only 26% of AI-written text. OpenAI said it should not be a primary decision-making tool, and withdrew it on 20 July 2023 for its low rate of accuracy. |
| Weber-Wulff et al. (2023), International Journal for Educational Integrity | A test of 12 public tools and two commercial systems found them neither accurate nor reliable, and more prone to miss AI text than to accuse a human writer. |
| Perkins et al. (2024) | Six detectors, 805 texts. Accuracy of 39.5% on unmodified AI writing, falling to 22.2% after simple evasion techniques. |
Why non-native English writers are flagged more often
The Liang study is the one to know, because the gap it found was not small. The same seven detectors were near-perfect on essays by US school students and wrong about most of the TOEFL essays, an average false positive rate of 61.3%. Then the researchers did something that explains the mechanism: they had GPT-4 rewrite the TOEFL essays with richer vocabulary, and the average false positive rate fell to 11.6%. The detectors were reacting to the range of the language, not to who had written it.
Two cautions about that study. It tested detectors available in March 2023, and it did not include Turnitin. Detectors have changed since, and no comparable public test of every current tool exists.
Why a small percentage is not a small number
A rate is not a count. Vanderbilt University, explaining in August 2023 why it switched off Turnitin's AI detector, did the sum: it had submitted 75,000 papers to Turnitin in 2022, so a 1% false positive rate would have meant about 750 papers wrongly labelled. Scale matters again at the other end: Turnitin reported in April 2024 that its detector had reviewed more than 200 million papers in its first year, and that about 11% carried at least 20% AI writing and about 3% at least 80%. Whatever the rate, a system running at that volume produces a steady stream of both errors.
What should I do if I am falsely accused of using AI?
Slow down, change nothing, and find out exactly what is being said and under which process. In order:
- Do not edit or delete anything. Leave the submitted file, your drafts and your notes exactly as they are. They are your evidence.
- Do not admit to something you did not do. Accepting a penalty to make it stop can be tempting under stress, and a finding can stay on your record.
- Ask, in writing, what the concern is. Which tool was used, what score it gave, which passages were marked, and whether this is an informal conversation or the start of a formal case.
- Read the academic integrity policy. Look for the stages, the time limits, your right to see the evidence, your right to respond, whether you may bring someone with you, and how to appeal.
- Get advice. A students' union advice service, an ombuds office or a personal tutor has usually seen this before.
- Check your references before the meeting. Upload a copy of the paper to Verify references on EdCitation, free and with no account. Each entry comes back verified, marked "check this", or not found, and every in-text citation is matched to the list. A wrong year is easy to explain; a source nobody can find is not, so learn about it before anyone asks, and note it rather than editing the submitted file.
- Reply on time, calmly, with your evidence. Offer to talk through the work. The next section lists what to bring.
Turnitin's own advice to instructors points the same way: the highlighted sentences are there "to initiate a conversation, not to draw a conclusion".
What does a fair process look like?
The burden should be on the university to show what it says happened, not on you to prove your innocence. The clearest published statement of that in the UK comes from the Office of the Independent Adjudicator, the independent body that reviews student complaints in England and Wales. Its Good Practice Framework puts the burden of proof on the provider: it is for the university to show that the student did what is alleged, and where regulations do not state a standard of proof it is reasonable to assume the balance of probabilities.
The framework also expects the student to be given copies of all the information the decision maker will consider, in advance, with reasonable notice of any hearing, a fair opportunity to respond to it, and the right to appoint a representative. It treats telling the student early, as soon as possible after the event, as good practice, and expects reasons to be given for any decision and any penalty.
Universities outside England and Wales are not bound by it, and your own institution's regulations are what govern your case. It is still a useful yardstick: if you are being asked to disprove a score you have not been shown, something has gone wrong with the process, and that is worth saying politely and in writing.
How do I show that I wrote it?
Show the process, because a process is hard to fake and a detector never saw it. Useful evidence, strongest first:
- Version history. Google Docs, and Word files saved to OneDrive, keep dated versions that show the text growing over days, with the false starts and deletions of real writing.
- Earlier drafts and files. Dated copies, emailed versions, a plan, an outline, a draft a tutor or friend commented on.
- Notes and reading. Handwritten or typed notes, annotated PDFs, library loans, saved searches, your reference manager, or the reading list and reference lists kept in a free EdCitation account.
- Your sources. Be ready to say what each one argues and where in your essay you used it. This is also where a clean reference list helps: every source real, and every one you can discuss.
- Your other writing. Earlier essays or exam scripts that show the same voice and habits.
- A conversation. Offer to explain your argument, why you structured it as you did, and what you would change.
If you did use an AI tool in a way your course allows, for example to check grammar, say so plainly, show what it did, and cite it as your course asks.
How can I lower the risk before I submit?
Keep a record as you write, because you cannot build one afterwards. Write in a tool that saves version history. Keep your outline and drafts. Keep notes on what you read. Know your course's rules on AI tools, including grammar and paraphrasing tools, which some courses count as AI use.
Five things are worth doing before any paper goes in, and a checker can do only some of them:
| Before you submit | How | Where EdCitation does it |
|---|---|---|
| Keep a record of the writing | Write in Google Docs, or in Word saved to OneDrive; keep the outline and drafts | Nowhere: only your own files show this |
| Know what this assignment asks | Read the brief and the marking criteria | Check your paper reads the brief into a checklist (word limit, sections, style, number of sources), free |
| Know the rules on AI tools | Read the course and university policy, and ask the course leader if it is unclear | Nowhere: no tool can answer this for you |
| Make sure every source is real | Look each reference up in the record | Verify references, free, no account |
| See how a detector reads the paper | Run one yourself, as a warning light | Check AI, Pro, shown only to you |
Do not write for the detector. Changing good sentences to chase a score makes the essay worse and proves nothing about who wrote it. No wording can guarantee how any detector will score a text, and the evidence above is the reason: the same passage can be scored differently by different tools.
Can I check my paper with an AI detector before I submit?
Yes, and it helps only when the result is treated as a warning light, never as a verdict. EdCitation's Check AI marks every passage of your paper as AI-written, AI-assisted or human and gives the score behind that call; nobody sees the result but you, so you learn where a detector reacts before anyone else runs one. It comes with Pro ($8 a month, 240 credits included), you see what a check will cost before you start it, and a check that fails costs nothing. Pricing has the detail.
Be clear about its limits. It is an estimate, like every detector. A low score does not prove that you wrote the paper, a high score does not prove that you did not, and your university's own detector may read it differently. A high score on your own honest work tells you one thing: get your drafts and version history in order before you hand in. EdCitation never writes or rewrites any part of a paper, so it will not "fix" a flagged passage for you, and nothing should.
The check that is not an estimate
For the references, EdCitation is the best check there is, and the reason is the gap between an estimate and a record. A detector judges the wording, and the six detectors in the Perkins et al. (2024) test averaged 39.5% accuracy even on unaltered AI text. Verify references judges nothing: every entry is looked up in the publisher's record and no reference is ever written for you, something no chatbot can offer. Each is returned as verified, as "check this", or as not found; a retracted paper carries a flag; and an entry the record could not answer for is labelled that way, never as missing. It is free with no account; Pro ($8 a month) adds Check AI and Check plagiarism, and Max ($24 a month) adds Theoretics QA and the Library. Why AI tools invent references explains why fabricated sources appear in the first place.
Quick questions
Can an AI detector prove that I used AI?
No. An AI detector gives an estimate of how machine-like a text looks, not a record of how it was written. Turnitin says its own false positive rate is not zero and that the highlighted sentences are there to start a conversation rather than to settle one. The same holds for EdCitation's own Check AI: its labels are for the writer's information, not evidence of anything.
Are AI detectors biased against non-native English writers?
The best-known study found that they were. Liang et al. (2023) ran seven detectors on 91 human-written TOEFL essays and found an average false positive rate of 61.3%, while essays by US eighth-grade students were classified almost perfectly. Rewriting the TOEFL essays with a richer vocabulary cut the rate to 11.6%.
What is Turnitin's false positive rate?
Turnitin stated in 2023 that its document-level false positive rate is less than 1% for documents with 20% or more AI writing, and that its sentence-level rate is around 4%. These are the company's own figures, and the company notes that false positives are commoner in documents that mix human and AI writing.
Should I rewrite my essay until a detector passes it?
No. Rewriting to please a detector does not show who wrote the essay, and another detector may still disagree. Keep the drafts, version history and notes that show how you wrote it.
Why did OpenAI withdraw its AI classifier?
OpenAI withdrew its AI text classifier on 20 July 2023 because it was not accurate enough. In the company's own tests it caught 26% of AI-written text and wrongly labelled human-written text as AI-written 9% of the time.
Who has to prove what in an academic misconduct case?
In England and Wales, the Office of the Independent Adjudicator's Good Practice Framework says the burden of proof rests on the provider, normally on the balance of probabilities, and that the student should see the evidence and have a fair chance to respond. Your own institution's regulations govern your case, so read them.
References
- Chechitelli, A. (2023a). Understanding false positives within our AI writing detection capabilities. Turnitin. https://www.turnitin.com/blog/understanding-false-positives-within-our-ai-writing-detection-capabilities
- Chechitelli, A. (2023b). Understanding the false positive rate for sentences of our AI writing detection capability (14 June 2023). Turnitin. https://www.turnitin.co.uk/blog/understanding-the-false-positive-rate-for-sentences-of-our-ai-writing-detection-capability
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), Article 100779. https://doi.org/10.1016/j.patter.2023.100779
- Office of the Independent Adjudicator for Higher Education. (n.d.). Good Practice Framework: Good disciplinary procedures. https://www.oiahe.org.uk/resources-and-publications/good-practice-framework/disciplinary-procedures/good-disciplinary-procedures/
- OpenAI. (2023, January 31). New AI classifier for indicating AI-written text. https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
- Perkins, M., Roe, J., Vu, B. H., Postma, D., Hickerson, D., McGaughran, J., & Khuat, H. Q. (2024). GenAI detection tools, adversarial techniques and implications for inclusivity in higher education [Preprint]. arXiv. https://arxiv.org/abs/2403.19148
- Turnitin. (2024, April 9). Turnitin marks one year anniversary of its AI writing detector with millions of papers reviewed globally [Press release]. https://www.prnewswire.com/news-releases/turnitin-marks-one-year-anniversary-of-its-ai-writing-detector-with-millions-of-papers-reviewed-globally-302111754.html
- Vanderbilt University. (2023, August 16). Guidance on AI detection and why we're disabling Turnitin's AI detector. https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/
- Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, Article 26. https://doi.org/10.1007/s40979-023-00146-z