An AI detection report comes back as one number and a page of coloured passages. The number looks like a verdict and the colours like accusations. Neither is. Each is an estimate of how the words read, made by a program that never watched you write, and the useful part is the passage-by-passage detail, read with its limits in mind.
This guide, published by EdCitation, explains the meaning of the labels, why one percentage for a paper misleads, which passages detectors misread, and what to do with a flagged one; every source was read on 23 September 2026. It ends with EdCitation's Check AI, part of Pro, which shows you every passage labelled before you hand in. The part of a paper that can be checked for certain, its references, goes through the free Verify references.
What does an AI detection report actually show?
An AI detection report shows a classifier's estimate of how closely each passage, and the paper as a whole, resembles text a language model would produce. It does not show who wrote the text, how, or with what help.
A guide by Guy Curtis in the academic integrity toolkit of TEQSA, Australia's higher education regulator, says detectors compare features such as perplexity, burstiness and sentence structure with patterns in human and machine text, producing "a probability estimate", and that people can write with the very features a detector associates with AI (Curtis, 2026).
A plagiarism report links a passage to a source you can open; EdCitation's Check plagiarism names the source of every match. An AI detector has no source to show. Weber-Wulff et al. (2023), who tested 14 detectors on 54 documents, found their verdicts could not be verified, so a student accused on that basis alone would have nothing to answer.
Why a score is a probability, not proof
A score is not the chance that you used AI. TEQSA's guide does the arithmetic: a detector with a 1% false positive rate gives about one essay in a hundred a high score, so in a class of 100 where nobody used AI, one essay still scores high and the real probability that it was AI-generated is zero (Curtis, 2026).
OpenAI (2023), which withdrew its own classifier in July 2023 for its low accuracy, warned that it could label human writing AI "incorrectly but confidently". Scores also move: Weber-Wulff et al. (2023) uploaded some documents twice and got different values, from the tools' own inconsistency.
What do AI-written, AI-assisted and human mean on a report?
The labels describe how each passage reads, not what you did. In EdCitation's Check AI, every passage is labelled AI-written, AI-assisted or human, in place in your text, with the score behind the label beside it.
AI-assisted and AI-written differ in a policy
Policies draw the line by what the tool did. The European Network for Academic Integrity calls tools that only influence a text's form, such as proofreaders, spelling checkers and a thesaurus, generally acceptable, and says any AI tool used should be acknowledged (Foltynek et al., 2023). The Quality Assurance Agency (2023a), the UK's quality body for higher education, describes "hybrid submissions" in which AI assisted: with initial ideas, grammar and spelling, or cutting text to meet a word limit. Whether that is allowed is decided by your course, not by a detector.
Why a detector cannot measure how much help there was
A detector reads the finished text, not the process. Jisc's National Centre for AI says even an accurate detector cannot tell whether AI wrote the whole text or only helped with phrasing and clarity (Webb, 2023).
Saha and Feizi (2025) tested it. Five language models polished 300 human-written texts (blogs, emails, reviews, news, abstracts, speeches) to set degrees, and 12 detectors read about 15,000 results. Every detector flagged more lightly polished texts than untouched originals; minimal polishing with GPT-4o was flagged at rates from 10% to 75%, depending on the detector; and most barely told minor polishing from major. So a paragraph tidied with a grammar tool your course allows can be flagged, and a human label does not certify that no tool was used.
Why can an AI detection percentage for the whole paper mislead?
A percentage for the whole paper hides where the flagged text sits, what kind of text it is, and what the tool counted.
- What was counted. Turnitin's figure covers only prose sentences in long-form writing. It does not detect AI text in bullet points, tables or annotated bibliographies, so a document mixing kinds of writing shows a gap between percentage and highlights; below 20% it shows no number, because false positives are commoner there (Turnitin, n.d.).
- How it was counted. Products often calculate and display their figures differently, by sentence, paragraph or document, which makes them hard to compare (Webb, 2025).
- What an average hides. Weber-Wulff et al. (2023) called their own overall accuracy figure misleading because it hid the spread between kinds of document: 96% on human writing, 20% lower on human writing machine-translated into English, 42% on hand-edited AI text against 74% on unedited, and 26% on machine-paraphrased AI text.
Check AI's figure is the share of the text that does not read as confidently human. It is not the share you wrote with a tool, nor the chance that you used one. Whether it comes from one long paragraph or a sentence in every section changes what you should do, and only the passage labels show which.
Which passages do AI detectors misread most often?
The sources name five hard cases: short passages, lists and formulaic writing, edited or mixed text, machine-translated text, and writing in a second language.
| Kind of passage | What the sources found | What to do with its label |
|---|---|---|
| Short | OpenAI's classifier was "very unreliable" below 1,000 characters; Turnitin needs 300 words of prose | Weigh it least |
| Lists, tables, bibliographies | Not detected by Turnitin; very predictable text cannot be reliably identified | For a reference list, check what matters, whether each entry exists, in Verify references |
| Formulaic or scientific | Accuracy fell from humanities to science: 0.86 to 0.51 (Turnitin), 0.96 to 0.58 (Originality) | Reread it; standard wording is often right |
| Edited or mixed | Both tools poor on half-and-half texts; sentence errors commonest at the join | Match it to your drafts |
| Translated or second-language | 61.3% average false positive rate on non-native essays in one test | Keep drafts and translation notes |
Short passages, and a disagreement
Short text is where even vendors stop trusting their tools: OpenAI (2023) said its classifier grew more reliable with length, and Turnitin needs 300 words of prose. TEQSA's guide lists documents under about 300 words among the cases where detectors are much less reliable (Curtis, 2026).
Longer is not simply safer. Hadra et al. (2026), testing Turnitin and Originality on 192 texts, found both did best on their shortest texts, of 300 to 330 words; but that group held only 23 texts, none under 300 words. The fair reading: below a few hundred words a label deserves little weight, and above that, length does not rescue a score. The University of Melbourne (n.d.) tells staff not to rely on Turnitin's sentence-by-sentence judgements. A passage label tells you where to reread, not what happened.
Lists, formulaic and technical writing
Predictable text gives a detector little to work with. OpenAI (2023) gave the extreme case: a list of the first 1,000 prime numbers reads the same whoever writes it. Academic writing has milder versions: a methods section, inclusion criteria, a standard definition. Melbourne warns staff that formulaic disciplines, and students whose style is regular or routine, may raise the chance of a false positive, and Hadra et al. (2026) suggest technical terms and formulaic sentences explain why science texts fared worse.
Edited and mixed passages
Mixed writing is common and the hardest case. Weber-Wulff et al. (2023) found hand-edited AI text almost undetectable, and Hadra et al. (2026) found both commercial detectors poor on texts half student writing and half AI. Turnitin puts its sentence-level false positive rate at around 4%, commonest where human and AI writing meet, with 54% of those sentences right next to actual AI writing (Chechitelli, 2023). A sentence you wrote, beside a passage a tool helped with, can share its label.
Writing in a second language, or translated
A smaller vocabulary produces plain, predictable English, which some detectors read as machine-made. Liang et al. (2023) ran seven detectors on 91 TOEFL essays by non-native writers and 88 essays by US eighth-grade students. The US essays were classified almost perfectly; the TOEFL essays drew an average false positive rate of 61.3%, which richer vocabulary cut to 11.6%. Weber-Wulff et al. (2023) warned that students who machine-translate their own writing risk false accusation.
The evidence is not uniform. Hadra et al. (2026) found no significant difference for Turnitin between writing in English as a foreign language and professional writing, and only a borderline trend for Originality. The risk is documented for some tools, not shown for all.
How do universities tell staff to treat an AI detection score?
As a reason to look further, never as proof on its own; they differ on whether to run a detector at all.
- University of Melbourne (n.d.). Look only at work where more than 20% is predicted AI; the report is "not enough on its own", and a second piece of evidence, such as false references or a gap from earlier work, is required. Staff may ask for drafts and notes. Students do not see Turnitin's percentage or highlights.
- University of Dundee (2026). Opted out of Turnitin's AI detection, tells lecturers not to use unauthorised detectors, and has no student consent to upload work to third-party sites.
- TEQSA's guide. The "AI score" alone, high or low, is insufficient for an allegation, and investigators should seek evidence that disconfirms AI use as well as evidence that confirms it (Curtis, 2026).
- Quality Assurance Agency (2023b). AI output "cannot reliably be detected", and a long-term approach to assessment avoids the need for unreliable detection software.
- Turnitin (n.d.) itself. Its model may misidentify text and should not be the sole basis for adverse action against a student.
What that means for the writer
The score your university sees is often hidden from you, so to know which passages a detector reacts to, you have to look yourself. The evidence staff are told to seek (drafts, notes, real references) is evidence you can have ready. Dundee's concern is staff uploading students' work; checking your own paper differs, but read your university's rules on outside tools first.
What should I do with a passage flagged as AI?
Reread it, check it against your drafts and notes, and declare any use your course allows; do not rewrite honest work to chase a lower score.
- Read the label with its score. In Check AI each passage carries its label in place. Note what kind of passage it is: short, a list, a definition, methods, translated or grammar-checked.
- Reread it as your marker would. If it says what you mean, in your words, leave it.
- Find its history: earlier drafts, your software's version history, the notes and sources behind it.
- Check its references. TEQSA's guide says fabricated references are typically misconduct in their own right (Curtis, 2026). Run the paper through Verify references, free.
- Declare any tool you used, as your course asks. How to declare AI use in an assignment gives the wording.
- If the passage is weak, rewrite it yourself in the same file, so the history records the change. Never use a "humanising" tool: TEQSA names them as a way of bypassing detection.
- If your university has flagged it, follow what to do if an AI detector wrongly flags your writing.
How does EdCitation's Check AI show your paper?
EdCitation's Check AI shows every passage of your paper labelled AI-written, AI-assisted or human, in place in your own text, with the score behind each label, and shows that report to you and nobody else.
What you give it, and what comes back
Paste your text or choose the file: a section or the whole paper. The report gives one figure for the paper, the share of the text that does not read as confidently human; every passage labelled in place with its score; and the passages to look at again, those marked AI-written or AI-assisted.
What it costs
Check AI is part of Pro, $8 a month, and runs on credits: Pro includes 240 a month and Max, at $24 a month, 720, and more can be bought from $10. The cost of a check is shown before it runs, a check that fails uses nothing, and the same text checked again within a day uses nothing. See pricing.
Why it is the right tool for this job
For a writer who wants to see, before the deadline, what a detector makes of each passage, Check AI is the best tool for the job, because the university's own reading is often hidden: Melbourne's students, for one, do not see Turnitin's percentage or highlights. Check AI gives the writer that reading privately, a label on every passage as well as one number. For a flagged passage its advice is to reread it and keep your drafts, where one detector in the Weber-Wulff et al. (2023) test told users to keep editing until less AI was detectable; EdCitation never writes any part of a paper. It also says plainly what it cannot do: predict your university's detector, which may read the same text differently, or turn any score into proof.
Before submitting, pair it with Check plagiarism, also Pro, on the same credits, which marks passages matching published or web text, the check a dissertation needs most. Check your paper, which reads your brief into a checklist, and Verify references are free with no account. The complete checklist puts every check in order.
Quick questions
Is an AI detection score proof that I used AI?
No. A score is a classifier's estimate of how the text reads. Turnitin says its score should not be the sole basis for action against a student, and TEQSA's guide says no score, high or low, is enough evidence by itself.
What does AI-assisted mean on an AI detection report?
It is a label for how a passage reads, not a record of what you did. Saha and Feizi (2025) found detectors flag even lightly AI-polished human writing and barely distinguish light polishing from heavy.
Why do two AI detectors give my paper different percentages?
They count differently, by sentence, paragraph or document, and some leave lists and tables out, as Jisc's National Centre for AI and Turnitin's own guide note.
Can a list or a short paragraph be flagged as AI?
Yes, and those labels deserve the least weight. OpenAI said its classifier was very unreliable below 1,000 characters and on highly predictable text, and Turnitin says its model does not detect AI text in bullet points or tables.
Who sees my Check AI report?
Only you. EdCitation's Check AI, part of Pro, shows the report only to the person who ran it, and it is not your university's detector.
References
- Chechitelli, A. (2023, June 14). Understanding the false positive rate for sentences of our AI writing detection capability. Turnitin. https://www.turnitin.com/blog/understanding-the-false-positive-rate-for-sentences-of-our-ai-writing-detection-capability
- Curtis, G. J. (2026). Detecting plagiarism of AI-generated text in student assessments and securing take-home written assessments. Tertiary Education Quality and Standards Agency. https://www.teqsa.gov.au/guides-resources/protecting-academic-integrity/academic-integrity-toolkit/risks-academic-integrity-ai/detecting-plagiarism-ai-generated-text-student-assessments-and-securing-take-home-written-assessments
- Foltynek, T., Bjelobaba, S., Glendinning, I., Khan, Z. R., Santos, R., Pavletic, P., & Kravjar, J. (2023). ENAI recommendations on the ethical use of artificial intelligence in education. International Journal for Educational Integrity, 19, Article 12. https://doi.org/10.1007/s40979-023-00133-4
- Hadra, M., Cambridge, K., & Mesbah, M. (2026). Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity, 22, Article 4. https://doi.org/10.1007/s40979-026-00213-1
- 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
- OpenAI. (2023, January 31). New AI classifier for indicating AI-written text. https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
- Quality Assurance Agency for Higher Education. (2023a, May 8). Maintaining quality and standards in the ChatGPT era: QAA advice on the opportunities and challenges posed by generative artificial intelligence. https://www.qaa.ac.uk/docs/qaa/members/maintaining-quality-and-standards-in-the-chatgpt-era.pdf
- Quality Assurance Agency for Higher Education. (2023b, July). Reconsidering assessment for the ChatGPT era: QAA advice on developing sustainable assessment strategies. https://www.qaa.ac.uk/docs/qaa/members/reconsidering-assessment-for-the-chat-gpt-era.pdf
- Saha, S., & Feizi, S. (2025). Almost AI, almost human: The challenge of detecting AI-polished writing. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 25414-25431). Association for Computational Linguistics. https://aclanthology.org/2025.findings-acl.1303/
- Turnitin. (n.d.). Using the AI Writing Report. Turnitin Guides. https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report
- University of Dundee. (2026). AI (artificial intelligence) in teaching and assessment. https://www.dundee.ac.uk/guides/ai-artificial-intelligence-teaching-and-assessment
- University of Melbourne. (n.d.). Turnitin's AI writing detection tool. Academic Integrity. https://academicintegrity.unimelb.edu.au/staff-resources/turnitins-ai-writing-detection-tool
- Webb, M. (2023, September 18). AI detection: Latest recommendations. Jisc National Centre for AI. https://nationalcentreforai.jiscinvolve.org/wp/2023/09/18/ai-detection-latest-recommendations/
- Webb, M. (2025, June 24). AI detection and assessment: An update for 2025. Jisc National Centre for AI. https://nationalcentreforai.jiscinvolve.org/wp/2025/06/24/ai-detection-assessment-2025/
- 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