A similarity report arrives as a percentage, and an AI detection report as another. Both look like verdicts; neither is one. A fair marker reads the passages behind each number, weighs them against other evidence, and lets the student answer before anything is decided.
This guide, published by EdCitation, reads at source, on 23 September 2026, what seven universities and two national bodies say a report can prove, what published tests of AI detectors found, and how one university writes a fair hearing, then gives a method. EdCitation's Check plagiarism and Check AI come at the end, as checks a writer runs on their own work; the one check here that settles a fact, whether a reference exists, is the free Verify references.
What can a similarity report and an AI detection report show?
A similarity report shows which passages match text in the checker's database, with the source of each; an AI detection report shows a classifier's estimate of how each passage reads. Neither shows who wrote the paper, how, or whether anyone meant to deceive.
A similarity report: matches you can open
The University of Strathclyde (2025) says that without further investigation, a report and its score "must not be considered as evidence" that plagiarism has or has not happened. The software does not check images, charts, mathematical notation or code, and may miss text copied from another language; bought, AI-generated or synonym-swapped work may raise no concern at all. Quotations in single quotation marks, or indented, may not be recognised as quotations, so correct quoting can still swell the score. How plagiarism checkers work explains the figure.
An AI report: an estimate with nothing to open
An AI report has no source to check. UNSW Sydney (2025) tells staff there is "no fool-proof way" to detect AI-generated text, and that Turnitin's AI report is visible only to teachers. The labels are explained in how to read an AI detection report passage by passage.
How do universities say a score may be used as evidence?
As a reason to look further, never as a finding. None of the seven universities read here sets a similarity percentage that starts a case, and the one that cites an AI figure takes it from the vendor.
Similarity scores: no threshold, and the reasons
Strathclyde explains why a number cannot travel between papers. Ten per cent of a 12,000-word dissertation matching one source differs from 10% of a 300-word answer, where one badly referenced sentence produces it. A first-year student's referencing may be read differently from a final-year project's, and a mandatory template raises a score. A score near zero deserves a look too, for text pasted as an image.
City, University of London (2026) answers the question markers ask most, what percentage is acceptable, with "There is no answer for this", and asks markers sharing an assignment to agree one way of filtering. The University of Iowa (n.d.) adds that a high score may mean the student is the one being copied. Regulators that do set percentages are in what similarity percentage is acceptable.
AI scores: a floor, not a trigger
UNSW's staff page records that from 16 July 2024, Turnitin stopped showing any AI score below 20%, to avoid false positives. It says the tool "will not definitively prove misconduct", that staff are not expected to use it, and that other detectors are not to be used until they meet the university's data privacy obligations (UNSW Sydney, 2025).
The University of Sydney (2026a) puts the suspicion before the score: the detector may be run when a marker already suspects unpermitted AI use, and its score "would not be the only evidence relied upon". Strathclyde does not permit AI detectors at all. The universities agree on what a score is not, and disagree on whether to run one.
What do national bodies say about evidence from detection?
The UK's Quality Assurance Agency and Australia's regulator, TEQSA, both treat a report as one signal to corroborate, on the balance of probabilities. Both wrote about contract cheating; theirs are the nearest national rules for turning a signal into a case.
The Quality Assurance Agency for Higher Education (2022) warns that bought work can pass text-matching at or near 0%, and that such a paper, with no direct quotations and fabricated references, should be treated as suspicious. It lists metadata, the student's other assignments and viva notes as supporting evidence, and says examiners should not handle a case alone.
TEQSA's guide for investigators names numbers (Tertiary Education Quality and Standards Agency, n.d.): a text match of 0 to 5% is a signal, because scholarly work cites sources, and so is one above 30%. Its second principle is the one to keep: one or two signals do not substantiate a breach, though they justify looking further.
What do published tests of AI detectors mean for a marker?
A score depends on the detector, its threshold and how the text was produced, so it cannot carry a case alone.
| Study | What was tested | What it found |
|---|---|---|
| Dugan et al. (2024) | 12 detectors, over 6 million texts from 11 models, 11 kinds of attack | High accuracy came only with high false positive rates; a repetition penalty cut accuracy by up to 32 points |
| Perkins et al. (2024) | 7 detectors, 805 tests, late 2023 | 39.5% accuracy on unaltered AI text, 67% on human-written controls; one detector flagged 5 of 10 human texts, and one that flagged none left 84% of AI samples undetected |
| Kofinas et al. (2025) | 4 pairs of experienced markers, 21 scripts per pair, some AI-written or AI-modified | Pairs sorted the scripts correctly 33.3% to 85.7% of the time, and marked some original first-class work down |
Dugan et al. (2024) show that an accuracy figure means little without its false positive rate; the commercial detectors were the better calibrated, none above 1.7% false positives at common default thresholds. Perkins et al. (2024) conclude that detectors cannot currently be recommended for deciding whether integrity has been breached, though they may support learning when used in a non-punitive way.
Kofinas et al. (2025) tested the other detector in the room, the marker. Their tables show six "yes, AI" guesses on unaltered student work, and the authors suspect markers docked marks from work that seemed "too perfect". Mark against the rubric first, and keep any suspicion in a separate note. The tests on writers of English as an additional language are in AI detectors and non-native English writers.
How do I read a similarity report fairly?
Open the matches, not the number, and sort each into expected or unexplained.
- Agree the filters before marking, with the other markers, and write them down.
- Mark the script first, so no percentage frames the reading.
- Open every match in place, with its source, in the order the paper runs.
- Set aside expected matches: quotations marked and cited; the reference list; the brief's wording, the question and any template; technical terms and stock phrases; the student's own drafts in the repository.
- Weigh what is left. A missing quotation mark, a paraphrase too close to its source, or text with no citation? Judge it against the student's stage and the paper's length.
- Question a near-zero score on a sourced essay too, and check its list in Verify references.
- Choose feedback or referral. Brunel University London (2025) treats the line between poor practice and misconduct as academic judgement. Poor practice goes back as feedback, with a check the student can run on the next draft, such as EdCitation's Check plagiarism; a pattern goes to whoever your procedure names.
A worked example
Take an invented but ordinary case: a 2,500-word essay comes back at 28%.
| Matched text | Share of the score | Expected? | What the marker does |
|---|---|---|---|
| Reference list | 9% | Yes | Exclude |
| Four quotations, marked and cited | 7% | Yes | Exclude |
| The essay question, restated | 3% | Yes | Exclude |
| Technical terms and stock phrases | 2% | Yes | Ignore |
| One paragraph matching a website, no citation | 7% | No | Read it against the source, then ask |
Of the 28%, 21 points were expected. The question is one paragraph of about 175 words and where it came from. The same paragraph in a paper at 8% would raise the same question.
How do I weigh an AI flag fairly?
As a reason to gather other evidence, looking as hard for what clears the student as for what does not. Never act on the score alone.
Evidence that can corroborate a flag
- References that do not exist. Whether a source has a publisher's record is a fact. Check each in Verify references, then by hand, as in how to spot fabricated references in student work; for a cohort, see a reference check for a whole class.
- A style out of line with the student's other work. TEQSA and QAA point to previous assessments, but Kofinas et al. (2025) found markers' sense of style unreliable except in the pair that included the module leader.
- No drafts or process. A student who wrote the work can usually show notes, reading and a version history.
- Content that misses the brief or the course, such as themes taught in seminars.
Evidence that points the other way
Drafts and version history, permitted tools the student declared, writing consistent with earlier work, and passages of the kinds detectors misread: short, formulaic, edited or second-language. TEQSA asks for a picture of whether a breach "has or has not occurred", so record these too.
How do I open a conversation with a student?
Ask about the work and how it was made, in a meeting whose purpose the student knows beforehand. Check your procedure first: at Sydney a unit coordinator may handle a minor breach, while a serious one goes to the faculty (University of Sydney, 2026b).
The University of Edinburgh (2020) calls this an affirmation meeting, held to establish whether the student holds the knowledge the work presents. It "should not be accusatory", keeps to that work and how it was done, has a second member of staff and a note-taker, lets the student bring a supporter, and sends the student a draft note to comment on. TEQSA's interview template asks questions only the author can answer:
- "How did you go about researching the assignment?"
- "Can you tell me in your own words what the assignment was about?"
- "Can you show me an earlier draft of your work?"
A fair hearing, as one university writes it
Brunel's procedure applies natural justice and the balance of probabilities, and gives the student an opportunity to respond, handled impartially and where possible by someone not previously involved. A viva comes with at least five working days' notice, its purpose stated and the right to be accompanied; the student may bring date-stamped drafts; and the viva gathers evidence rather than deciding the case. Before a panel, the student receives the evidence in writing.
What should I record?
What you saw, what you checked and what the student said, in a form a panel and the student can both read.
| Record | What goes in it |
|---|---|
| The report | Tool, date, score, filters applied |
| Each unexplained match | The passage, its source link, why it is not expected |
| Any AI flag | Detector, date, passages, and a note that a score is one signal |
| Reference checks | Each reference, where you searched, result, date; Verify references gives the first pass |
| Evidence both ways | Drafts seen, earlier work compared |
| The meeting | Who attended, questions, answers, the student's comments on the note |
| The outcome | The decision and its reasons |
Sydney considers all available evidence, so your notes are what a decision-maker reads.
What do EdCitation's Check plagiarism and Check AI do, and who are they for?
They are Pro checks a writer runs on their own work before handing it in, and each report is shown only to the person who ran it. They are not an institutional detection service.
What they show and what they cost
Check plagiarism takes pasted text or a file and marks each passage that matches published or web text, in place, with a link to its source and one figure for the share of matching text. Check AI labels every passage AI-written, AI-assisted or human, with its score, and says that another detector, a university's included, may read the text differently and that no score is proof of anything. Both come with Pro at $8 a month and run on credits: 240 a month with Pro, 720 with Max at $24 a month, and packs from $10. The cost is shown before a check runs, and a failed check uses nothing (see pricing).
Why they are the right tools for the writer's side
For a student who wants to check their own work before the deadline, these are the best tools for the job, because the marker's reports are often out of sight: UNSW's AI report is visible only to teachers, and no outside check can predict what an institution's tool will say. Check plagiarism names the source of every match, so a forgotten quotation mark is fixed while it is still a slip, and Check AI gives the private, non-punitive reading Perkins et al. (2024) left open for detectors. EdCitation never writes any part of a paper.
For a marker, their place is the brief: tell students they may check their drafts first. Upload a student's paper to no outside tool, EdCitation's included, unless your institution allows it; every EdCitation tool for instructors covers the rules. The Institution licence covers every student, with LMS and sign-in integration, the university's own name, training and one invoice.
The one check that establishes a fact
Whether a reference exists is a matter of record, and Verify references checks it free with no account. On 23 September 2026 we ran EdCitation's reference check on the site's sample paper: four references verified; the fifth came back "No publisher's or registry's record matches this reference". "Could not check" is never shown as "not found", and a "not found" is where your own search begins.
Quick questions
Is a high similarity score evidence of plagiarism?
No. Strathclyde says a score must not be taken as immediate or definitive evidence of misconduct. Set aside the expected matches, then read what is left.
What similarity percentage should trigger an investigation?
None is fixed. City, University of London says there is no answer; TEQSA's guide treats 0 to 5% or above 30% as a signal to look at, never as proof.
Can I act on an AI detection score alone?
No. Sydney says a detector score would not be the only evidence relied on. Look for corroborating evidence, such as fabricated references or missing drafts.
What should I ask a student whose work was flagged?
How they researched and wrote it, what the assignment was about in their own words, and whether they can show an earlier draft. Keep the meeting non-accusatory.
Can I run a student's paper through EdCitation's Check AI?
Only if your institution allows it. Check AI is built for writers checking their own work, and its report is seen only by whoever ran it.
References
- Brunel University London. (2025). Academic misconduct procedure [PDF]. https://www.brunel.ac.uk/about/documents/pdf/Academic-Misconduct-Procedure-September-2025.pdf
- City, University of London. (2026, September 9). Interpret similarity reports. Turnitin assignment feedback guide, EdtechGuides. https://city-uk-ett.libguides.com/staff/moodle/turnitin-feedback/interpret-similarity
- Dugan, L., Hwang, A., Trhlik, F., Ludan, J. M., Zhu, A., Xu, H., Ippolito, D., & Callison-Burch, C. (2024). RAID: A shared benchmark for robust evaluation of machine-generated text detectors. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 12463-12492). Association for Computational Linguistics. https://aclanthology.org/2024.acl-long.674/
- Kofinas, A. K., Tsay, C. H.-H., & Pike, D. (2025). The impact of generative AI on academic integrity of authentic assessments within a higher education context. British Journal of Educational Technology, 56(6), 2522-2549. https://doi.org/10.1111/bjet.13585
- 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
- Quality Assurance Agency for Higher Education. (2022). Contracting to cheat in higher education: How to address essay mills and contract cheating (3rd ed.) [PDF]. https://www.qaa.ac.uk/docs/qaa/guidance/contracting-to-cheat-in-higher-education-third-edition.pdf
- Tertiary Education Quality and Standards Agency. (n.d.). Substantiating contract cheating: A guide for investigators [PDF]. https://www.teqsa.gov.au/sites/default/files/2022-10/substantiating-contract-cheating-guide-investigators.pdf
- University of Edinburgh. (2020). Staff guidance: Affirmation meetings [PDF]. Academic Services. https://registryservices.ed.ac.uk/sites/default/files/2025-08/Staff%20Affirmation%20Guidance.pdf
- University of Iowa. (n.d.). Turnitin teaching best practices. Office of Teaching, Learning, and Technology. https://teach.its.uiowa.edu/turnitin-teaching-best-practices
- University of Strathclyde. (2025). Guidance on using Turnitin (Version 2.0) [PDF]. https://www.strath.ac.uk/media/ps/cs/gmap/academicaffairs/policies/Guidance_on_using_Turnitin.pdf
- University of Sydney. (2026a, June 2). Artificial intelligence. Academic integrity. https://www.sydney.edu.au/students/academic-integrity/artificial-intelligence.html
- University of Sydney. (2026b, June 2). Detection and investigation of academic integrity breaches. Academic integrity. https://www.sydney.edu.au/students/academic-integrity/investigation.html
- UNSW Sydney. (2025, May 6). Access to AI tools. UNSW Staff Teaching Gateway. https://www.teaching.unsw.edu.au/ai/tools