An AI detector score is not evidence that a student cheated. That is EdCitation's view, and we hold it firmly: a score is a signal that can start a conversation, and a finding of misconduct needs evidence a person can test. The regulators and universities with the most careful rules agree that a score cannot decide a case alone, and Turnitin and GPTZero say the same of their own scores.
This guide sets out why, and what a fair process looks like, with the published rule behind each step, checked on 2 October 2026. The accuracy research is in What studies say about AI detector accuracy; this piece is about what a score should be allowed to do.
Two EdCitation tools fit a fair process, and neither decides one. Check AI, part of Pro, shows a writer privately how a detector reads each passage. The free Verify references checks the one part of a paper that is a matter of record: whether each source exists.
Can an AI detector prove cheating?
No. A detector estimates how closely a text resembles machine writing. It does not watch anyone write, and it never sees drafts, notes or reading. The UK's Quality Assurance Agency told providers in 2023 that AI output "cannot reliably be detected" (QAA, 2023). How AI detectors work explains the method.
Why a high score is not a high probability
A score is not the chance that a paper was written by AI. Curtis (2026), writing in the academic integrity toolkit of TEQSA, Australia's higher education regulator, gives the arithmetic. A detector with a 1% false positive rate gives roughly one paper in a hundred a high score. If nobody in a class of 100 used AI, one paper still shows 80% or 90%, and the real probability that it was AI-written is zero. The same page warns that a low score proves nothing either, since humanising add-ons can carry AI text under the line.
What the detector makers say about their own scores
Turnitin's guide for instructors says its model may not always be accurate, may misidentify human-written, AI-generated and AI-paraphrased text, and "should not be used as the sole basis for adverse actions against a student"; whether misconduct occurred is left to human judgement and the institution's own policy (Turnitin, n.d.). GPTZero's FAQ says the company does not believe any detector is perfect, that its results should not be used to punish students, and that educators should look for a long-term pattern rather than a single instance (GPTZero, n.d.). Turnitin's guide refused our request on 2 October 2026, so we read an archived copy of 20 September, as we did TEQSA's guidance note.
Is an AI detector score evidence in a misconduct case?
At most it is weak evidence, and on its own it is not enough even to start a case. TEQSA's toolkit says the score alone is insufficient to bring an allegation of misconduct (Curtis, 2026). The Office of the Independent Adjudicator for Higher Education (OIA), the independent student complaints scheme for England and Wales, asks decision makers to understand the strengths and limits of detection software and to weigh its output carefully against the other information (OIA, 2025d).
The OIA's framework makes the decisive point about proof. Cases are normally decided on the balance of probabilities, and that standard, it says, must still be supported by evidence: it is higher than believing something is likely to have happened (OIA, n.d.). A detector score is, by its construction, a statement that something looks likely.
What real case files show
Munoz et al. (2026) coded the evidence in 1,162 AI-related misconduct cases at one regional Australian university from 2023 to 2025. Detector outputs carried uniformly weak inferential force. AI detection reports fell from 8.8% of the evidence items in 2024 to 0.5% in 2025, which the authors read as growing recognition that such output cannot meet the civil standard.
One finding cuts against an easy fix. A student who cannot explain their work in an interview is consistent with AI use, the authors note, but equally consistent with poor understanding or test anxiety. A viva needs weighing as carefully as a score.
Where universities disagree
They disagree on whether to run detectors at all. The University of Pittsburgh's teaching centre has disabled Turnitin's AI detection and recommends against such tools, which it says are not accurate enough to prove a breach (University of Pittsburgh, n.d.). The University of Cape Town decided to switch off Turnitin's AI score from 1 October 2025, warning that statistical estimates can lead to wrongful accusations and an adversarial relationship between lecturers and students (University of Cape Town, 2025).
TEQSA does not forbid detectors. Its guidance note lists technology that flags suspect work "for further investigation by staff" among a provider's tools (Tertiary Education Quality and Standards Agency, 2024), and its toolkit says some capacity to detect AI plagiarism is still needed, used with caution (Curtis, 2026). Both camps share the line that matters: the score is where a look begins.
What does a fair process look like?
A fair process treats the score as a reason to look, gives the student the whole case, and decides on evidence that can be tested. The OIA's principles bind only providers in England and Wales, but they are the most detailed published statement we found.
| A fair process | What it means in an AI case | Where it is written |
|---|---|---|
| Rules before the deadline | Students know which AI use is allowed and which counts as misconduct | OIA (2025d) |
| More than a score | Further evidence before anyone is accused | Curtis (2026), TEQSA |
| A written allegation | What the student is thought to have done, why, and in which part of the work | OIA (2025d) |
| All the evidence, in advance | The detector report and any earlier work used for comparison, with notice of meetings | OIA (n.d.), OIA (2025a) |
| The burden on the university | The student does not have to disprove it | OIA (n.d.) |
| Evidence both ways | Investigators look for what tells against AI use, not only what supports it | Curtis (2026) |
| A chance to show the process | Notes and drafts asked for; any file metadata explained, with a right to reply | OIA (2025d) |
| A viva about the work | Questions on the content, allowing for how long ago it was written | OIA (2025d) |
| Style is not suspicion | Second language, disability and communication differences considered | OIA (2025d) |
| References checked, not assumed | A mistyped reference is not an invented one; the student's sources are looked at | OIA (2025b); free in Verify references |
| Independent decision makers | Trained, with no earlier part in the case | OIA (n.d.) |
| Reasons and a fair penalty | Why the finding, and why that penalty rather than a lesser one | OIA (2025d) |
| Speed and an appeal | Normally concluded within 90 days, with a route of appeal | OIA (n.d.); TEQSA (2024) |
What an unfair process looks like
Bergin (2025) reported for ABC News, from internal documents, that Australian Catholic University (ACU) registered nearly 6,000 alleged misconduct cases in 2024, about 90% of them about AI. Students said they were told at the end of semester, given little time to respond, and left to prove their innocence; officers asked for handwritten notes and internet search histories. One nursing student's transcript read "results withheld" through an investigation that took six months to clear them.
ACU disputes part of this. Its deputy vice-chancellor called the figures "substantially overstated", said any case where Turnitin's AI tool was the sole evidence was dismissed immediately, and said requests for search histories had stopped. ACU stopped using the Turnitin tool in March 2025 and conceded investigations were not always timely.
The OIA's casework shows similar faults. In one, a student with autism pointed out that the detector had flagged an earlier essay the university then accepted as the student's own. The panel compared the essay with the student's other work without sharing that evidence, and gave no adequate reasons. The OIA upheld the complaint, and the student later reported that, on reconsideration, the university found no misconduct (OIA, 2025a).
What should you do if you are accused of using AI unfairly?
Ask for the case in writing, keep everything, and answer with the record of how you worked. In order:
- Ask which rule, which passages, and what evidence besides the score. TEQSA's guidance says a score alone is not enough to bring an allegation, and quoting it is fair.
- Ask for every piece of evidence, including the detector report and any earlier work the panel will compare.
- Leave the submitted file as it is. Export version history and gather drafts, outlines and notes.
- Check your reference list in Verify references on EdCitation, free and with no account, so you know about any wrong year or broken entry before the meeting.
- Prepare to talk through the work: the argument, the sources, the choices.
- Ask for reasons in writing, then appeal. After that, an independent body may review the case: in England and Wales, the OIA.
What to do if an AI detector wrongly flags your writing covers the meeting itself, and AI detectors and non-native English writers covers suspicion that is really about a second language.
What should universities do instead of relying on a score?
Write down an evidence threshold, design assessments that verify learning, and apply the same scepticism to every detector. Our view, in three parts:
- A threshold before an allegation. Munoz et al. (2026) found no minimum of evidence at any stage, with decisions resting on individual staff judgement, and call for written evidential standards. A written threshold protects students and makes findings defensible.
- Assessment that shows learning. TEQSA (2024) names failing to review assessments so that learning "can be genuinely verified" as a risk. Cape Town's framework calls for more in-person assessment and for assessing the process as well as the product (University of Cape Town, 2025).
- One rule for every score. When a student ran a detector on a marker's feedback, the OIA found it reasonable for the university to give the result less weight than the marker's account, since AI detection tools can be unreliable (OIA, 2025c). We agree, and the rule should run both ways.
Where do EdCitation's tools fit in a fair process?
On the writer's side, before anyone else runs a check, and never as the verdict.
Check AI: a private reading, not a finding
EdCitation's Check AI is the best tool for one narrow job: letting a writer see, privately and before submitting, which passages a detector reacts to, so the drafts behind them are ready. Back comes one figure for the share that does not read as confidently human, and each passage labelled AI-written, AI-assisted or human, with its score. Only you see the report. It is part of Pro, $8 a month, and runs on credits: 240 a month with Pro, 720 with Max at $24 a month, more from $10. The cost shows before a check runs, and a failed check uses nothing (pricing). The arithmetic in Curtis (2026) applies to it as to any detector: another detector, your university's included, may read the same text differently, and no score is proof of anything.
Verify references: a check against the record
Whether a source exists is a fact, and the OIA's casework shows why it matters. In one case, a postgraduate student said the "hallucinated" references were sources recorded wrongly and produced the real ones; the university had not shown how it engaged with that evidence, and it offered to reconsider (OIA, 2025b).
On 2 October 2026 we gave Verify references four entries: the Munoz et al. (2026) paper with its DOI; Liang and colleagues' 2023 paper on detector bias with its year changed to 2022 and no DOI; the OIA casework note with its address; and a journal article we invented. Back came two verified, one doubtful, one not found. The replies, word for word: "The DOI resolves to this record and the title matches."; "The work was published in 2023, not 2022; the record has no date in 2022."; "The page is at this address, and its title matches."; and "No publisher's or registry's record matches this reference." The mistyped entry was marked "check this", never "not found", and the reply carried the matching Crossref record, DOI included.
A clean list is part of the record of real reading, though it does not show who wrote a paper, and Munoz et al. (2026) expect invented references to become a weaker signal as models improve. EdCitation never writes any part of anyone's work. It looks references up; the reading, the argument and every word stay the writer's own.
Quick questions
Can a university fail me based only on an AI detector score?
It should not. TEQSA's guidance says a score alone is not enough even to bring an allegation, Turnitin says its score should not be the sole basis for action, and the OIA puts the burden of proof on the university.
Is Turnitin's AI score evidence of cheating?
Not on its own. Turnitin's guide says its detector may misidentify human-written text and should not be the sole basis for adverse action, leaving the decision to human judgement and institutional policy.
Do I have to prove I did not use AI?
Not under the OIA's framework, which places the burden on the university. In practice, drafts, notes and version history are the strongest answer to a score, so keep them from the first day.
What if a reference was wrong but I did not use AI?
Say so and show the source you used. EdCitation's free Verify references tells a mistyped reference from a missing one: a real paper with the wrong year comes back "check this", with the reason, and an invented one comes back not found.
Can I see what a detector says before I submit?
Yes. EdCitation's Check AI, part of Pro, labels each passage of your own text with its score, in a report only you see. It is one detector's estimate, not proof either way.
References
- Bergin, J. (2025, October 9). University wrongly accuses students of using artificial intelligence to cheat. ABC News. https://www.abc.net.au/news/2025-10-09/artificial-intelligence-cheating-australian-catholic-university/105863524
- 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
- GPTZero. (n.d.). Frequently asked questions. https://gptzero.me/faq
- Munoz, A., Hinchcliff, M., Langfield, C., & Rogerson, A. (2026). How strong is the evidence in generative AI-related academic misconduct allegations? A mixed-methods analysis. International Journal for Educational Integrity, 22, Article 26. https://doi.org/10.1007/s40979-026-00235-9
- 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/
- Office of the Independent Adjudicator for Higher Education. (2025a). AI and academic misconduct: CS072501 [Case summary]. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072501/
- Office of the Independent Adjudicator for Higher Education. (2025b). AI and academic misconduct: CS072505 [Case summary]. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072505/
- Office of the Independent Adjudicator for Higher Education. (2025c). AI and academic misconduct: CS072506 [Case summary]. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072506/
- Office of the Independent Adjudicator for Higher Education. (2025d). Casework note: Complaints relating to AI and academic misconduct. https://www.oiahe.org.uk/resources-and-publications/learning-from-our-casework/ai-and-academic-misconduct/casework-note-complaints-relating-to-ai-and-academic-misconduct/
- Quality Assurance Agency for Higher Education. (2023). Reconsidering assessment for the ChatGPT era: QAA advice on developing sustainable assessment strategies [PDF]. https://www.qaa.ac.uk/docs/qaa/members/reconsidering-assessment-for-the-chat-gpt-era.pdf
- Tertiary Education Quality and Standards Agency. (2024). Guidance note: Academic and research integrity (Version 2.0) [PDF; archived copy]. https://web.archive.org/web/20251107073827/https://www.teqsa.gov.au/sites/default/files/2024-02/guidance-note-academic-and-research-integrity.pdf
- Turnitin. (n.d.). Using the AI Writing Report [Archived copy]. Turnitin Guides. https://web.archive.org/web/20260920025637/https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report
- University of Cape Town. (2025). UCT framework for artificial intelligence in education: Generative and other AI in teaching, learning and assessment [PDF]. https://cilt.uct.ac.za/sites/default/files/media/documents/cilt_uct_ac_za/2486/uct-ai-in-education-framework-june-2025-final.pdf
- University of Pittsburgh, University Center for Teaching and Learning. (n.d.). Encouraging academic integrity. https://teaching.pitt.edu/resources/encouraging-academic-integrity/