Academic integrity policy wording for AI and references

Most academic integrity policies say that references must be accurate and that AI use must be acknowledged. Neither sentence can be enforced. "Accurate" is a judgement, and "acknowledged" does not say what the acknowledgement contains or where it goes.

Wording that works is wording that can be checked: a reference either resolves to its source or it does not, and a tool was either declared at the level the brief allowed or it was not. Below: what the sector bodies ask for, how universities in the UK, the US and Australia word it now, our own template clauses for a syllabus, an assignment brief and a rubric, and a fair process for the day a reference cannot be found. That day comes: in the Walters and Wilder (2023) test, more than half the references GPT-3.5 produced (55%) and nearly a fifth of GPT-4's (18%) did not exist. A clause promising that references will be verified is only as good as the check behind it, and that check is now cheap: EdCitation's free Verify references looks up every entry in a pasted list against the publisher's record.

What should an academic integrity policy say about AI and references?

Three things, in the assessment brief where students read it: what makes a reference checkable, which AI uses are permitted at what level and how they are declared, and what happens when a reference cannot be found.

None of these documents says how a reference is made checkable; the first template clause below fills that gap.

How do universities word it now?

In three different ways, and only some mention references. Every page was read on 22 September 2026.

United Kingdom

The University of Edinburgh's guidelines, revised March 2026, list "Citing and referencing AI-found sources without reading and verifying them" as misconduct, and give a one-sentence acknowledgement naming the model, the platform and the task. The University of Glasgow permits generative AI in unsupervised assessments by default, with an acknowledgement giving the tool, its version and publisher, how it was used and which parts are the student's own. Its page warns that a model "generates text that looks like a citation" and tells students to find and read every source.

United States

Stanford's guidance, adopted 16 February 2023, sets the opposite default: where a course says nothing, AI use is treated as help from another person, students acknowledge any use "other than incidental use", and they disclose when in doubt.

Australia

The University of Sydney's page, updated 2 June 2026, requires an acknowledgement of "any use of software or tools", tells students to keep AI outputs "as evidence of your research and writing process", and says a detector score "would not be the only evidence relied upon". Monash University gives a fill-in format: the tool, the task, the number of iterations and how the output was changed. For brainstorming, it says, "a citation would not be appropriate".

Where they disagree, and what we could not read

Glasgow permits by default; Stanford treats undeclared use as unauthorised help; Edinburgh restricts use for "much of your assessed work" and leaves detail to each course. A clause must therefore state its own default for a silent brief. The pages of Oxford, UCL, Cambridge, Harvard, Michigan, Melbourne and UNSW could not be read on the day of writing, so nothing is claimed about them.

Which AI Assessment Scale level should the brief name?

One of the five: a number with a fixed meaning is easier to teach, apply and defend than a paragraph. Perkins et al. (2024) define them in Table 1 as follows.

Three details matter for wording. The levels are cumulative, so a Level 3 task also permits everything at Level 2. The authors recommend Level 1 only under supervision or for low-stakes work. And the paper's own Level 2 example includes an internet-connected model suggesting sources, which is where invented references enter a paper. The scale says nothing about references, so a referencing clause must apply at every level, including Level 5.

Perkins et al. (2025) have since revised the scale, removing the traffic-light colours and adding a level called AI exploration, so a policy should say which version it names.

What wording makes references checkable?

A reference is checkable when its identifier resolves to the source and the student can produce the source on request. The clause below is our own template, not any university's; adapt it, and say that it applies at every AI level.

Template clause: references (EdCitation, for adaptation)

1. Every reference gives a DOI where one exists, and otherwise a URL to the publisher's or organisation's own copy. A print book gives an ISBN or a library holding.
2. By submitting, you confirm that you have read each source you cite and that it says what you cite it for.
3. On request, within five working days, you supply a source log: for each reference, where you found it, when, and a copy, link or library record.
4. A source suggested by an AI or search tool is found, read and cited from the publisher's record, not from the tool's output.
5. A reference that cannot be found will be put to you before any conclusion is drawn.

Clauses 1 and 4 are quick for a student to meet. Given a DOI, a title, a PubMed ID or an ISBN, EdCitation's Cite a source builds the reference from the publisher's record in the style the brief sets, so the identifier comes from the publisher and not from a chatbot's memory.

The exceptions to write in

A DOI is not proof on its own. Linardon et al. (2025) checked 176 references written by GPT-4o: 35 did not exist, and of the 141 real ones, 64 carried errors, most often an incorrect or invalid DOI. So the clause requires a DOI that resolves to the source cited, and the marker opens it and compares title and authors; what is a DOI explains this for students. Reports, theses, older chapters and sources in other languages are covered poorly by the indexes, so "could not be checked" stays separate from "not found". A wrong page number on a real paper is a referencing error, not fabrication.

How should the AI-use clause be worded?

It names the level, states the default for a silent brief, and fixes the form of the declaration. Again, this is our template.

Template clause: AI use (EdCitation, for adaptation)

1. Each brief names its level on the AI Assessment Scale (Perkins et al., 2024, Table 1). Where a brief names none, Level 2 applies: AI for ideas and structure, no AI-generated content in the submission.
2. Any use beyond incidental spelling and grammar checking is declared in a statement at the end of the work: the tool and version, the task, how much of the submission it affected, and how the output was changed. Model statement: "I used [tool, version] to [task]. It affected [which parts]. I changed the output by [what]."
3. At Level 4 or above, AI-generated content in the submission is also cited; see how to cite ChatGPT and AI tools.
4. Keep prompts, outputs and drafts until the mark is confirmed. You may be asked for them.
5. Sources found with a tool are subject to the references clause in full.

Rubric lines for referencing

QAA (2023a) suggests rubrics that reward higher-level skills. Referencing is not one, but a rubric line is where a student learns that checkable sources carry marks.

What is a fair process when a reference cannot be found?

Evidence first, then a question to the student, and no conclusion until the student has answered. The steps follow the Office of the Independent Adjudicator's Good Practice Framework for England and Wales, a fair yardstick anywhere.

1. Record the check. The reference, where you searched, what you found, the date, and whether the result is "no record where one should be" or "could not be checked". Where a checker such as EdCitation's ran first, note its verdict and put your own search beside it: a tool's result opens the question and never settles it.
2. Ask for the source, in writing, with the deadline the clause set. A student who read the paper can usually send a PDF, a link or a library record.
3. Meet, with the evidence in advance. The OIA expects the student to receive everything the decision maker will consider, with reasonable notice and a right to a representative.
4. Keep the burden where it belongs. The framework puts the burden of proof on the provider, on the balance of probabilities unless the regulations say otherwise; the student "should not have to disprove the allegation".
5. Give reasons, for the finding and for any penalty. QAA (2023b) asks that misconduct policies be applied "sensitively and sparingly", and that a first case may be best dealt with through student support.
6. Allow an appeal, heard within 30 days of its being lodged.

Why a detector score is never the proof

A missing reference is a fact about a source. A detector score is an estimate about prose, and says nothing about whether a reference was invented, or by whom. Weber-Wulff et al. (2023) put 14 detectors to the test, 12 public and two commercial, and judged them "neither accurate nor reliable", tending above all to call text human-written. Sydney says its detector score is never the only evidence; a policy should say the same in one sentence. A student may put the published error rates to you; they are collected in the guide on detector false positives.

How do you check every reference against the clause?

With a lookup, not a judgement. EdCitation's Verify references is the best tool for enforcing a checkability clause, because it tests what the clause asks: not whether a reference carries a DOI, but whether the record behind it is the paper cited. Linardon et al. (2025) show why that is the right test, since the commonest error in the real references GPT-4o wrote was a wrong or invalid DOI. A chatbot cannot do this job; it writes references and does not look them up. Paste a list or upload the paper: every entry is reported as verified, as doubtful (shown as "check this") or as not found, and retractions are flagged. "Could not check" is never shown as "not found", which keeps the policy's own distinction intact. It is free and needs no account; Pro, at $8 a month, and Max, at $24, add paid tools for the writer and are set out on the pricing page.

A tool covers some of the clause and not the rest, and the policy should say which is which.

The result is evidence about the references; how they got there is the student's to answer. For a whole institution, the Institution licence covers every student under one invoice, connected to the LMS and the university's sign-in. The marker's own spot check is in how to spot fabricated references in student work.

Quick questions

Should a policy ban AI outright?

Usually not. QAA's May 2023 advice prefers clear rules on permitted use to a ban, and the AI Assessment Scale lets a brief name a level from No AI to Full AI for each task.

What must an AI declaration contain?

The tool and version, what it was used for, how much of the work it affected, and how the output was changed. Glasgow, Edinburgh, Sydney and Monash all require an acknowledgement, and Glasgow and Monash list those elements.

Is a fabricated reference the same as unauthorised AI use?

No. A reference that does not exist is a fact about the source; how it got into the paper needs the student's answer. Record the two separately, and never infer the second from a detector score.

Does every reference need a DOI?

Every reference that has one should give it, because a DOI resolves to the publisher's record and a title search may not. A source without one gives a URL to the organisation's own copy, an ISBN or a library holding, so a marker has somewhere to look.

Should the brief tell students to check their own references first?

Yes, and it costs them nothing to do. Pasting the list into EdCitation's free Verify references before submission shows a borrowed DOI or a missing paper while it can still be fixed. The marker's own check still follows; the student's run is preparation, not a substitute for it.

References

• International Center for Academic Integrity. (2021). The fundamental values of academic integrity (3rd ed.). https://academicintegrity.org/aws/ICAI/asset_manager/get_file/911282?ver=1
• Linardon, J., Jarman, H. K., McClure, Z., Anderson, C., Liu, C., & Messer, M. (2025). Influence of topic familiarity and prompt specificity on citation fabrication in mental health research using large language models. JMIR Mental Health, 12, Article e80371. https://doi.org/10.2196/80371
• Monash University. (n.d.). Acknowledging the use of AI. https://www.monash.edu/student-academic-success/build-digital-capabilities/create-online/acknowledging-the-use-of-generative-artificial-intelligence
• 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/
• Perkins, M., Furze, L., Roe, J., & MacVaugh, J. (2024). The Artificial Intelligence Assessment Scale (AIAS): A framework for ethical integration of generative AI in educational assessment. Journal of University Teaching and Learning Practice, 21(6). https://doi.org/10.53761/q3azde36
• Perkins, M., Roe, J., & Furze, L. (2025). Reimagining the Artificial Intelligence Assessment Scale: A refined framework for educational assessment. Journal of University Teaching and Learning Practice, 22(7). https://doi.org/10.53761/rrm4y757
• QAA. (2023a). 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
• QAA. (2023b). 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
• Russell Group. (2023). Russell Group principles on the use of generative AI tools in education. https://www.russellgroup.ac.uk/sites/default/files/2025-01/Russell%20Group%20principles%20on%20generative%20AI%20in%20education.pdf
• Stanford University. (2023, February 16). BCA guidance & recommendations: Generative AI policy guidance. Office of Community Standards. https://communitystandards.stanford.edu/generative-ai-policy-guidance
• University of Edinburgh. (2026). Using generative AI in your studies: Guidelines for students. https://information-services.ed.ac.uk/computing/comms-and-collab/elm/generative-ai-guidance-for-students/using-generative
• University of Glasgow. (n.d.). Artificial intelligence: Info for students. https://www.gla.ac.uk/myglasgow/sld/ai/students/
• University of Sydney. (2026). Artificial intelligence. https://www.sydney.edu.au/students/academic-integrity/artificial-intelligence.html
• Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5
• 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