What the Dubai Accord Actually Says About AI Detection

The short answer. The Dubai Accord, launched 27 August 2026 at the ENAI Summer School in Dubai, is a voluntary international framework for academic integrity in the age of AI. Its own text is careful about what it is not: "not a universal misconduct code, a technical standard, a product certification scheme or a substitute for law", but "a shared ethical floor". On detection it is blunt: "No automated detection output or probabilistic indicator should be treated as conclusive evidence of misconduct." Several of Britain's strictest universities were already there: UCL does not use detectors when marking, Manchester says detector output "cannot currently be used as evidence of malpractice", and King's College London switched Turnitin's AI score off. Endorsing the Accord is not certification, and nothing in it approves any tool, ours included.

If you run a department, you probably assume the hardest line on AI detectors is drawn by the people who refuse to use them, and the softest by the people who sell them. Last week a document appeared that scrambles that assumption. The Dubai Accord on Academic Integrity in the Age of Artificial Intelligence was written by academic-integrity researchers, deliberated at the ENAI Summer School, and launched on 27 August 2026, and it lists Turnitin, the biggest detection vendor in education, among its industry partners. I read the full 42-page inaugural edition, and it contains the clearest statement on detection I have seen in any framework, from anyone.

A floor, not a rulebook

First, what the document is. The Accord describes itself as "a concise but actionable international framework" and then spends a paragraph saying what it is not: "not a universal misconduct code, a technical standard, a product certification scheme or a substitute for law". Institutions endorse it voluntarily, and the text is explicit that "endorsement does not certify compliance". What endorsement does carry is a pledge to review your own policies against the Accord's principles within twelve months, using a self-review checklist it publishes as an appendix.

That structure matters for how you read everything below. Nothing in the Accord binds a school, approves a product or certifies anyone. It is a set of commitments an institution measures itself against. So when I quote it, I am quoting a benchmark, not a regulation.

The detection line

The Accord's section on surveillance and detection contains this sentence: "No automated detection output or probabilistic indicator should be treated as conclusive evidence of misconduct. Such outputs may, where justified, form one element of an inquiry, but findings should rest on a broader evidentiary record and independent human judgement."

The student-rights section says the same thing from the other side: "Automated detection scores or probabilistic indicators should not by themselves establish misconduct". And the institutional commitments on due process add the machinery: students should receive "notice of concerns, access to the substance of the evidence, a meaningful opportunity to respond and a route to independent review or appeal", with records corrected where allegations are not substantiated.

And the procurement section turns it on vendors: institutions should "require vendors to provide relevant evidence about performance, interoperability, limitations, data practices, accessibility, security and foreseeable bias". The ask is evidence with the limitations left in, not marketing claims.

Britain's strictest universities got there first

While the Accord was being drafted, several UK universities had already written versions of its detection position into their own guidance. These are their words, from their own pages, at three different levels of formality:

• UCL's Generative AI Hub, university-wide teaching guidance: "UCL does not use GenAI detectors when marking. Instead, if a marker has any concerns about your work, they'll usually request a quick chat to get more information."

• Manchester's University AI Guidelines, owned by its AI Strategy Group and last edited in April 2026: "Tools to detect AI-generated content are unreliable and biased and cannot be relied on to identify academic malpractice in summative assessment. Output from such tools cannot currently be used as evidence of malpractice."

• King's College London's AI guidance, which KCL labels guidance rather than policy: "King's took the decision not to enable the AI detection % in Turnitin due to concerns about its reliability and potential for false positives."

Not every university goes that far. UCD's Academic Integrity Policy, a university-wide regulation, permits detection software but says it "should be used with caution and regarded only as one tool in assisting an examiner to make a judgement about academic misconduct". Cambridge's policy office page concentrates on the student side: unacknowledged AI content in a summative assessment is misconduct unless the brief says otherwise. The range runs from "one tool, with caution" to "not evidence at all". What I could not find, anywhere in this set, is a university saying a detector score alone can carry a finding.

Schools already have the same instruction, most just have not read it

For UK schools the equivalent line has existed since 2023. JCQ's guidance on AI use in assessments, which we went through line by line in what JCQ actually says about AI detection in coursework, treats a detector result as one piece of evidence inside a holistic judgement of whether work is authentic, made by the teacher, never by the tool. The Accord's contribution is not a new rule for schools. It is that the same position now has an international framework, a student-rights framing and a self-review checklist behind it, with the biggest vendor in the sector listed as a partner on the back page.

The Accord also offers something a school can steal directly: a six-question test it says institutions should apply "before introducing or materially expanding an AI-enabled practice". Educational purpose, human agency, evidence and quality, rights and equity, transparency and challenge, review and exit. Question five asks whether affected people will "be able to question or appeal consequential outcomes". If your school runs a detector and a student wrongly flagged by it would not know how to challenge the result, that question is currently failed. It is why we publish a guide for exactly that situation, wrongly flagged: what to do.

Where that leaves a detection vendor

We sell an AI detector, so the fair question is what this document costs us. Held against the Accord's procurement test, here is our own account, with the unflattering parts left in. Our measured error rates are published in full on the methodology page: 4 wrongly flagged essays in 400 real student essays, 5 in 517 professional pieces, and a list of what we miss that includes every one of 70 half-AI, half-human submissions. Every result we return says it is a screening signal for a teacher to weigh, never proof. Submitted text is never stored. And our position on penalties matches the Accord's: no decision about a student should rest on a score alone, ours included.

I would not claim the Accord for our side, and the document itself forbids that reading: it certifies nobody. What I can say is narrower. A framework written by integrity researchers, partnered by the largest vendor in the market, has now put in writing that detection output can never be more than one element of a human inquiry. The vendors this is uncomfortable for are the ones whose product pages promise certainty. The measured position was already the honest one, and as of 27 August it has 42 pages of international backing.