Exam boards asked teachers to authenticate work, not detect AI

The classroom note

Since March 2023, a UK teacher marking coursework or a non-exam assessment has had a joint document to work from: the JCQ guidance on AI use in assessments, produced by the Joint Council for Qualifications, the body representing England's major awarding bodies. An early version of the same page, archived in March 2023, already carried the title 'AI Use in Assessments: Protecting the Integrity of Qualifications' and described its purpose as giving 'teachers/assessors involved in delivering JCQ qualifications the information they need to manage use of AI in assessments'. For a head of department, the change was a shared vocabulary and a shared expectation across specifications, rather than each subject or centre inventing its own approach.

What the evidence says

The guidance, in the version dated 30 April 2025, is explicit about mechanism rather than intention: a student must retain 'copies of prompts and outputs in non-editable format' and reference the specific tool and generation date, and unacknowledged AI use 'constitutes malpractice with severe sanctions including disqualification'. It also states what a teacher must not do: rely on AI as a sole marking method, or accept work the teacher 'cannot confidently authenticate'. This is the guidance's own account of its requirements, not a study of how often centres comply; JCQ does not publish, on this page, a measured rate of undetected misuse or malpractice cases before and after the guidance.

The implementation question

The document asks teachers to authenticate work through comparison with a student's known writing pattern, review of intermediate drafts, and brief conversation to check understanding, treating AI-detection software as supplementary evidence rather than proof. That is a workload and a skill: a teacher must already know what a student's unaided writing looks like well enough to notice a change, across every student in a class, on top of marking. Detector output is explicitly not a substitute for that judgement, which shifts the practical cost from a licence fee for software onto a teacher's time spent watching drafts develop over a term.

What holds and what fails

What holds is the design choice to build malpractice policy around authentication of the whole process, not a single scan of the final document, which does not depend on any one detector's accuracy claims holding up. What is likely to fail, editorially, is an assumption that authentication scales without additional time: a teacher with large classes and infrequent contact with early drafts has fewer of the comparison points the guidance assumes are available. The policy is a rule about what evidence counts, not a guarantee that the evidence will exist in every classroom.

• Does this department actually collect intermediate drafts, or does the guidance's authentication method exist only on paper here?
• What happens when a teacher's honest answer is that they cannot confidently authenticate a piece of work?
• Has this year's version of the guidance changed since last year's, and who is checking?

A joint policy that asks teachers to know their students' writing, rather than trust a score from a detector, is asking for something that takes a term to build and seconds to lose. That is the real implementation cost behind a short, confident-sounding malpractice rule.