Exam boards split assessment into what AI can and cannot reach

The classroom note

A teacher setting a piece of coursework this term is designing it inside a landscape three regulators have already mapped. Reading their guidance side by side, as retrieved on 16 September 2026, shows they agree on one structural point before any policy detail: some forms of assessment are already protected from generative AI, and others are not.

What the evidence says

Ofqual's policy paper, updated 16 July 2026, states that "supervised exams remain largely protected," since students cannot access AI tools during exams, while "non-exam assessment...faces greater pressure," since unsupervised, extended coursework is where undisclosed use is hardest to see. The Joint Council for Qualifications' guidance, updated 30 April 2025, addresses that gap with a specific mechanism: students must reference AI-generated content as they would "any quote or material from another source," naming the tool and the date the content was generated. The International Baccalaureate's statement, first published 1 March 2023, makes a related point: it will not ban AI tools, but "does not regard" work produced even partly by such tools "to be their own," so an essay that is "predominantly quotes," AI-generated or otherwise, "will not get many, if any, marks."

The implementation question

The mechanism at work here is process evidence, not detection. Citing an AI tool by name and date, as JCQ specifies, converts an invisible act into a visible one a marker can weigh, much as a bibliography already does for any other source. That reframes the design question for whoever sets the task: an assessment fully answerable by an unedited AI response, with nothing distinctive left for the student to add or cite, is not testing what it once tested, whether or not misuse is ever detected. Supervised conditions solve this by removing the tool; a citation requirement solves it by making the tool's contribution visible when it cannot be removed.

What holds and what fails

What holds: the underlying distinction between controlled and uncontrolled assessment conditions, which predates generative AI and still explains most of the risk difference these regulators describe. What fails: assuming a citation rule alone will surface every undisclosed use - none of these documents claims that, and JCQ's own guidance still leans on teachers comparing work against a student's known style. This is an editorial reading of three separate documents, not one shared policy: each covers a different qualification system, and a school should confirm which actually governs its own assessments.

• Is this task supervised, or unsupervised coursework where AI use would need to be self-disclosed?
• Does our policy specify exactly how a student should cite an AI tool, by name and date, as JCQ recommends?
• Could this assessment be substantially completed by an unedited AI response, with no distinctive contribution required?

None of these documents bans a tool outright. Each instead asks a design question first: what, precisely, can this assessment still measure once the tool is available.