A lesson-planning trial saved minutes, not the authorship line

The classroom note

A science teacher building next week's lesson on cell division types a request into ChatGPT instead of opening a blank template. That scenario is measured in a Teacher Choices trial run by the Education Endowment Foundation with evaluator NFER: 259 science teachers across 68 secondary schools in England were split between a ChatGPT-guide group and a group asked not to use generative AI for lesson preparation. The same clock, minutes producing material, is what a feedback tool changes when the producer is a student, a boundary the trial does not test.

What the evidence says

The result is specific enough to check rather than take on faith: ChatGPT-group teachers averaged 56.2 minutes of weekly preparation against 81.5 minutes in the comparison group, a 31% reduction rated high-security evidence, with a blind expert panel finding no drop in quality. Teachers used the tool lightly, mostly for quiz questions, and the guide came from a corporate social-impact team, not the evaluator. The trial does not measure authorship of assessed work; it measures a teacher's labour on unassessed preparation, and stretching that finding to a graded essay is an editorial leap the source does not make.

The implementation question

For assessed work, the exam boards' own joint guidance draws the line the trial leaves open: a student must identify text reproduced from an AI response, and reproducing or paraphrasing it so a submission is no longer the student's own is malpractice, as is using AI so the work no longer reflects the student's own analysis. Declared use requires naming the tool and showing how it was used, sometimes by screenshot. A detector such as Turnitin's writing-detection tool sits inside that boundary rather than settling it: it states a false-positive rate under 1% but does not itself rule on misconduct, leaving that to a teacher, and notes that detectors score revised AI text as less detectable, so an edited draft can defeat the check meant to catch it.

What holds and what fails

What holds is that a fast, quality-neutral saving on preparation labour is a real, measured effect for unassessed material, and the same behaviour, a first pass then an edit, is what JCQ malpractice findings describe when undeclared on assessed work. What fails is treating ‘AI helped write it faster’ as one verdict: the trial shows fast and undamaged when the output is not scored as attainment; the guidance shows the same speed becomes a breach once it substitutes for a mark meant to certify someone's own analysis.

• Is the AI-assisted work marked as the student's own attainment, or used as unassessed preparation?
• If a detector flags a draft, does the school ask the student first, as the guidance recommends?
• Has anyone told the class, in writing, which parts of an assignment must be free of undeclared AI drafting?

A minutes-saved statistic and a malpractice definition answer different questions; a school that quotes one to settle the other has skipped the harder work of deciding, subject by subject, where drafting help ends and a mark meant to certify someone's own effort begins.