England case v0.1 · Prepared
Coursework: who actually decides the mark?
One human choice · England · Evidence checked 12 November 2026
Imagine a teacher assessing a student's written GCSE project. An AI tool offers a mark and an explanation. The teacher could use that suggestion to guide the assessment, or make a judgment first and ask the tool to check it. Both arrangements include a person. They give the person different work to do and create different opportunities to notice a mistake.
For students, parents and teachers, the useful question is what the person must actually do—and whether there is enough time and support to do it well.
1. What exact work are we choosing about?
Here the task is awarding a mark for internally assessed non-examination work. JCQ's instructions, effective from 1 September 2026, cover relevant GCSE, AS and A level specifications in England and Wales; this case uses England as its example. They rule out AI as the main or only means of marking. Teachers must examine the complete work, apply their own professional judgment and retain responsibility for the result. JCQ instructions, section 6.1.
Ofqual's policy, updated in July 2026, tells awarding organisations that AI must not be the only marker of assessments within regulated qualifications. It also recognises potential uses in classroom assessment. These are different settings: a practice exercise and work determining a qualification can justify different arrangements. Ofqual's published approach.
This choice concerns a particular assessment function. It leaves wider questions about teaching, lesson preparation and employment terms to be considered separately.
2. Why keep a person involved?
A student needs a judgment that can be explained against the assessment criteria and reconsidered when there is a problem. In Human by choice's framework, that is a reason of answerability: someone has to take responsibility for a consequential decision.
Ofqual's January 2026 discussion also identifies possible uses of AI in checking human marking and training markers. It calls for evidence suited to the particular assessment context. It does not present a universal verdict that either people or machines mark best. Ofqual's explanation.
The strongest reason to retain judgment with a teacher is therefore more specific than a preference for familiar methods. The person should be able to recognise what the student has achieved, explain the assessment and correct an error. Whether a particular arrangement delivers that has to be tested.
3. What must the human role contain?
The JCQ approach gives the teacher both substantive assessment work and responsibility for the mark. A signature on an AI-produced result would be too thin a description of that role.
One possible arrangement to test would let the teacher record a provisional judgment before seeing an AI check. The teacher would then consider any disagreement and record the reason for the final mark. This is a suggested design for evaluation, not a requirement stated in the instructions or a tested method offered by this programme.
That arrangement has a real cost: it still requires substantial human assessment. Its possible benefit is that the school could examine what the AI check added, where it helped and where it distracted from a sound judgment. Different designs should be compared on the work they actually require.
4. Who has the authority to choose?
Ofqual regulates qualifications in England; its published approach applies its existing regulatory framework to AI use. The JCQ document supplies instructions for the covered assessments. Those are the starting conditions for a school's workflow. Ofqual policy; JCQ instructions.
A useful next step would be for centre leaders, teachers and assessment staff to describe a proposed workflow in ordinary language, with students able to question how their work would be judged. Awarding organisations could provide examples showing how permitted assistance fits their assessment requirements. A proposal to change the underlying division of responsibility would need to address the regulator's requirements.
This article proposes questions for that discussion. Human by choice has no mandate to speak for students, teachers or examining bodies.
5. Who pays, who gains, and what is the best opposing case?
The serious case for greater automation is that assessment takes time. If an assisted process could give students useful feedback sooner, improve consistency and free teachers for other work, keeping every existing task unchanged would have a cost too. These are potential benefits to demonstrate, not savings established by this case.
A workable budget would count reading, checking, resolving disagreements, training, subscriptions and administering reviews. It would also ask where any saved time goes. Does the teacher gain time for teaching, does the school reduce paid hours, or does the tool add another layer of checking?
Schools with fewer resources could face a particularly difficult choice if preserving the human role required extra capacity. Funding that role should be part of the decision. Passing responsibility to a teacher without providing time to exercise it would leave the central problem unresolved.
No cost estimate or measured effect on jobs is established here. The comparison worth making is between complete arrangements for delivering reliable assessment, including the people and resources each needs.
6. How would we know it worked, and when should we reconsider?
A bounded evaluation could compare lawful assessment workflows on final marking quality, differences between student groups, teacher time, correction effort and students' understanding of their marks. An independent assessment of quality would be needed: agreement with the AI's suggestion would not itself show that the mark was right.
Review is already part of the system. JCQ requires centres to tell candidates their internally assessed marks in time to seek a review before submission; moderation can subsequently change marks. JCQ instructions, section 6.1.
A trial should set its criteria and reconsideration date in advance, including what would count against adopting the tool. Evidence that a specific assisted design improved assessment without an excessive burden would strengthen its case. Evidence that checking consumed the promised savings, or that particular students were disadvantaged, would weaken it.
The human role should be described clearly enough to examine: what judgment the teacher makes, what assistance the system supplies, who can change the result and who has the resources to put an error right.