When the work is fluent, what tells us the pupil understands?

01 / Begin with the purpose

Three uses of AI require three different decisions

An exercise in which pupils critique an AI-generated explanation is different from coursework submitted as evidence of independent achievement. A teacher using AI to suggest feedback is a third activity. Each has its own purpose, permissions and risks. DfE's wider policy describes opportunities and limitations in education; it should not be read as overriding qualification-specific assessment rules.4

Ask the exams officer and subject leaders to map the assessments that contribute to qualifications, including the non-examination components. This follows the practical starting point in Ofqual's March leadership briefing. Record which guidance applies, who authenticates work and how pupils are informed. A whole-school statement is useful only if departments can explain its application to a real task.3

Include work completed outside supervised lessons and any arrangements for private candidates where relevant. Identify the stages at which staff can see the pupil's thinking. The outcome is a proportionate assessment map, not a new surveillance system or an assumption that unsupervised work is necessarily dishonest.

02 / Use the right version

Check the actual document before briefing staff

The source check matters here. JCQ's October-hosted v4 PDF states that it is effective from September 2026; an older HTML page continues to display an April 2025 edition. We checked the current PDF directly. Leaders should retain the edition used for their briefing and check the awarding body's instructions for the qualification concerned, rather than relying on a saved search result.2

One explicit current requirement is to give students the relevant AI information for candidates before they begin non-examination assessment. Check that the communication happened, that pupils understand it and that new starters are included. A document uploaded to a website does not establish that the candidate received and understood the instructions.2

Explain boundaries through the task itself: what is permitted, what must be acknowledged and what the pupil must demonstrate independently. Acknowledging AI does not make every use permissible or earn credit for understanding the pupil has not demonstrated. Avoid contradictory messages in assemblies, departmental guidance and homework instructions.2

03 / See the thinking

Build authentication into the work rather than adding it at the end

Use the normal production process to understand authorship: planning, discussion, intermediate work and appropriately supervised activity within the assessment rules. Staff need a credible picture of how the final submission developed. This is more useful than discovering at the deadline that the only evidence available is a polished document and a signed declaration.

For a proportionate local review, ask a department to trace one assessment from the initial instructions to authentication. Where did a teacher see the pupil make a choice, explain a source or revise an argument? What would happen if the final work differed sharply from that process? These are review questions, not additional assessment conditions to impose without checking the applicable rules.

Keep access arrangements and legitimate support visible. A change in presentation or fluency is not itself proof of AI misuse. Staff should understand the pupil's permitted support and usual work before drawing conclusions. Protecting integrity includes protecting pupils from an allegation based on an unsupported impression.

04 / Handle doubts fairly

Investigate concerns with evidence and the correct procedure

JCQ explains that detection tools can produce false positives and false negatives and should form part of a broader consideration of authenticity. A numerical score cannot settle the case. Do not present a detector output as a finding of malpractice or buy a product on the assumption that it removes the need for professional judgement.2

Ofqual's leadership briefing supports a conversation in which a pupil explains their thinking and process, followed by the appropriate malpractice procedure if concerns remain. Staff should know who advises them, what to retain and how the stage of authentication affects the reporting route. Use the current awarding-body process rather than improvising a sanction.3

Give staff time and support to handle concerns consistently. Record the evidence considered, the pupil's explanation and the decision through the approved process. Keep sensitive material within authorised access. A fair investigation protects both qualification integrity and the pupil's ability to understand and challenge what has been concluded.

05 / The assessor's responsibility

Human review must determine the judgement, not approve a suggestion

For assessment within JCQ's scope, its guidance permits centres to decide whether AI assistance in marking is appropriate after considering privacy, but prohibits AI as the sole marker. The human assessor must review the whole work and determine the mark. The responsibility remains with that assessor; agreement with a generated suggestion is not sufficient assurance by itself.2

Before any staff use, establish whether the tool and processing are approved. DfE's data guidance asks schools to consider how personal information is used, seek the relevant DPO or IT advice and check outputs. Removing a pupil's name may not remove all identifying or sensitive content from their work. Approval should address the actual information and purpose, not simply the product name.5

For ordinary formative feedback, judge usefulness separately from speed. EEF's guidance directs attention to whether pupils can use feedback to improve learning. A faster comment that is inaccurate, inaccessible or disconnected from the taught task does not achieve that purpose. Review examples and subsequent pupil action before claiming an educational benefit.6

06 / The leadership outcome

Test the controls through one assessment cycle

Bring the assessment, digital, inclusion and data leads together for a focused review. Check a small sample of instructions, candidate communications, authentication records and any approved staff use. Identify contradictions and give each correction an owner. Use existing quality assurance where possible, so that the review strengthens teaching rather than creating another parallel reporting burden.

At the next review, ask whether staff can explain the boundaries, pupils understand them and judgements remain defensible. Check workload as well as compliance: a tool may shift effort from drafting to checking rather than remove it. The practical outcome is confidence in what the assessed work demonstrates, with a process capable of examining doubt fairly.

Leadership questions

Questions that turn the development into a governing conversation.

What we are monitoring next

Publication is the beginning of the watch.

• JCQ and awarding-body revisions to AI, authentication, candidate information and malpractice procedures.
• Formal Ofqual decisions, distinguished from speeches and proposals about future assessment.
• The school's evidence on consistent instructions, fair handling of concerns and defensible human judgements.

Continue the evidence trail

How this develops our earlier coverage

• Our autumn AI feature: evidence, data boundaries and responsible use ↗
• AI adoption: review the whole lifecycle, not only the pilot ↗
• Data governance: make human review a working capability ↗