Keeping Judgement Human: A Practical AI Checklist for UK Assessors

Artificial intelligence is no longer a distant concern for vocational assessment. Assessors are encountering tools that draft feedback, summarise learner evidence, generate assessment questions, check for text similarity and analyse video or audio recordings. Some of these tools can reduce administrative burden. None of them removes the assessor’s accountability for the assessment decision.

The practical question for UK assessors is not whether AI will appear in assessment practice. It is how to use it without weakening authenticity, evidence validity, data protection or learner confidence. The checklist below is designed as a pre-use and review framework. It does not replace awarding organisation requirements, legal advice or internal quality assurance procedures, but it provides a structured way to test whether a tool belongs in the assessment process at all.

Start with the decision, not the tool

Before opening a new platform, define the assessment task and the decision that must be made. Is the tool supporting administration, such as scheduling or formatting? Is it helping to organise evidence? Or is it being asked to make or recommend an assessment judgement? The higher the tool’s influence on the outcome, the stronger the justification and oversight required. An assessor may use software to sort evidence, but the professional judgement must remain traceable to a competent person.

1. Confirm awarding organisation and regulatory requirements

Check the current guidance from your awarding organisation and, where relevant, the regulatory framework for the qualification. Some qualifications prohibit certain tools during assessment conditions. Others require specific declarations from learners. Apprenticeship end-point assessment organisations may have separate rules. Record the requirement you are working to and the date you checked it. If the guidance is silent, do not assume permission. Seek clarification and keep the response on file.

2. Define the boundary between support and judgement

Write a short statement describing what the tool will and will not do. For example: it may summarise a learner’s written evidence, but it will not determine sufficiency; it may suggest feedback wording, but the assessor will review and rewrite it; it may flag similarity, but it will not decide malpractice. This boundary should be understood by assessors, internal quality assurers and learners where appropriate. It also makes audit review easier because the role of the tool is explicit.

3. Protect authenticity and learner authorship

Authenticity is central to vocational assessment. If learners are permitted to use AI, the assessment conditions should say so clearly. Consider whether a declaration is required, whether the assessment environment can be controlled and whether the learner can explain the evidence they have produced. A tool that generates polished text can obscure authorship. Assessors should look for corroborating evidence, such as professional discussion, observation, practical demonstration or earlier drafts. Do not accept AI-generated content as proof of competence without testing the learner’s own understanding.

4. Test evidence validity and sufficiency

An AI summary is not primary evidence. It is a representation of evidence and may omit context, nuance or contradiction. Assessors should review the original source wherever possible, especially where the decision concerns safety, professional conduct or competence in a regulated setting. Ask whether the evidence is valid for the qualification criteria, sufficient in breadth and depth, and current. If the tool has translated, transcribed or summarised material, check for distortion. Keep the original evidence or a clear route to it.

5. Keep moderation and standardisation robust

AI-assisted assessment should be subject to the same moderation and standardisation as any other assessment. Sample decisions where tools have been used. Compare outcomes with decisions made without AI support. Ask whether the tool introduces bias through the data it was trained on, the language it privileges or the way it scores similarity. Internal quality assurers should be able to see where the tool influenced the process and what the assessor did next.

6. Treat malpractice and detectability carefully

AI detection tools are not infallible. A high similarity score is not proof of malpractice, and a low score is not proof of authenticity. Assessors should follow the awarding organisation’s malpractice procedure, gather evidence, invite the learner’s explanation and avoid making accusations based on a single automated output. Record the concern, the evidence reviewed and the decision reached. The same caution applies to deepfake audio or video, which may require specialist verification.

7. Apply UK data protection principles

Assessment evidence often contains personal data and may include special category data. Before using an AI tool, establish the lawful basis for processing, minimise the data entered and check where the data will be stored or transferred. Do not paste learner scripts, video, audio or personal details into public AI tools. Use approved systems with appropriate contracts, access controls and retention arrangements. A data protection impact assessment may be required for higher-risk processing. If in doubt, consult your organisation’s data protection lead and take advice from the Information Commissioner’s Office guidance.

8. Be transparent with learners

Learners should understand how AI may be used in their assessment and what is expected of them. Transparency supports confidence and fairness. Explain whether tools are used for administration, feedback support or evidence review; whether learners must declare AI use; and how they can raise a concern or appeal. Check that the use of a tool does not disadvantage learners with particular needs or create an additional barrier to reasonable adjustments.

9. Maintain an audit trail

Record enough information to reconstruct the process. This may include the tool used, the purpose, the date, the prompt or input category, the output reviewed, the assessor’s decision and any changes made. Do not retain unnecessary personal data. The audit trail should show that a competent assessor exercised judgement, not that a tool produced an outcome. Where feedback has been drafted with AI assistance, keep the final version and a note that it was reviewed by the assessor.

10. Review, record and refresh your competence

Technology changes quickly; assessment standards do not. Build AI-related questions into your CPD cycle and standardisation meetings. Review incidents and near misses. Ask whether the tool is still necessary, whether it has changed terms or functionality, and whether awarding organisation guidance has moved. Assessors do not need to become technologists, but they do need enough understanding to challenge a tool’s output and recognise its limits.

A short pre-use checklist

Before adopting an AI-supported assessment process, confirm that awarding organisation requirements have been checked and recorded; the tool’s role is clearly limited; learner authenticity can be verified; original evidence remains accessible; moderation and standardisation cover AI-assisted decisions; malpractice procedures are followed rather than automated scores; UK data protection principles have been applied; learners have been told what is happening; an audit trail can be produced; and staff competence has been reviewed.

AI can help assessors manage workload, but it cannot carry professional accountability. The safest approach is neither blanket refusal nor uncritical adoption. It is a documented, proportionate use of technology that keeps evidence valid, learners informed and judgement human.