AI for academic assessment: what staff can hand over, and how to design for it

AI for academic assessment raises two separate questions, and most institutions only discuss one of them. The first is what staff can sensibly use AI for when they set, mark, and moderate work. The second is how to design assessment that still tells you what a student can do, now that HEPI’s 2026 survey found 94% of UK undergraduates use generative AI to help with assessed work¹.

This guide covers both, for lecturers, programme leaders, and assessment and quality leads. It assumes your institution already has an AI policy. If it doesn’t, start with our guide to writing a university AI policy.

What can lecturers use AI for in assessment?

Most of the useful jobs sit around the mark rather than in it, saving time without putting standards at risk.

Drafting rubrics and marking guides. Give AI your learning outcomes, the assessment brief, and your grade descriptors, and ask for a first draft of a rubric with a level descriptor for each criterion. Then edit hard: AI tends to write descriptors that sound different but mean the same thing, and markers need clear steps between bands.

Writing exemplars for calibration. Ask for three short responses to the same question, pitched at a borderline pass, a solid 2:1, and a first. Markers then grade them blind before a moderation meeting, which shows quickly where the team disagrees. Label every exemplar as AI-generated, and never present one to students as real student work.

Building feedback banks. AI can turn your rough notes into a bank of clear, constructive feedback statements tied to each rubric criterion, with a next step in each one. You still choose and adapt each one.

Preparing for moderation. AI can summarise markers’ comments on a sample, flag where rationales disagree, or draft the moderation report from your notes. Anonymise first, and keep the judgement about the sample with the moderator.

Generating questions. For quizzes, formative checks, and question banks, AI can produce varied questions at different levels of difficulty, plus plausible wrong answers for multiple choice. Check every answer key yourself: AI can be confidently wrong about which option is correct.

Accessibility adjustments. AI can rewrite an assessment brief in plainer English or numbered steps without changing what’s being assessed. Decisions about individual reasonable adjustments stay with your disability and inclusion processes.

What shouldn’t you hand over to AI?

Three things stay with people, whatever the tool.

The final judgement on a grade. A mark is an academic judgement that affects a student’s progression, and someone has to be able to explain and defend it at appeal. AI can help you draft the words around a mark. It shouldn’t decide the mark.

Unverified feedback. If AI drafted a comment, read it against the student’s actual work before it goes out.

Student work in unapproved tools. A student’s submission is their work, and it often contains personal data. Don’t paste it into a free personal chatbot or any tool your institution hasn’t approved for that purpose. Our guide to responsible AI in higher education covers the data rules in more detail.

Do AI detection tools work?

Not well enough to prove misconduct on their own. Jisc’s National Centre for AI has said since 2023 that detectors can’t conclusively show a piece of text was written by AI, and its 2025 update found no major improvement in detection technology². There’s also a fairness problem. A Stanford study found that widely used detectors flagged more than half of a sample of essays by non-native English writers as AI-generated³.

Treat a detector flag as a reason to look more closely, not as evidence. A conversation with the student about their work, their drafts, and their sources tells you far more. Use only detection tools your institution has approved, and follow its misconduct process exactly.

How do you design assessment that holds up when students use AI?

Detection is a weak defence, so the stronger response is in the design. QAA’s advice on reconsidering assessment for the AI era asks providers to look at assessment across the whole programme, not one module at a time⁴. Three approaches come up again and again in UK and international practice.

The two-lane approach: secure and open assessment

The University of Sydney’s two-lane model sorts assessment into two kinds⁵. Lane 1 assessments are secure and supervised, such as in-person exams, practicals, and vivas, and they confirm that a student has the core knowledge and skills the programme promises. Lane 2 assessments are open: students may use AI, are taught to use it well, and are marked on what they do with it.

The model accepts that an unsupervised essay can’t be secured against AI. The design work moves to programme level: which outcomes must be checked in a secure setting, and where does that happen?

The AI Assessment Scale

The AI Assessment Scale (AIAS), developed by Mike Perkins, Jasper Roe, Leon Furze, and Jason MacVaugh, gives staff a shared language for how much AI each task allows⁶. Its current version has five levels:

1. No AI: completed in a controlled setting without AI
2. AI Planning: AI for exploring topics and planning, with the student’s own ideas assessed
3. AI Collaboration: AI for drafting and feedback, with the student assessed on how they evaluate and use its output
4. Full AI: AI use expected, with the student assessed on how well they direct it
5. AI Exploration: creative use of AI to solve problems or produce new ideas in the discipline

The scale works best as a way to communicate expectations clearly in every assessment brief. It isn’t a security measure: a level-two rule on an unsupervised task still relies on students following it.

Process evidence and vivas

When the finished product can’t show who did the thinking, assess the process too. Ask for annotated drafts, a short reflective account of how the student used AI and what they changed, or a log of key decisions. Short vivas or recorded walkthroughs, even five minutes on a sample of submissions, let students show they understand what they submitted.

How to bring this into your institution

Assessment change usually stalls between the policy and the module handbook. A few steps help:

1. Map each programme’s learning outcomes to secure and open assessments, so every core outcome is checked somewhere supervised.
2. Put the permitted level of AI use, and an example of what it means, in every assessment brief.
3. Agree how staff themselves may use AI in marking and feedback, and write it down.
4. Train markers together, using real briefs and anonymised exemplars from your own programmes.

That last step is where our AI for Academic Assessment course fits. Pitched from foundation to intermediate, it’s built around five learning outcomes: designing assessments that stay authentic and robust in an AI-enabled environment; developing marking guides, rubrics, and feedback statements with AI responsibly; generating exemplar responses for calibration and moderation; identifying potential AI misuse and applying institutional guidance; and streamlining marking and feedback while upholding academic standards. Like all our courses, it can be delivered as a workshop or as a self-paced online course built around your own policy. For the wider rollout, see our guide to implementing AI at your university.

Find out where your staff are

The quickest way to see how your academic staff use AI in assessment now, and where to start, is our free AI Capability Check. It’s a 30-minute call over Teams or Zoom, with no obligation. We’ll place your team on the AI Use Hierarchy, give you two or three prioritised next steps, and send you a written report to keep.

Book your free AI Capability Check

You can also explore the courses, including AI for Academic Assessment.

Frequently asked questions

Can lecturers use AI to mark essays? AI can help draft rubrics, feedback statements, and moderation notes, but the final mark should be an academic’s judgement. Use only tools your institution has approved, and don’t paste student work into personal accounts.

Are AI detectors accurate enough to use in misconduct cases? Not on their own. Jisc has said detectors can’t prove text was AI-written, and studies have found they flag non-native English writers unfairly. Treat a flag as a prompt to look further, and follow your misconduct process.

What is the AI Assessment Scale? A five-level framework by Perkins, Roe, Furze, and MacVaugh, from No AI to AI Exploration, that helps staff state clearly how much AI use each assessment allows.

What is the two-lane approach to assessment? A model from the University of Sydney that splits assessment into secure, supervised tasks that confirm core skills, and open tasks where students may use AI and learn to use it well.

How should universities redesign assessment for AI? Work at programme level: make sure every core outcome is checked somewhere secure, state the permitted AI use in each brief, and assess process as well as product.