Assessment in the Age of AI: Design for Process, Explanation and Judgement

A polished final submission is no longer reliable evidence of learning on its own. A student may have drafted it independently, collaborated with peers, used generative AI thoughtfully, accepted AI output without scrutiny, or combined several approaches. The practical response is not to treat every student as suspicious. It is to redesign assessment so that learners show how they reached an outcome, explain their decisions and demonstrate what they can do with support removed or changed.

This is the central challenge of assessment in the age of AI: assess the learning you value, rather than only the product that is easiest to submit. Guidance from UNESCO calls for a human-centred approach to generative AI in education, while assessment-reform guidance from Australia’s Tertiary Education Quality and Standards Agency (TEQSA) argues that institutions should reconsider how assessment assures learning in an AI-enabled environment. UNESCO’s guidance and TEQSA’s assessment-reform resource are useful starting points for this shift.

Move from a single product to a body of evidence

A final essay, worksheet, design or code file can still matter. It may show synthesis, communication and technical quality. But when it is the only evidence collected, it carries too much weight. A more resilient design gathers a small body of evidence across the learning journey.

That body of evidence can include:

• a brief plan or proposal that identifies the student’s intended approach;
• annotated notes, source selection or worked examples;
• a draft with feedback and a revision memo;
• a transparent record of any AI use;
• a final product tailored to a real audience or context; and
• a short explanation, conference or oral defence.

None of these elements needs to be burdensome in isolation. Together, they give the teacher a better basis for professional judgement. They also make assessment more useful to students because the task rewards planning, critique and improvement—not merely producing a plausible answer at the deadline.

Start with the learning, not the tool

Before deciding whether AI is allowed, identify the knowledge and capability students are meant to demonstrate. Ask: What should a learner be able to understand, decide, create, explain or perform by the end of this task?

Then separate the assessment into components. Some may need independent performance: for example, applying a method, explaining a concept in the learner’s own words or making a disciplinary judgement. Other components may appropriately include AI: brainstorming alternatives, generating practice questions, critiquing an initial draft or comparing explanations.

A practical framework for communicating those boundaries is the AI Assessment Scale, a proposed model that distinguishes levels of AI involvement in assessment. You do not need to adopt its labels wholesale. Its useful principle is simpler: tell students clearly whether AI is prohibited, permitted for limited support, expected as a collaboration tool, or part of the task itself.

Build five kinds of evidence into the task

1. Process evidence

Process evidence shows how a learner approached the work. It need not mean monitoring every keystroke. Instead, request purposeful checkpoints: a question map, a project plan, a selected set of research notes, a worked calculation, an outline, a design rationale or a short screen recording of a practical workflow.

Make checkpoints assessable only where they serve the learning goal. A student should understand why they are being asked to show a plan: not to prove innocence, but to practise a method that experts use.

2. Explanation evidence

Ask students to explain important choices. Prompts such as “Why did you choose this evidence?”, “What alternative did you reject and why?”, or “How would the conclusion change if this assumption were false?” require judgement rather than surface-level production.

Explanation can be written, audio-recorded or delivered in a short discussion. It should focus on a few consequential decisions, not become a second full assignment.

3. Revision evidence

Revision turns feedback into visible learning. Require students to submit an early version, receive feedback, revise, and add a concise memo that explains what changed. A useful revision memo names one improvement made, one piece of feedback not adopted, and the reason for that decision.

This protects room for agency. Students are not rewarded simply for obeying every suggestion; they are rewarded for evaluating feedback and improving deliberately.

4. Oral defence or live application

A short oral defence, live demonstration or teacher conference can test whether students can apply and explain their work when the context changes. For example, after submitting a lesson plan, a trainee teacher might justify one adaptation for a learner with different needs. After submitting a business proposal, a student might respond to a new constraint introduced by the assessor.

Keep this proportionate. A two-to-five-minute check-in may be enough for a low-stakes task. Use a shared question bank, a simple rubric and predictable scheduling so that the format is manageable and fair. Offer equivalent accessible ways to demonstrate explanation where appropriate, in line with local support arrangements.

5. Transparent AI-use evidence

If AI is permitted, ask students to disclose its role. The goal is not a performative confession. It is to build the evaluative habit of recognising where tools influenced the work and what the learner did with the output.

A short declaration can ask students to state:

• which tool or tools they used, if any;
• which stage of the task involved AI;
• one output, suggestion or claim they accepted, changed or rejected;
• how they checked accuracy, bias, relevance or quality; and
• which final decisions remained their own.

Do not grade students on whether AI made their prose sound more polished. Grade the relevant learning outcomes: the quality of their analysis, verification, decision-making and communication.

A practical assessment pattern

For many extended tasks, a four-stage pattern is enough:

Consider a history inquiry in which students create a public-facing exhibit panel. They submit a proposal identifying their question and sources; an annotated draft explaining how they selected evidence; a final panel; a revision memo; and a two-minute response to a follow-up question about a source they did not use. AI may be allowed for idea generation or language feedback, but students must identify the use and verify factual claims against approved sources. The final grade can prioritise historical reasoning and evidence use rather than treating fluent wording as the whole achievement.

Write instructions students can act on

Ambiguity invites uneven practice. Put AI expectations beside the task instructions, assessment criteria and examples—not in a policy document students may never revisit. Use direct language.

In this assessment, you may use generative AI to brainstorm questions and receive feedback on clarity. You may not submit AI-generated analysis as your own. If you use AI, include a 150-word use-and-evaluation note describing the tool, purpose, key output and how you checked or changed it. You must be able to explain your final choices in a short follow-up discussion.

Adapt the wording to the subject and age group. Most importantly, make the permitted use realistic. If students are expected to use AI, teach them how to question outputs, check sources, protect personal information and recognise that confident wording is not the same as reliable content. UNESCO’s guidance highlights ethical, safe and meaningful use as core considerations for education. Read the full UNESCO publication.

Use a rubric that rewards judgement

A rubric should make the design visible. Alongside criteria for the final product, consider criteria such as:

• Reasoning: the student explains choices with relevant evidence or principles.
• Process: planning and development materials show a purposeful approach.
• Revision: changes respond thoughtfully to feedback or self-evaluation.
• Evaluation of AI use: where AI is allowed, the student identifies limitations, checks outputs and retains responsibility for decisions.
• Communication: the student can explain the work appropriately for the intended audience.

Weight criteria according to the actual learning outcomes. Do not add process points merely because AI exists. In a timed calculation task, independent application may be central. In a professional writing task, responsible tool use and editorial judgement may deserve more prominence.

Avoid the two unhelpful extremes

The first extreme is designing every task as if AI does not exist. This can leave teachers trying to infer authorship from a final product alone. The second is treating every task as an AI-detection exercise. That shifts attention away from teaching and can damage trust.

A better approach is assessment by design: vary the evidence, state the conditions, teach responsible use and retain teacher judgement. Students should know what is being assessed, what support is allowed and how they can demonstrate their own learning.

Make the workflow sustainable

Assessment redesign should not create an impossible marking load. Reuse a small set of checkpoint templates, revision-memo prompts and oral-defence questions across a course. Use short, targeted feedback at draft stages and reserve detailed feedback for the most consequential work. Sample process evidence where a full review is unnecessary, while ensuring your method remains consistent and transparent.

SubSchool can help educators automate repetitive parts of assessment and feedback workflows while teachers keep authorship and the final educational decision. If you are building clearer rubrics, feedback routines and review points around AI-enabled work, explore SubSchool’s AI grading features and decide where automation supports—rather than replaces—your professional judgement.

Bottom line: In the age of generative AI, the strongest assessment asks students to make their thinking visible. Design for process, explanation, revision, oral defence and transparent tool use. The final product still matters, but it becomes one piece of a more credible and more educational picture.