What Counts as Evidence of the Student?
Student assessment in a GenAI-infused environment needs more than a better detector or a larger pile of activity data. The specific problem here is how to connect a claim about what a student can do to observations that actually support that claim. Evidence-Centred Design provides a useful way to make that connection explicit, so educators can choose evidence that reveals human understanding without pretending that every trace of activity is a transparent record of thought.
Start with the claim, not the format
Assessment design often begins with a familiar container: an essay, a quiz, a project, an oral presentation, or a final code submission. The container matters, but it does not tell us what conclusion the evidence can support. The first question should be a sentence about the learner. For example: “The student can diagnose a clinical case by asking relevant questions and justify a provisional conclusion.” Or: “The student can select a numerical method, implement it, explain its assumptions, and adapt it when the conditions change.” Those claims are more precise than “the student completed the assignment.”
Evidence-Centred Design, or ECD, is a way of keeping the claim, the task, the observations, and the inference connected. An assessor first identifies the capability to be certified. They then design a task that gives the student an opportunity to demonstrate it, decide what observations would count as evidence, and determine what those observations justify—and do not justify—concluding. The value of the framework is not a special formula. It is the discipline of asking whether the task can elicit the intended capability and whether the evidence can bear the weight of the judgment.
A broad claim usually needs to be decomposed before evidence can be selected. “Can write well” might include understanding a question, forming an argument, using sources, organising ideas, revising in response to criticism, and communicating clearly. A system could help with grammar while leaving the argument to the student, or it could generate the argument and leave the student to correct punctuation. Those two situations produce different evidence. If the claim is not separated into its meaningful parts, the final score can conceal which capability was demonstrated and which was supplied by the tool.
The task conditions belong in the claim as well. An assessment that permits documentation and AI can legitimately ask whether a student can verify, adapt, and defend an assisted result. An assessment that intends to establish unaided recall must provide a condition in which the system cannot supply the answer. Neither condition is automatically more rigorous. They support different inferences. Making the condition explicit prevents an institution from assessing responsible tool use while secretly scoring students as though they had worked without tools, or from claiming independent mastery based on a task that allowed the system to perform the central reasoning.
Consider an assessment in which a student interacts with a virtual patient and submits a diagnosis. If the only observation is the final diagnosis, the assessor has product evidence: the answer is right, wrong, or partly right. That evidence may matter, but it does not show how the student took a history, what information they considered, or whether they noticed uncertainty. If the assessment also records the conversation, the sequence of questions becomes process evidence. If it captures how the student monitored their approach or reflected on the decisions they made, it adds metacognitive evidence. The claim can then be broader than “the student reached the correct endpoint.”
The OECD Digital Education Outlook 2026 describes this kind of medical-education example. Students worked with GenAI-powered virtual patients, and the assessment model combined product evidence such as diagnostic accuracy, process evidence such as the history-taking conversation, and metacognitive evidence such as interaction and clickstream data. In that example, combining the three streams predicted learner performance more reliably than product measures alone, while process data was the strongest standalone predictor. That is a useful result, but its meaning is specific. It does not show that any clickstream is a good measure of learning. It shows that, for this task and evidence model, how the student interacted carried information that the final diagnosis did not contain by itself.
The distinction can be applied to a writing assignment. The finished essay is product evidence. A sequence of drafts can show how the argument changed. A brief defence can show whether the student understands the central claim and can respond to a challenge. A reflection can reveal why they accepted one source or rejected an AI suggestion. A later essay on a related but unfamiliar problem can test transfer. Each observation answers a different question. None should be smuggled in as proof of all the others.
This matters more when a generative system participates in production. A clean final product may show that the student can direct, edit, or accept a system’s output. It does not automatically show that they can generate the reasoning independently. Conversely, an AI-assisted product may be legitimate evidence if responsible tool use is part of the intended claim and the assessment observes the student’s selection, verification, and judgment. The right evidence depends on what the institution is promising when it awards the grade.
A practical test is to write two columns before selecting an assessment format. In the first, state the human capability: explain a concept, solve a new problem, critique evidence, collaborate, communicate, or use a tool responsibly. In the second, list the smallest observations that would persuade a reasonable assessor that the capability is present. “The prose sounds polished” may be relevant to communication, but it is weak evidence of independent reasoning. “The student explains the structure, identifies a counterargument, and revises the claim when new evidence appears” is more closely tied to the intended capability.
The relationship is deliberately directional. A claim defines what the task should elicit; the task produces observations; observations become evidence only when they are relevant to the claim; and the inference remains qualified by its boundary. A large volume of observations does not repair a mismatch at the start. If the claim is independent mathematical reasoning, collecting more evidence about formatting or time spent in a document does not solve the problem.
From final products to human contribution
The shift toward process evidence is not an invitation to record everything a student does. More data can create new problems: privacy exposure, unequal surveillance, noisy measures, and false confidence in automated interpretation. The OECD describes learning analytics that can make strategies such as planning, searching, note-taking, and self-monitoring more visible, but it also notes that educators and learners need substantial data literacy to interpret analytic representations. A trace is an observation. It becomes meaningful evidence only through a defensible connection to the capability being assessed.
This distinction is especially important for interaction with AI. A conversation log may show that a student asked for a hint rather than a finished answer. That is useful context, but it does not prove that the student understood the hint. The stronger observation is what happens next: the student applies the idea, explains the choice, detects a limitation, or transfers the reasoning to another problem. Similarly, a record showing many revisions does not establish productive learning. Revisions can reflect thoughtful evaluation, random trial and error, or repeated acceptance of generated alternatives. The assessor needs a rubric that distinguishes these possibilities.
A good rubric maps observations to claims rather than rewarding whatever is easiest to count. Suppose the claim is that a student can evaluate an AI-generated explanation. Relevant criteria might include identifying a factual or logical weakness, checking it against an appropriate source, explaining the consequence, and producing a corrected version. The number of prompts used is not the capability. A polished correction without a reasoned account is incomplete evidence. The rubric can give separate weight to accuracy, verification, reasoning, and communication because those dimensions support different parts of the claim.
Evidence should also be read as a pattern rather than as a passcode. A student might produce a correct answer but give an explanation that contradicts it; that combination suggests a different inference from a correct answer paired with a coherent explanation and a successful variation. A student might make a small execution error while correctly identifying the method, diagnosing the cause, and repairing it. If the assessment claims to measure problem solving, the repair may be more informative than flawless first-pass execution. These judgments require an assessor to interpret relationships among observations, not simply add up visible artifacts.
Authentic context strengthens the design when it makes the target judgment meaningful. A public-health student might analyse a changing case rather than repeat a definition. A programmer might debug a function, explain an assumption, and modify the approach when an input condition changes. A language learner might respond in an adaptive conversation rather than select answers from a fixed list. These tasks resemble meaningful practice, but resemblance alone is not enough. If the system performs the substantive reasoning and the student merely approves the result, the task may be authentic as a workplace simulation while still weak as evidence of the student’s competence.
Staged work can close that gap. An assignment may ask for an initial position, a record of the evidence or feedback considered, a revised product, and a short explanation of the changes. The stages create several opportunities to observe diagnosis, judgment, and adaptation. They also support learning because students receive information while there is still time to use it. Yet staged work has its own boundary: drafts can be generated, edited, or reconstructed. A process record is not automatically genuine merely because it has dates attached. A brief live explanation or a later transfer task may be needed when the capability is consequential.
Oral defence and live performance provide another route. Asking a student to explain a decision, answer a variation, or recover from an error makes the relationship between product and understanding more visible. This does not make spoken performance universally fair or superior. It can introduce demands related to anxiety, language, disability, speed, or unfamiliarity with the format. Those demands must be addressed as part of validity. If rapid speaking is not the target skill, the assessment should not quietly score it as though it were.
The same reasoning applies to a closed-book or tool-removed check. It can be valuable when the claim concerns unaided recall or independent application. It can be misleading when the target is competent work with documentation or AI. The question is not whether the check is strict. The question is what the changed conditions reveal. A new problem can test transfer; an explanation can test conceptual ownership; a controlled task can test whether the student can perform the core operation without a system supplying it. Each choice carries a different inference.
Privacy is another boundary on process evidence. A platform that logs prompts, searches, revisions, and interaction patterns may help an educator support self-regulated learning, but it also collects information about students beyond the final answer. The educational benefit does not erase the need to explain what is recorded, why it is needed, who can see it, and how long it is retained. Data minimisation is not merely a compliance preference. Unnecessary traces increase the chance that an institution will mistake behavioural visibility for cognitive certainty or use data for purposes the student did not understand.
Fairness requires the same precision. If some students have better devices, faster connectivity, more experience prompting AI, or more opportunities to practise the assessment format, the observed performance may combine the target capability with unequal conditions. The assessment can provide standard access, permitted alternatives, practice opportunities, and accommodations, but those safeguards must be designed before scoring begins. A rubric cannot rescue a task that asks students to display a capability they had no equitable opportunity to develop.
Minimum sufficient evidence does not mean the smallest possible grade calculation. It means enough distinct, relevant observations to support the claim and its important boundary. Evidence should be added when it answers a different question or resolves a plausible ambiguity. If a final answer and an explanation already establish the intended competence, collecting every prompt may add surveillance without adding validity. If the answer could have been generated, a transfer task may add more information than a longer activity log. This is how an evidence model stays rigorous without turning the assessment into an indiscriminate data-collection exercise.
The documented conversational exam puts this architecture into a live setting. Its students code while explaining their reasoning, respond to prepared follow-up questions, and receive a rubric-based judgment across technical and conceptual dimensions. The next chapter examines how that format turns process and explanation evidence into a workable interaction—and why its success depends on careful preparation rather than on conversation alone.