How Should Schools Assess Learning When Students Use AI?
Schools should not try to make every assignment “AI-proof.” They should design assessment so that important learning remains visible when artificial intelligence is prohibited, permitted within boundaries or deliberately included as part of the task.
That requires more than inspecting the final product. A robust system combines evidence from what the student produces, how the work developed, what the student can explain, whether the learning transfers and how the student judges AI output. The central question is therefore not simply, “Did the student use AI?” It is: What knowledge, reasoning, skill and judgment can the school validly attribute to this learner?
This question has become urgent because generative AI can raise the quality of a finished essay, solution, presentation or code sample without guaranteeing durable learning. The OECD Digital Education Outlook 2026 distinguishes successful task performance from learning gains and warns that cognitive outsourcing may create fluent performance without secure capability. Assessment must therefore look beyond completion and recover evidence of the learner’s own understanding.
Begin with the claim the assessment needs to support
Every assessment makes an implicit claim. A mathematics task may claim that a student can select and justify a method. A history essay may claim that the student can evaluate evidence. A science investigation may claim that the student can design a fair test and interpret data. A language task may claim that the student can communicate appropriately for a purpose and audience.
AI complicates these claims only when the assessment has not specified which performance must be attributable to the learner. If a chatbot selects the method, drafts the explanation and corrects the errors, the final answer cannot by itself support a strong claim about the student’s reasoning. If the learning goal is to evaluate AI-generated arguments, however, the use of AI is not interference; it is part of the construct being assessed.
Assessment design should therefore start with three decisions. The school identifies the precise capability, decides which forms of assistance are compatible with that capability and selects the evidence needed to justify the judgment. Technology rules follow from the construct, not the other way around.
Five channels of evidence for AI-era assessment
No single channel is sufficient in every context. A finished product may show quality but conceal authorship. A process log may show activity but not understanding. An oral explanation may reveal reasoning but be affected by language, anxiety or accessibility needs. Transfer is powerful but must remain aligned with the original target. AI judgment is essential when critical tool use is itself an outcome.
| Evidence channel | What it can show | What it cannot prove alone | Useful design move |
|---|---|---|---|
| Product | Quality, disciplinary conventions, accuracy, coherence and completeness of the final work. | Who performed each intellectual step or whether the learner can reproduce the reasoning. | Mark the artifact against clear criteria, then connect selected criteria to another evidence channel. |
| Process | Planning, drafts, revisions, feedback use, source checking and changes in reasoning. | Understanding merely because many steps or screenshots were recorded. | Request a small number of meaningful checkpoints rather than an exhaustive surveillance record. |
| Explanation | The learner’s ability to justify choices, connect evidence and reconstruct key reasoning. | Long-term retention or transfer to a genuinely different context. | Use a brief viva, conference, annotation or written rationale focused on high-value decisions. |
| Transfer | Whether the learner recognizes and adapts the underlying knowledge in a changed case. | Every element of a large project or the full quality of collaborative work. | Change surface features while preserving the intended principle, method or disciplinary practice. |
| AI judgment | Purposeful use, verification, bias awareness, attribution, revision and the decision to reject output. | Independent subject mastery unless paired with unaided evidence. | Ask the learner to critique a generated response and explain how assistance changed the final decision. |
Triangulation does not mean adding five separate tasks to every assignment. It means selecting a proportionate combination. A low-stakes homework task may need only the product and one reflective explanation. A major project may justify process checkpoints, an oral defense and a short transfer task. A regulated qualification may prescribe different rules entirely. The evidence burden should rise with the stakes and with the difficulty of attributing performance.
Use three assessment modes instead of one universal AI rule
A coherent school assessment system includes tasks completed without AI, tasks with bounded AI support and tasks where critical AI use is itself assessed. Treating one mode as universally correct produces predictable weaknesses. If every task bans AI, students may never learn to use it responsibly and enforcement becomes unrealistic. If every task permits unrestricted use, schools lose evidence of unaided knowledge and skill. If every task requires AI, access, privacy and pedagogical fit are ignored.
| Assessment mode | When it is appropriate | Evidence priority | Typical risk |
|---|---|---|---|
| Independent performance | When recall, fluent execution, foundational knowledge or unaided reasoning is the construct. | Direct observation, timed or supervised work, explanation and transfer under accessible conditions. | Confusing independence with removal of legitimate accommodations or overusing controlled tests. |
| Bounded AI support | When AI may help with brainstorming, feedback, language support, examples or revision without replacing the target thinking. | Product plus selected process evidence, disclosure, explanation and an independent check. | Rules remain vague, so students cannot tell what support is legitimate. |
| AI-integrated performance | When evaluating, directing or improving AI output is part of the learning outcome. | Quality of verification, disciplinary judgment, revision, attribution and defense of decisions. | Rewarding polished prompting while neglecting subject knowledge and independent capability. |
The three modes should appear across a curriculum rather than inside every lesson. Students need opportunities to build foundations without AI, to use support under clear boundaries and to demonstrate sophisticated collaboration with AI. This progression is consistent with the OECD’s human-centered emphasis on developing knowledge and skills before and alongside purposeful AI use.
A practical assessment-design sequence
The design process begins by writing the intended learning in language that names the cognition or practice. “Write an essay” describes a product. “Build and defend a causal interpretation using contextualized evidence” describes learning. “Create a presentation” is a format. “Select, organize and communicate evidence for a defined audience” describes capability.
The teacher then identifies the vulnerable intellectual moves: the parts AI could perform while leaving the student’s competence uncertain. In an essay, those moves may include selecting evidence, structuring the argument and interpreting sources. In coding, they may include decomposing the problem, choosing an algorithm and debugging. In mathematics, they may include recognizing the problem type and justifying the method.
Next, the permitted role of AI is defined in operational language. “AI allowed” is not enough. Students need to know whether AI may generate questions, propose examples, critique a draft, edit language, write code, solve steps or create the final text. The boundary should state both what the tool may do and what the learner must do.
The teacher then selects the smallest combination of evidence that can support the assessment claim. A short explanation may be enough to confirm one key decision. A transfer problem may be more appropriate when imitation is a major risk. A process checkpoint may matter when planning and revision are learning outcomes. The final stage is to check fairness: whether the design preserves accommodations, provides realistic access, minimizes unnecessary data and can be explained consistently to students and families.
The AI-era assessment design canvas
| Design question | Decision to record | Example |
|---|---|---|
| What is the learning claim? | Name the knowledge, reasoning, skill or judgment that the result should support. | The student can evaluate how evidence supports a causal historical claim. |
| Which intellectual moves matter? | Identify the decisions that must be attributable to the learner. | Select evidence, interpret its significance and consider source reliability. |
| What may AI do? | Define permitted assistance precisely. | Generate counterquestions and offer feedback on clarity; not select the evidence or write the final analysis. |
| What must the student disclose? | Specify tool, purpose, meaningful prompts or outputs retained, verification and contribution. | A short statement explaining where AI feedback changed the draft and what was rejected. |
| Which evidence channels are needed? | Choose a proportionate combination of product, process, explanation, transfer and AI judgment. | Final paragraph, source annotation, three-minute defense and one new source comparison. |
| How will fairness be protected? | Preserve accommodations, alternatives, access and data minimization. | Provide an approved school account or an equivalent non-AI feedback pathway. |
| What will happen if evidence conflicts? | Define review, clarification and academic-integrity procedures before the task. | Request further explanation and examine drafts; do not infer misconduct from style alone. |
A shared canvas can improve consistency across departments without forcing identical tasks. The history, mathematics and art departments can make different decisions while using the same design questions.
What this looks like in different subjects
In mathematics, an AI-supported problem set can still generate valid evidence if students identify the method, annotate the reasoning, locate an error in a generated solution and solve a structurally similar problem independently. The assessment should not award full credit merely because the final answer is correct. It should make method selection and justification visible.
In science, students might use AI to brainstorm hypotheses or critique the clarity of a method, but measured observations must remain distinct from generated claims. A short defense can test whether the learner understands variables, uncertainty and the relationship between evidence and conclusion. A transfer question can place the same scientific principle in a new system.
In history and social sciences, students can compare an AI-generated interpretation with primary and secondary sources. The important evidence lies in provenance, contextualization, corroboration and the student’s defense of why one interpretation is stronger. Generated quotations and citations require direct checking against the original source.
In language and writing, AI may support feedback, vocabulary exploration or revision, but the assessment claim must determine how much authorship remains with the learner. If independent composition is the goal, schools need a sample produced under appropriate conditions. If editing and rhetorical judgment are the goals, students can explain which revisions they accepted and why. The broader challenge of applying understanding beyond the practiced example is developed in From Brain Games to Real Learning.
In project-based learning, process evidence can be collected at decision points rather than through continuous logging. A proposal, annotated source choice, prototype review and final defense often reveal more than hundreds of screenshots. The purpose is to understand the development of the work, not to reconstruct every keystroke.
Disclosure should be meaningful, not bureaucratic
A disclosure statement should help the assessor interpret the work. Merely writing “AI was used” gives little information. A useful statement identifies the tool, the purpose, the parts of the process affected, the important output retained, the checks performed and the student decisions that followed.
Schools should keep the burden proportionate. Requiring complete prompt transcripts for every low-stakes task can create privacy problems, encourage performative compliance and overwhelm teachers with records they cannot meaningfully evaluate. More detailed retention may be appropriate where qualification rules explicitly require it. For ordinary classroom work, a concise contribution statement and selected evidence are often sufficient.
Students also need examples of acceptable and unacceptable use before assessment. They should not discover after submission that brainstorming was allowed but generated wording was not, or that translation assistance required acknowledgment. Clarity supports integrity more effectively than ambiguous warnings.
Academic integrity needs evidence, not guesswork
A sudden change in style, an inaccurate citation or a detector score may justify a conversation; none should automatically determine misconduct. AI systems and detection methods operate under uncertainty, while students may legitimately receive editing, translation or accessibility support. Schools should examine the full evidence: the assignment rules, drafts, sources, disclosures, teacher observations, student explanation and applicable institutional procedure.
A brief oral or written defense can clarify authorship and understanding, but it should not become an improvised interrogation. Questions should focus on the learning claim: why a source was selected, how a method works, what changed during revision or how the student would handle a nearby case. If concerns remain, schools should follow a fair, documented academic-integrity process.
Generative AI also creates a positive integrity outcome when assessment asks students to verify claims, expose uncertainty and reject fabricated evidence. Responsible skepticism becomes part of the curriculum rather than an enforcement message delivered only after a problem occurs.
High-stakes and regulated assessment must be treated separately
Classroom innovation does not override qualification rules. In England, Ofqual’s AI in coursework resources for schools, published on 9 March 2026, emphasizes that students must not submit AI-generated coursework as their own and that permitted use must be disclosed. The Joint Council for Qualifications’ coursework guidance effective from 1 September 2026 requires students using AI-generated content in permitted circumstances to acknowledge the tool and retain relevant prompts and generated material for review.
These requirements apply within their stated jurisdiction and qualification context. Other countries, examination boards and institutions may use different rules. Schools should maintain a clear separation between formative classroom experiments, internally assessed work and regulated qualifications. Teachers and students need to know which rule set governs the task before work begins.
Accessibility remains essential in high-stakes assessment. Reasonable adjustments and approved assistive technologies should not be confused with unauthorized AI assistance. The construct being assessed and the governing rules determine what support is legitimate.
A proportionate school implementation plan
Implementation should begin with an audit of existing assessment, not a new AI policy written in isolation. Departments identify tasks where the finished product carries too much of the judgment, where AI boundaries are ambiguous and where independent evidence already exists. Schools may discover that oral questioning, practical demonstrations, drafts, performances and teacher observation already provide useful triangulation.
A small pilot can then redesign one recurring task in each participating subject. Teachers use the design canvas, agree the AI mode and collect a manageable combination of evidence. After marking, they compare whether the new design improved confidence in the assessment claim, changed workload, affected accessibility or exposed new misconceptions.
The next step is moderation. Teachers compare judgments using real student work and refine prompts, descriptors and disclosure expectations. Students should also be invited to explain where the rules were clear, where they were burdensome and how AI actually entered the process. Policy becomes more credible when it reflects observed practice rather than imagined behavior.
School leaders should review the system on a schedule because tools and external rules change. The review should examine validity, reliability, workload, equity, privacy, student learning and academic-integrity cases. It should not count the number of AI declarations as a measure of success.
This work connects directly to building student independence in the classroom. Assessment is not only a way to certify what students know; it shapes what they practice. If every important intellectual move can be delegated without later explanation or transfer, the curriculum may reward completion while weakening independence. If students learn when to work unaided, when to use bounded support and how to evaluate AI critically, assessment can strengthen both integrity and capability.
Examples of AI experimentation in schools around the world show that institutions are trying different models. The most defensible models will be those that can answer not only what technology was used, but what students learned and how that learning was established.
Schools do not need an impossible AI-proof assessment system. They need a valid, transparent and proportionate evidence system. When product, process, explanation, transfer and AI judgment are combined deliberately, teachers can make stronger claims about learning without turning every classroom into a surveillance environment.
Educators who want to strengthen diagnostic questioning, observation, feedback and the progressive transfer of responsibility can explore the Become a Learning Coach training pathway, the book Become a Learning Coach, and the wider Learning Coach English article library.
