VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

AI and Assessment | What Schools Should Recognise as Learning

Three students sit around open books and worksheets at a classroom table, reading, writing and discussing the work together.

Schools should recognise learning through evidence of what a student can explain, decide, produce and adapt under stated conditions. When AI is available, a strong finished product can show successful work with assistance. To establish what the learner understands, the assessment also needs carefully chosen opportunities to explain a decision, check a claim or respond to a changed requirement.

A useful starting point is to write the intended learning claim before designing the task: “This student can use capacity and time constraints to justify a feasible plan.” Then specify the evidence, permitted help and criteria that would support that claim. AI can expand the complexity students explore, while assessment keeps their developing knowledge and judgement visible.

  • State the capability being assessed and the decisions the result will inform.
  • Publish the conditions, including AI use, collaboration and access supports, before work begins.
  • Separate the quality of a product from evidence about each learner’s understanding.
  • Use a checked planning case to examine calculations, sources and practical constraints.
  • Moderate contrasting responses against shared criteria, then report strengths, limits and next steps.
  • Treat the rubric below as a classroom proposal, not a validated instrument or an official examination scheme.

CHAPTER 1 OF 18 · Define learning evidence

1. Begin with the claim a result will support

Back to contents

An assessment result makes a claim about a learner. “Can produce a timetable with an approved planning tool” is different from “can identify a bottleneck independently.” Both capabilities may be valuable. They require different evidence, and neither should quietly stand in for the other.

Make the claim specific enough to guide teaching. “Understands AI” is too broad for one assignment. “Checks an AI-generated schedule against the supplied room limits and explains a necessary correction” gives the student something concrete to practise. It also gives the teacher something observable to assess.

The purpose matters. A short classroom check may help a teacher choose tomorrow’s explanation. A result used to place a student in a course carries wider consequences and needs stronger supporting evidence. One engaging project should not become a permanent label about a learner’s intelligence, independence or future potential.

Not every useful observation needs a mark. During practice, a teacher can stop, explain a missing idea and let the student try again. The resulting response shows what happened after that teaching. Keeping this distinction visible allows assessment to guide learning without making every request for help feel like a failure that will follow the student into a final report.

For the case in this article, the main claim is that a learner can justify whether a plan meets several stated constraints. Supporting claims concern source selection, clear communication and responsible use of assistance. These are connected, but a weakness in one should remain visible rather than disappearing inside a broad judgement of “good work.”

This builds on AI-powered assignments that expand student thinking. The additional institutional question is what a school may reasonably conclude from the evidence those assignments produce. The answer should be useful to the next teacher as well as to the person awarding the mark.

Contents · Next chapter

CHAPTER 2 OF 18 · Define learning evidence

2. Use validity, reliability and fairness as design questions

Back to contents

The 2014 Standards for Educational and Psychological Testing, developed jointly by AERA, APA and NCME, connect validity to evidence supporting intended interpretations and uses of results. They also address reliability or precision and fairness. Applied cautiously here, these principles ask whether the evidence supports the intended claim, whether relevant judgements are sufficiently dependable, and whether irrelevant barriers distort what students can demonstrate. The Standards do not validate this article’s rubric or prescribe an AI classroom method.

For the proposed planning task, ask three practical questions. Does the task actually require reasoning about constraints? Would two informed teachers interpret the same response similarly? Can learners access the documents and express their reasoning without an avoidable obstacle becoming the main determinant of success?

A beautifully designed poster is weak evidence for mathematical reasoning if every calculation came ready-made and no explanation was requested. A correct spoken explanation may be strong evidence for that reasoning even when its speaker needs support to produce an organised written report. The report itself can still be assessed separately.

Dependable marking also needs a manageable set of criteria. If teachers are asked to judge creativity, leadership, ethics, confidence, reasoning, design and independence from one short performance, each label may receive very little evidence. Select what the lesson genuinely teaches and samples.

These questions should shape the task before students begin. Revising criteria after seeing an impressive submission makes expectations unstable. An assessment becomes easier to explain when its intended claim, required evidence and scoring decisions follow the same educational purpose.

Contents · Previous chapter · Next chapter

CHAPTER 3 OF 18 · Define learning evidence

3. Distinguish successful assistance from learning evidence

Back to contents

Keep four questions available: what can the student accomplish with the permitted tools; what can the student explain or do without answer-generating help; what remains available after a delay; and what can the student adapt to a changed problem? A single task need not answer all four, but its report should say which were examined.

Original research illustrates why the distinction matters. Bastani and colleagues’ 2025 high-school mathematics experiment found improved assisted practice performance with two GPT-4 tools. The more unrestricted tool was followed by weaker unaided performance than the control condition; teacher-informed safeguards largely removed that negative effect without establishing an unaided benefit. The unaided examinations followed practice within the sessions. This is evidence from a particular subject, setting and tool design, not a universal verdict on AI or a demonstration of long-term retention.

For assessment design, the practical implication is to collect the evidence that matches the claim. If the goal is competent planning with AI, examine the quality of that assisted work. If the goal includes understanding why the plan works, ask the learner to apply or explain the underlying constraint in a separate check.

Independence needs a precise definition. A calculator, accessible document format or established communication support may remain available. What is withheld in a particular check is assistance that supplies the reasoning being assessed. Record that distinction so a later reader does not mistake “no chatbot” for “no support of any kind.”

Contents · Previous chapter · Next chapter

CHAPTER 4 OF 18 · Define learning evidence

4. Publish the assessment conditions before work begins

Back to contents

Give students a short conditions statement alongside the task. For the example below, a teacher might permit an approved AI tool to suggest schedules during drafting, allow discussion within a planning group, and require every submitted claim to be checked against the supplied documents. Students would know that an individual follow-up uses those documents and a calculator, but no generated answer or peer solution.

Explain what must be submitted: a proposed timetable, a capacity argument, a note resolving conflicting information, and a brief record of one important revision. Explain which components receive group feedback and which provide evidence about individuals. Students should not discover the individual check only after investing their effort in a different kind of task.

Make the assistance statement functional. “AI was used” gives little information. “The tool suggested a rotation table; we checked every room allocation and corrected its claim about extra participants” identifies the role of the tool and the human decision. A concise record is sufficient for this proposal; collecting a large transcript is not an educational goal.

Provide a route for students without individual tool access. The teacher can supply the same prepared AI draft for everyone to critique. If evaluating that draft is the target, this can give the class a common starting point. It does not establish that each student can independently operate an AI application, so the report should not make that additional claim.

Finally, tell students how to ask for clarification and how changed conditions will be recorded. An interrupted connection or an unreadable source file needs a practical response before it becomes an unexplained difference in results.

Contents · Previous chapter · Next chapter

CHAPTER 5 OF 18 · Work through the planning case

5. Set a fictional planning task with explicit constraints

Back to contents

Consider a fictional Discovery Morning for 84 learners. Every learner must attend three different workshops: Design in the Studio, Evidence in the Seminar Room, and Data in the Computer Lab. Each workshop lasts 25 minutes. The event runs from 9:00 to 10:35, with a five-minute opening briefing in each group’s first room and two five-minute transitions between workshops.

The current coordinator’s sheet gives learner capacities of 30 in the Studio, 28 in the Seminar Room and 28 in the Computer Lab. These figures already allow for the facilitators; students should not subtract them again. One dedicated facilitator is available in each room throughout the event. Groups stay together, and rooms cannot be exchanged or enlarged within this task.

The supplied pack contains the current coordinator’s sheet, a room list confirming these limits, and a clearly dated flyer from the previous year that lists 32 places in the Seminar Room. It also contains a fictional AI suggestion that adding room capacities establishes how many people can attend every workshop. All documents, people and operational limits in this case are invented for teaching.

The assignment asks students to produce a feasible plan for 84 learners, explain whether 86 could attend every workshop under the same conditions, and identify which source controls the room limit. This combination requires calculation, interpretation and judgement about the relevance of evidence.

State the simplifications. The five-minute movement allowance is given rather than measured. This is a classroom model, not permission to run a real event. A real organiser would need locally verified arrangements; the learner’s job here is to reason accurately from the supplied constraints.

Contents · Previous chapter · Next chapter

CHAPTER 6 OF 18 · Work through the planning case

6. Build and check a feasible plan for 84 learners

Back to contents

Divide the learners into three groups of 28. Each group fits all three rooms, including the two smaller spaces. A rotation can then give every group one visit to each workshop without placing two groups in the same room at the same time.

One feasible timetable for the fictional Discovery Morning
TimeGroup A: 28 learnersGroup B: 28 learnersGroup C: 28 learners
9:00–9:05Opening briefingOpening briefingOpening briefing
9:05–9:30Design: StudioEvidence: Seminar RoomData: Computer Lab
9:30–9:35TransitionTransitionTransition
9:35–10:00Evidence: Seminar RoomData: Computer LabDesign: Studio
10:00–10:05TransitionTransitionTransition
10:05–10:30Data: Computer LabDesign: StudioEvidence: Seminar Room
10:30–10:35Remaining bufferRemaining bufferRemaining buffer

The required time is 5 + (3 × 25) + (2 × 5) = 90 minutes. The 95-minute window therefore leaves five minutes. Calling that remaining time a buffer does not allow a student to shorten a required workshop elsewhere in the plan.

Check the table in both directions. Reading a row checks simultaneous room use. Reading a group’s column checks that its learners attend all three different workshops. Adding all the numbers in the table would not answer either question.

A learner can demonstrate these checks with annotations, a short explanation or a teacher-observed walkthrough. The representation may vary, but the evidence should identify why no room is overfilled, no workshop is missed and the event fits the time window.

Contents · Previous chapter · Next chapter

CHAPTER 7 OF 18 · Work through the planning case

7. Examine the bottleneck and the conflicting source

Back to contents

The sum 30 + 28 + 28 = 86 describes the available learner places across the rooms during one workshop period. It does not show that 86 different learners can each attend every workshop across the morning. The Computer Lab can accommodate only 28 learners per period. Across three periods it provides 3 × 28 = 84 opportunities to attend Data.

Because every learner needs that workshop, 86 learners cannot all complete the programme under the stated conditions. The Seminar Room gives the same limit. Two spare Studio places cannot replace a missing Data or Evidence place. The schedule for 84 demonstrates that this upper limit is achievable in the simplified model.

Adding a fourth full workshop period would require another 25-minute session and another five-minute transition: 120 minutes including the briefing. That exceeds the 95-minute window. A proposal to add staff, change rooms or shorten sessions would change the problem’s conditions and needs to be identified as a requested change.

The previous year’s flyer does not establish this year’s Seminar Room capacity. The current room list and coordinator’s sheet directly address the present arrangement. Students should explain that relevance; simply calling one document “official” does not show that it governs this event.

If two current, equally applicable documents disagreed, a defensible answer could state the uncertainty and seek clarification. For this fictional pack, the applicable limit is supplied clearly. The assessment should recognise both a correct calculation and the judgement about which information that calculation may use.

Contents · Previous chapter · Next chapter

CHAPTER 8 OF 18 · Gather and interpret evidence

8. Collect process evidence that explains a decision

Back to contents

The final timetable answers whether the proposed event can run within the model. A small amount of process evidence can answer a different question: which decisions did the learners inspect and why did they revise them? Ask for one meaningful before-and-after example rather than a diary of every action.

A useful record might say: “Our draft used the total of 86 seats to claim that 86 learners could attend. We checked how many Data places were available across the morning and changed the limit to 84.” It names the initial claim, the check and the resulting decision.

A weak record might say: “We improved our answer after reviewing it.” That statement may be sincere, but it does not reveal the reasoning. Feedback can ask the group to identify the particular claim and the evidence that changed it. This is a request for educationally relevant information, not a judgement about honesty.

A neat process record is still a submitted product. Use it alongside the plan and an appropriate individual check rather than treating its existence as proof of independent understanding. Similarly, a messy draft can contain a valuable correction even when the finished explanation needs work.

Keep the workload proportionate. The proposed evidence packet is a timetable, a short capacity explanation, a source note and one revision. Students should spend their effort analysing the problem, not manufacturing an impressive record of activity. The teacher should be able to locate the decisive evidence without reading an entire conversation history.

Contents · Previous chapter · Next chapter

CHAPTER 9 OF 18 · Gather and interpret evidence

9. Separate group achievement from individual evidence

Back to contents

A group can produce a feasible schedule even when different members understand different parts. Recognise the shared achievement: the team assembled a usable plan. Then gather individual evidence for any claim about each learner’s reasoning. Membership of a successful group does not, by itself, answer that question.

Give every member a small common check. For example, reduce the Computer Lab capacity to 26 while keeping the other conditions. Ask how many learners can now attend every workshop and request a justification. Three periods provide 78 Data places; three groups of 26 show that 78 is feasible within the existing rotation.

Students may explain in an agreed accessible format. The question should examine capacity reasoning, rather than surprise them with a presentation skill that was never taught or announced. Follow-up questions can clarify meaning, but record when a question supplies the crucial idea rather than merely asking the learner to express it.

If collaboration is also an intended learning goal, collect evidence of collaboration itself. An observed exchange in which a student invites a peer’s challenge and uses it to revise the plan is more informative than counting how many messages that student posted. Decide beforehand what observations are feasible to collect.

The separate article on building shared understanding with AI develops those group processes. Here the assessment principle is to keep the unit of the claim clear: a team product, an individual explanation and a collaborative contribution are related but distinct pieces of evidence.

Contents · Previous chapter · Next chapter

CHAPTER 10 OF 18 · Gather and interpret evidence

10. Use a modest rubric tied to this task

Back to contents

The following proposed rubric is an instructional aid for this fictional assignment. It has not been validated, trialled for dependable scoring or adopted as an official standard. Its purpose is to make expectations discussable. Teachers would need to test and refine its descriptors before using it for consequential decisions.

Proposed criteria for the Discovery Morning evidence packet
CriterionEvidence to examineDemonstrated for this taskA reason to seek more evidence
Constraint reasoningCapacity argument and calculationsExplains why 84 is feasible and why 86 cannot complete every workshop.Uses total seats alone or overlooks a required workshop.
Use of sourcesSource note and applied limitsUses the current applicable documents and explains why the old flyer does not control.Chooses a convenient number without resolving conflicting information.
Usable planTimetableSpecifies all groups, rooms and periods without clashes, omissions or excess time.A reader must invent a missing allocation or repair the schedule.
Review and revisionOne decision recordIdentifies a substantive check and explains whether it changed the conclusion.Reports activity without showing what was checked.

For each criterion, record demonstrated, partly demonstrated or not yet demonstrated, with a brief reason. Add “not assessed” when no relevant opportunity was provided; absence of an opportunity is different from an unsuccessful response. These categories describe evidence from this task, not fixed levels of a person’s ability.

Keep the individual follow-up alongside the rubric rather than silently treating the group packet as an individual result. Later retention or transfer evidence also belongs in its own dated entry. If a school requires a numerical mark, it should establish and explain its weighting and any essential requirements before the assignment. This example does not supply a universal conversion into marks.

Contents · Previous chapter · Next chapter

CHAPTER 11 OF 18 · Moderate and report learning

11. Compare two contrasting fictional responses

Back to contents

Response A: a complete plan with a weak extension

Response A presents a clear timetable that correctly serves 84 learners. It uses the current room figures, includes both transitions and identifies the five-minute buffer. Its source note correctly rejects the old flyer. However, it concludes that 86 learners could attend every workshop because the room capacities add to 86.

Its revision record explains a genuine correction to the timing of the second transition. The group checked an important detail successfully, but the capacity argument remains incomplete. In the individual follow-up with a Lab capacity of 26, the fictional learner again adds room capacities and gives 84 as the answer.

Response B: strong reasoning with an incomplete handover

Response B explains that the Lab and Seminar Room each supply only 84 workshop places across three periods. It states that three groups of 28 can each visit all three rooms by rotation. It uses the current documents appropriately and rejects the 86-learner claim. Its revision record shows how it corrected that claim after checking the AI suggestion. The timetable, however, leaves the final period’s group-to-room allocations blank.

In the same individual follow-up, the fictional learner calculates 3 × 26 = 78 and explains that three groups of 26 fit every room. This supports the specific capacity claim under the stated check conditions. It does not repair the incomplete group timetable.

Both responses contain achievement and a next step. A needs further work on the bottleneck. B needs to complete the operational communication. The difference is educationally useful precisely because the assessment does not allow attractive presentation, or one strong explanation, to settle every criterion at once.

Contents · Previous chapter · Next chapter

CHAPTER 12 OF 18 · Moderate and report learning

12. Moderate the judgement before final reporting

Back to contents

For a small classroom trial, two teachers could independently review these responses using the proposed criteria. Each would identify the evidence behind a judgement before discussing it. This is a suggested moderation process, not a claim that two raters or one discussion establishes statistical reliability.

A first teacher might initially favour A because the plan is ready to use. A second might favour B because the bottleneck explanation is stronger. The criteria resolve the apparent disagreement: A demonstrates the usable-plan criterion but only partly demonstrates constraint reasoning; B demonstrates constraint reasoning but only partly demonstrates a usable plan.

Both demonstrate appropriate source use. Both provide a substantive review example, although neither example proves that every part of the submission was checked. The individual follow-up adds separate evidence: A has not yet demonstrated the revised-capacity reasoning in that check, while B has.

Moderators should compare the meaning of “partly demonstrated.” Does it cover a correct base case with an incorrect extension? Does a missing final allocation count as an omission even when a reader could infer it? Agree those interpretations, retain the examples as local anchors, and revisit other affected submissions if a clarification changes how they should be judged.

Agreement about two samples also leaves other questions open. The same learner may face a different demand when the plan involves unfamiliar quantities or a different source conflict. If the school wants to report a broader planning capability, it needs suitable evidence across more than this one problem. Repeated marking of the same narrow sample cannot provide that wider coverage.

Do not resolve disagreement by averaging unrelated impressions. If a required grade combines criteria, apply the previously announced scheme and keep the underlying profile available. Where evidence is insufficient, obtain a focused clarification or record the limitation. A confident-looking number should not conceal an unresolved question about what the learner demonstrated.

Contents · Previous chapter · Next chapter

CHAPTER 13 OF 18 · Moderate and report learning

13. Report a learning profile that supports action

Back to contents

A useful report names the capability, the evidence and the conditions. It also distinguishes a demonstrated weakness from an area that has not yet been assessed. This lets parents and future teachers understand what to build on without needing to reconstruct the entire assignment.

Illustrative report for the learner associated with Response B
AreaEvidence and conditionsNext step
Group planning productAI assistance permitted; capacity argument sound, final room allocations incomplete.Complete the final period and check each group’s route.
Individual understandingWith source pack and calculator, without generated or peer answers, justified the 78-learner limit.Explain how a changed requirement could alter the limiting resource.
Later retentionNot yet assessed.Revisit the underlying capacity relationship after a stated interval.
TransferImmediate adaptation within the same workshop setting observed; a different context not yet assessed.Try a new planning problem with the relevant information supplied.

This record does not assign an “AI competence” number. Operating a tool, checking a source, reasoning mathematically and explaining a decision are different claims. A broad composite would require a clear purpose, an explicit rationale and evidence that it supports its intended interpretation.

The report can still be concise. One sentence about a demonstrated strength and one about the next learning action may be enough for routine feedback. The essential detail is what the learner did under which conditions. “Strong with AI” gives the next teacher much less to use.

Where reports include group achievement, label it visibly. Where an individual received a hint that supplied the bottleneck idea, record that before describing the answer as independently reasoned.

When reporting progress, compare the demands as well as the results. An earlier assisted group product and a later individual explanation are not interchangeable measurements. Describe the new capability evidenced by the later check; a difference between their marks would not, by itself, quantify how much the student has learned.

Contents · Previous chapter · Next chapter

CHAPTER 14 OF 18 · Moderate and report learning

14. Add later evidence without promising universal transfer

Back to contents

A teacher could return to the capacity relationship the following week, using a different set of numbers and the same agreed access supports. That interval is a practical example, not an experimentally established schedule for this task. Record any relevant intervening teaching or practice when interpreting the result.

A later correct answer supports retention of what that check actually sampled. It does not establish that every part of the original assignment was retained. If the student solves the arithmetic but cannot identify which resource limits the plan, the next lesson should address the interpretation rather than repeat only multiplication practice.

For a different context, consider an invented demonstration programme where each visitor must use one specialist equipment station. Supply its operating periods, places and timing. Ask students to identify the limiting opportunity and justify a feasible arrangement. Changing the context should still leave enough information for learners to reason without needing untaught specialist knowledge.

The degree of change matters. New numbers in the same workshop scenario offer a narrower adaptation check than an unfamiliar problem with a different structure. Label the evidence accordingly. Successful performance on one variation is a useful observation, not proof of general problem-solving ability.

The companion article on learning transfer in new situations develops this distinction. For school reporting, the practical rule is simple: say what changed, when the check happened and what help remained available. Those details make later evidence interpretable.

Contents · Previous chapter · Next chapter

CHAPTER 15 OF 18 · Design fair school practice

15. Design fair opportunities to demonstrate the target

Back to contents

Singapore’s MOE AI in Education Framework identifies Agency, Inclusivity, Fairness and Safety as guiding principles. It provides a responsible-use framework; it does not endorse this assignment or validate its scoring. The following choices are proposed ways to consider those concerns in this particular classroom design.

Start with access to the task. Give learners readable documents, explicit vocabulary and a usable timetable format. Where the aim is capacity reasoning, unnecessary difficulty decoding a crowded page should not become the main challenge. Keep established access supports available and record any support that directly supplies the reasoning being assessed.

A spoken explanation, annotated table or supported communication format may make the reasoning visible. Decide whether the format preserves the intended claim. If concise written communication is itself an assessed objective, provide a separate opportunity to demonstrate it rather than silently dropping the objective.

Tool access also needs a plan. A common teacher-prepared AI response makes critique possible without requiring students to buy access or reveal personal information to an external service. If operating the tool is an explicit target, provide an appropriate supported opportunity to do that and assess it separately.

Review who could participate and what prevented participation. A missed check caused by inaccessible materials calls for an accessible opportunity, not an inference about weak reasoning. Designing learning that more students can use explores these arrangements further. Fair assessment begins while the task is being designed, before a teacher has a score to explain.

Contents · Previous chapter · Next chapter

CHAPTER 16 OF 18 · Design fair school practice

16. Keep teacher use of AI accountable and inspectable

Back to contents

AI can be assigned a limited role in preparing this assessment: suggesting an alternative fictional timetable, generating a deliberately flawed capacity argument, or helping draft feedback against teacher-selected criteria. Each generated item needs checking against the mathematical model and the learning objective before students encounter it.

For example, a generated variation might reduce the Lab capacity but accidentally change the number of workshop periods as well. That changes the reasoning demand. A teacher who intended a single-variable follow-up should repair the item before using it to compare learners’ responses.

If AI assists with comments, the responsible teacher should verify the evidence reference and final judgement. A proposed comment that says “You ignored the source date” is inappropriate when the learner explicitly explained it. The teacher should be able to point to the actual passage supporting any reported strength or gap.

For this proposal, retain the task, criteria, selected anchor responses and a short note of important moderation decisions. Avoid collecting personal data that the assessment does not require. Follow the institution’s approved arrangements for any student material handled by a tool; a fictional demonstration pack can be used when real submissions are unnecessary.

Give learners a route to request clarification or correction of a judgement. A student may show that a table was misread or a response was overlooked. The review should return to the evidence and criterion. Neither an automated suggestion nor an initial teacher interpretation should make a correctable error permanent.

Contents · Previous chapter · Next chapter

CHAPTER 17 OF 18 · Design fair school practice

17. Keep classroom proposals distinct from examination rules

Back to contents

A classroom assignment can deliberately assess responsible work with AI when the teacher has authorised that use. This does not establish what is permitted in a formal examination, national qualification or externally assessed coursework. The relevant current rules and instructions govern those settings.

As one jurisdiction-specific example, Ofqual’s advice note published on 27 April 2026 explains that AI use may itself form part of what some qualifications assess, while other components require protection of their intended assessment. The note applies to England and supports existing requirements; it does not create new regulatory requirements. It is not Singapore examination guidance.

For a Singapore classroom, communicate the distinction directly: these are the conditions for this learning activity; formal assessments follow their own published instructions. Students should not have to infer permission from seeing an AI-supported activity in another subject or setting.

There is also a distinction between feedback and consequential certification. A teacher may use an exploratory task to learn what students need next. Turning that task into a qualification decision requires evidence and procedures appropriate to the larger consequence. The classroom rubric here should not acquire that authority simply because its descriptors are clear.

Contents · Previous chapter · Next chapter

CHAPTER 18 OF 18 · Design fair school practice

18. Make recognition of learning useful to the institution

Back to contents

A school can begin with one department, one existing assignment and a small set of intended claims. Ask teachers to identify where the current evidence is strong and where it depends on an unstated assumption about assistance. Change the evidence collection only where it resolves a real uncertainty.

Try a small common follow-up, moderate a few contrasting responses, and review whether the resulting feedback is more useful. Examine the time required, the accessibility of the task and whether teachers can explain disagreements. This is a local improvement cycle, not a research design that by itself establishes a causal benefit from AI.

Teachers can start by rewriting one broad criterion into an observable capability. Students can ask which decisions they must explain and what help is permitted. Parents can ask what a strong finished product tells the school, what has been checked individually and what the next learning step will be. These questions turn a mark into information that can guide further work.

The principle connects to curriculum progression before and alongside delegation. Foundations give learners something with which to question an output. AI can then help them explore more alternatives, inspect more demanding problems and produce useful work. Assessment should make both the expanded performance and the growing human capability visible.

When schools recognise those capabilities accurately, later teachers and institutions receive a clearer account of what a learner can contribute and what support remains useful. That is a concrete way for human–AI coexistence to strengthen education: the work can become more ambitious while the claims about learning become more carefully grounded.

Contents · Previous chapter · Continue the series

Sources and scope

The event, submissions, rubric, moderation procedure and follow-up checks are original teaching proposals. They have not been experimentally validated. Sources were checked on 15 November 2026.

Continue the AI, Education and Civilisation series

Follow the argument from human learning into the institutions that shape shared life. Each article develops its own questions through practical examples and clear evidence limits.

Learning and human capability

Culture and shared life

Knowledge and public institutions

Economics, organisations and work

Learning design and participation

Assessment, feedback and school practice

Return to How Super Intelligence Works.