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 School Leadership | Turning Classroom Trials into Reliable Practice

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

A school can begin using AI responsibly by naming the learning problem it wants to solve, deciding what evidence would count as improvement, and giving someone responsibility for reviewing the result. Before extending a classroom trial, leaders should be able to explain what learners can now do, which support they needed, what the approach cost in staff time, and what still needs to change.

This guide develops a practical way to make those decisions. It follows an explicitly fictional trial involving twenty-four learners and two teachers, separates supported work from individual learning checks, and shows how a school could respond to uneven access, misleading outputs and hidden workload. The proposed methods are examples for professional discussion; official frameworks and research guidance are identified separately where they inform the argument.

  • Start with a specific learning difficulty and a bounded teaching approach.
  • Plan evidence, access arrangements, staff support and review responsibilities before the trial.
  • Read learning results alongside the conditions under which they were produced.
  • Count preparation, checking, support and follow-up when examining workload.
  • Keep the approach correctable, with clear decisions about continuing, revising, pausing or expanding it.

CHAPTER 1 OF 19 · Define a worthwhile trial

1. Begin with a learning problem that people can recognise

Back to contents

A useful starting question is concrete: can learners distinguish what a short source actually supports from an attractive claim that goes beyond it? That question gives a department something to teach and observe. It also helps a leader ask whether an AI tool contributes something useful to the lesson, rather than treating the presence of the tool as sufficient evidence of progress.

Imagine that teachers notice learners quoting a sentence accurately but drawing an unsupported conclusion from it. One proposed response is to let learners examine several candidate explanations, including a teacher-checked AI response, and identify the exact words that justify or weaken each explanation. The intended learning is the connection between evidence and judgement. Generating more explanations is only one possible activity within that purpose.

A leader can ask the department to write a short problem statement using three parts: the observed difficulty, the proposed change in teaching, and the learner action that would demonstrate progress. For example: learners will compare a claim with a source, revise an overstatement, and explain why their revision fits the available evidence. That statement is specific enough for colleagues to question and improve.

Keep the first claim modest. A successful lesson on one short source does not establish general critical thinking, mastery of every subject, or readiness for unrestricted AI use. It provides evidence about a particular capability under particular conditions. The broader educational ambition can remain large while the first teaching decision stays precise.

Contents · Next chapter

CHAPTER 2 OF 19 · Define a worthwhile trial

2. Read official guidance at its proper scope

Back to contents

Singapore’s Ministry of Education describes its AI in Education Framework through four principles: Agency, Inclusivity, Fairness and Safety. Its public information page, updated on 10 September 2026, also directs teachers planning small-scale localised trials in one or a few schools to an implementation guide and checklist requiring MOE Intranet access. That public statement identifies an existing institutional process. This article has not accessed the restricted checklist and does not substitute for it. Read the MOE framework information.

The Education Endowment Foundation’s 2024 implementation guidance offers a flexible process organised around Explore, Prepare, Deliver and Sustain. It is guidance for implementing educational approaches, rather than an experiment establishing that a particular AI product improves learning. A school can use it to organise questions about suitability, preparation, support and continuation while still examining the proposed teaching approach on its own merits. Read the EEF implementation process.

These distinctions matter in a leadership meeting. A principle can guide a decision; a study can provide evidence under stated conditions; a local trial can reveal practical difficulties; an institutional requirement can determine what processes must be followed. None should quietly become a claim that all the others have been satisfied.

For the rest of this guide, the worked trial, records and decisions are original teaching proposals. Leaders should place any such proposal within the current requirements and responsibilities of their own institution before using it.

Contents · Previous chapter · Next chapter

CHAPTER 3 OF 19 · Define a worthwhile trial

3. Define the trial before enthusiasm expands its scope

Back to contents

A bounded trial needs an audience, a learning purpose, a time window, a defined use of AI and a review decision. In the fictional case here, two teachers plan a short lesson sequence for one class of twenty-four learners. The learning purpose is to compare a claim with a short source and explain a correction. The AI component supplies candidate explanations for scrutiny; it does not decide marks or place learners into ability groups.

The teachers choose a small source pack they can check carefully. They specify when learners may consult an AI response, when they should discuss evidence with a partner, and when they will attempt an individual task. They also prepare a teacher-mediated route for learners who cannot use the planned live interface. These conditions belong in the trial description because they affect what the eventual results mean.

The review question is similarly limited: should this teaching sequence be revised and tried again with another comparable task? A department-wide rollout, a procurement commitment and a claim about examination attainment would require different evidence and decisions. Keeping those questions distinct prevents a useful small experiment from carrying conclusions it was never designed to support.

Write down an ending as well as a beginning. A trial can conclude with a useful lesson, an unresolved question, a decision to stop, or a revised teaching approach. It should not become permanent simply because nobody scheduled a review or because the original organiser moved to another responsibility.

Contents · Previous chapter · Next chapter

CHAPTER 4 OF 19 · Define a worthwhile trial

4. Map the complete teaching workflow

Back to contents

Leaders need to see the work surrounding an AI interaction. In the proposed sequence, someone selects the sources, checks the task instructions, tries candidate responses, prepares a fallback, teaches the lesson, reviews learner explanations and decides what to do next. A quick demonstration may reveal only the moment when an answer appears. The school has to organise everything that makes that answer educationally useful.

A practical planning conversation can follow one learner from the first instruction to the next lesson. What does the learner receive? What must they already understand? Who notices if the response conflicts with the source? What happens when the interface is unavailable? Where does the teacher record a misunderstanding? Which part of the next lesson changes because of that observation?

Follow the teacher’s experience as well. A generated activity may take little time to produce yet require substantial editing. An apparently helpful dashboard may create extra review tasks. Conversely, a carefully prepared common source pack may become easier to adapt over time. These are possibilities to examine in the local workflow, rather than benefits or costs to assume in advance.

The leader’s role is to make responsibilities workable. If every teacher must independently recreate the same checking process, the trial may be testing an unsustainable workload. If one specialist silently does all the difficult work, the department may misunderstand what continuation requires. Record who completes each necessary action and whether another colleague could reasonably take it over.

Contents · Previous chapter · Next chapter

CHAPTER 5 OF 19 · Prepare evidence and teaching

5. Decide what evidence will answer the question

Back to contents

The fictional teachers agree to collect several kinds of evidence because each answers a different question. A supported product shows what a learner can produce with the permitted resources. An individual check shows what the learner can explain under its specified conditions. A later check gives another observation after an interval. A task with changed conditions can examine whether the learner adapts the reasoning.

These observations should connect to the same learning purpose without becoming identical repeated questions. If learners have memorised the correction to one sentence, repeating that sentence may show recall of the example. A new short source with a different overstatement asks them to apply the underlying distinction. The teacher should still check that the new task does not introduce unrelated reading demands that overwhelm the intended comparison.

Agree what an adequate explanation looks like before reading the results. Here it requires a defensible claim, a relevant piece of source evidence, and an explanation of the relationship between them. Record support conditions alongside the judgement. A correct response after a substantial hint is useful evidence about supported performance, but the hint remains part of its interpretation.

The school can also collect brief learner comments about usability and confusion. Those comments may explain an obstacle or suggest a revision. They should sit beside evidence of learning rather than silently replace it. Enjoyment, completion and conceptual understanding are different questions, even when they all matter to the lesson.

Contents · Previous chapter · Next chapter

CHAPTER 6 OF 19 · Prepare evidence and teaching

6. Develop teachers' judgement through shared examples

Back to contents

UNESCO’s 2024 AI Competency Framework for Teachers describes fifteen competencies across five dimensions, covering a human-centred mindset, ethics, AI foundations and applications, pedagogy, and professional learning. It offers Acquire, Deepen and Create progression levels as a global reference. It is useful for thinking about the breadth of teacher development; it is not a local mandatory qualification or evidence that a particular workshop changes classroom practice. Read the UNESCO framework.

For the proposed trial, a focused staff session could begin with the actual source pack and three candidate explanations. Teachers identify which claims are supported, which need qualification, and which should be removed. They then compare the questions they would ask a learner who accepted an overstatement. This makes professional development closely connected to the judgement the lesson requires.

Include an example where the AI response is largely helpful but contains one consequential error. A teacher who rejects every generated response and a teacher who accepts fluent prose without checking both need a more discriminating method. The shared task is to locate the claim, examine its basis and decide what educational action follows.

A useful outcome of the session is a small set of agreed examples with explanations, not a large collection of prompts. Colleagues can return to those examples during moderation and adapt them when the source material changes. Leaders should allow time for teachers to disagree productively, revise an interpretation and ask for subject expertise where the answer remains uncertain.

Contents · Previous chapter · Next chapter

CHAPTER 7 OF 19 · Prepare evidence and teaching

7. Prepare materials and conditions that colleagues can understand

Back to contents

The common lesson pack should make its teaching logic visible. It can include the learning purpose, the short source, candidate explanations, likely misconceptions, the independent check and a note about permitted support. A colleague should be able to explain why each item is present. A resource that works only when its author is in the room is difficult to carry into ordinary departmental practice.

Version labels can remain simple and useful: the date, the editor, the reason for a change and the current version. Suppose a source sentence is revised to remove an ambiguity. The teacher should also inspect the candidate explanations, the expected answer and the independent check. Changing one piece may alter the meaning of the others.

Agree which elements are central and which can vary. In this trial, requiring a learner to connect a claim with evidence is central. The exact discussion grouping or choice between typing and handwriting might vary when it preserves the learning purpose. A record of sensible adaptations helps the team distinguish changes that improve access from changes that remove the reasoning being taught.

Prepare the fallback as part of the lesson, with comparable source evidence and an opportunity to make the same judgement. A printed candidate response may support the core task when a live conversation is unnecessary. If generating and refining a question is itself the learning goal, the alternative will need closer consideration because a static response may not provide equivalent practice.

Contents · Previous chapter · Next chapter

CHAPTER 8 OF 19 · Review the fictional trial

8. Examine a fictional trial without turning it into a success story

Back to contents

The following observations are invented to illustrate a leadership discussion. They do not describe a real school, product evaluation or research study. The class contains twenty-four learners, and two teachers organise the sequence. Twenty learners use the planned live route. Four use the prepared teacher-mediated fallback. All twenty-four complete the supported task and the immediate individual check.

ObservationFictional resultQuestion for review
Access to the planned live route20 of 24 learners; 4 use the fallbackDid both routes provide the intended learning opportunity?
Supported products meeting the stated criteria21 of 24 productsWhich support contributed, and what did learners understand?
Immediate individual checks meeting the criteria14 of 24 checksWhich reasoning still needs teaching?
Later checks meeting the criteria11 of the 22 learners present; 2 absentWhat does the new task show, and what evidence is missing?
Teacher review of selected AI responses3 substantive errors among 18 reviewed responsesWhat errors occurred, and how should the lesson change?
Invented observations for a small classroom trial. Counts describe different activities and conditions; they do not estimate an AI effect.

A leader could reasonably find this trial worth further investigation. Many supported products meet the criteria, and the teaching sequence has produced information about access and misunderstanding. The same record also leaves substantial questions. Several learners cannot yet explain the reasoning adequately in the immediate check, and the later observation does not resolve whether they can retain and adapt it.

Ask to see a few representative pieces of work alongside the counts: a supported success with a strong individual explanation, a supported success with a weak explanation, and an initially weak response that improves after teaching. These are examples for interpretation, selected with an explicit purpose. They should not be presented as a representative statistical sample of the whole class.

Contents · Previous chapter · Next chapter

CHAPTER 9 OF 19 · Review the fictional trial

9. Interpret denominators, task differences and missing evidence

Back to contents

The later check concerns twenty-two learners who were present. Reporting eleven successes out of twenty-four as the success rate among learners who attempted that check would incorrectly count the two absences as unsuccessful attempts. Reporting eleven out of twenty-two, alongside the two missing observations, gives the observed success rate; any whole-class summary should keep those missing outcomes explicit. The team can arrange an appropriate later opportunity for those learners without guessing their results.

The supported total of twenty-one and the immediate total of fourteen do not prove that exactly seven particular learners were hiding a lack of understanding. Aggregate counts alone do not identify the same individuals across activities. The team needs matched records, comparable criteria and a careful account of the different task conditions before describing an individual pattern.

Likewise, three errors among eighteen selected responses do not establish a general error rate for the product. Teachers may have selected unusual or difficult responses for review. The relevant first question is what those errors reveal about this lesson: perhaps the model adds an unsupported detail, misreads a qualifier, or combines evidence from different sources. Each would suggest a different correction.

This trial has no randomised comparison and cannot isolate a causal effect of AI. Teaching, task design, discussion, practice and access arrangements all contribute to the experience. The observations can support a local decision to revise and investigate. Claims about improved attainment across the school would require a stronger evaluation suited to that question. Being precise about these limits makes the record more useful for the next decision.

Contents · Previous chapter · Next chapter

CHAPTER 10 OF 19 · Review the fictional trial

10. Make a decision that follows from the evidence

Back to contents

The team should leave the review with a decision, reasons and an owner. In this fictional case, a defensible decision would be to revise the lesson and continue a bounded trial. The supported products suggest the activity is feasible for many learners, while the individual checks identify teaching that remains incomplete. The access differences and errors also require attention before any broader claim.

Possible decisionReason to consider itConcrete next action
Continue within the existing scopeThe purpose remains worthwhile and the conditions are manageableRepeat the approach with a new source and planned review
ReviseEvidence identifies a teachable misunderstanding or a workable access problemChange the relevant explanation, support or task condition
PauseRequired checking or access arrangements cannot be provided reliablyUse another teaching route while the unresolved condition is addressed
Consider a wider trialThe team has a clearer account of learning, feasibility and remaining uncertaintyDefine the new audience, support and evaluation before extending
Proposed decision options for discussion, with no automatic numerical threshold for expansion.

The team might retain the source-comparison activity, add explicit teaching about qualifiers, improve the fallback and reduce the number of candidate responses. Each revision should connect to an observed issue. Adding features because they are available can make it harder to discover what helped.

A decision to pause can preserve useful work. Teachers may keep the checked examples and continue the reasoning task without the live interface. The record should explain the unresolved condition and what would justify reconsideration. This prevents a pause from becoming an unexplained rejection of every future use of AI, while protecting the class from an approach the school cannot currently support.

Contents · Previous chapter · Next chapter

CHAPTER 11 OF 19 · Review the fictional trial

11. Count the work that makes the lesson possible

Back to contents

For leadership purposes, staff time includes preparation, checking, support and follow-up. The next table is another invented part of the same case. Minutes are summed across the two teachers; they are not the elapsed duration of the lesson or a separate amount for each teacher. This record covers only the four listed activities. Classroom teaching time and other unmeasured work are excluded, so it is not a complete workload total. The later cycle reuses material but still requires checking and review.

ActivityFirst cycle: total staff minutesLater cycle: total staff minutes
Prepare or adapt the common source pack9025
Check candidate outputs3525
Technical preparation and support2510
Review learner checks4040
Total190100
Invented staff-time record. The difference is 90 minutes; these observations do not establish that AI caused a time saving.

The later total is lower, but reuse, experience and different support needs have changed. There is no comparable non-AI sequence here from which to estimate a causal saving. A leader can still use the record to ask whether the ongoing work is manageable and which parts should be shared.

Time also has an educational purpose. Removing forty minutes of review would lower the recorded total while weakening the evidence available for the next lesson. The question is whether the school can support the necessary work at a reasonable opportunity cost. A teacher spending an evening repairing avoidable output errors has less time for other teaching preparation.

Record paid services, printing or equipment where they are relevant to the local decision, using actual costs when available. Avoid combining invented minutes with assumed prices to present a precise financial return. The first useful outcome is an honest account of what continuation requires.

Contents · Previous chapter · Next chapter

CHAPTER 12 OF 19 · Manage access and responsibility

12. Treat access as part of the design and the evidence

Back to contents

The four learners using a fallback should remain visible in the review without being marked as a problem group. The team needs to understand whether the alternative offered a meaningful opportunity to practise the intended judgement. A completion count alone cannot answer that question. Teachers can inspect the work and ask what each route made easier or harder.

If the live route supplies many conversational hints while the fallback provides one static response, the conditions differ. That difference may be appropriate for some learners, but it should be documented before comparing outcomes. Equally, identical interfaces do not guarantee equivalent opportunities when learners differ in reading access, language experience or familiarity with the task.

A useful leader question is: what must remain common for the learning claim to be meaningful, and what may vary to make participation possible? In this example, everyone needs a source, a claim to examine and a way to explain the relationship. The form of presentation or response can be adapted where it preserves that purpose.

The AI and Educational Access article develops this distinction in greater detail. For the school trial, leaders should ensure that the review does not reward a teacher for excluding learners whose support needs make implementation harder. A useful trial makes those conditions understandable and helps the school improve them.

Contents · Previous chapter · Next chapter

CHAPTER 13 OF 19 · Manage access and responsibility

13. Ask product questions that connect to the teaching task

Back to contents

A product demonstration becomes more informative when the school brings its own bounded task. Can the tool distinguish the source from an added inference? Can teachers inspect a generated explanation before learners use it? What happens when the response is wrong? Which settings change the amount of help? The answers should be demonstrated under the conditions the school expects to use.

Singapore’s September 2026 beta guidelines for EdTech product developers concern AI products seeking whitelisting for schools’ use through SLS access or integration with SLS. The document says it is not a checklist requiring every criterion to be met in full. Its specific scope should remain clear when leaders or developers discuss it; it does not by itself establish permission for any classroom deployment. Read the beta developer guidelines.

For a local decision, distinguish a promised feature from something the team has inspected. A system may advertise useful feedback, but the department still needs to examine feedback on its own sources and learner responses. A claim about reduced workload needs an account of the checking and follow-up that remain.

Ask what changes when the product, account settings or underlying service changes. Who informs the school, who rechecks the lesson and what fallback is available during uncertainty? These questions can be discussed before any purchasing decision. A small trial should not quietly create dependence on a feature that colleagues cannot maintain, explain or replace.

Contents · Previous chapter · Next chapter

CHAPTER 14 OF 19 · Manage access and responsibility

14. Keep authority and information use explicit

Back to contents

In the fictional sequence, AI offers candidate explanations. Teachers remain responsible for the instructional task and the interpretation of learner evidence within the school’s processes. That boundary should appear in ordinary language in the lesson plan: generated content is material to examine, and a model’s judgement does not automatically become a recorded judgement about a child.

The NIST AI Risk Management Framework, released in 2023, is intended for voluntary use. Its core functions are Govern, Map, Measure and Manage, applied iteratively across an AI system’s lifecycle. It offers a general way to organise responsibility and uncertainty; it is not education legislation or a certification of this proposed trial. Read the NIST overview and the framework’s core functions.

A school can translate those broad questions into a short local account: who owns the activity, what information it uses, what decisions it informs, what checks are required and who can correct an error. Follow the institution’s applicable arrangements for data, access and records. Avoid sending an entire learning history merely because a tool can accept a large prompt.

Use the smallest set of information that serves the teaching purpose. For this source-comparison task, a fictional response can often support staff discussion without identifying a learner. Where real work is needed, the school should use its authorised processes and make the purpose of the review clear to the people involved.

Contents · Previous chapter · Next chapter

CHAPTER 15 OF 19 · Manage access and responsibility

15. Make correction a visible part of ordinary practice

Back to contents

Consider an invented incident within the trial: a generated summary changes a task deadline from Thursday to Friday. The teacher discovers the discrepancy after some learners have read the summary. The immediate educational need is a clear correction that reaches affected learners through the school’s normal communication process. The teacher should explain which instruction applies and ensure that learners are not disadvantaged by the school’s error.

The team then examines how the wrong date entered the material. Did the source contain conflicting versions? Was the prompt ambiguous? Did a teacher copy a response without comparing it with the approved instructions? The purpose is to improve the process, while maintaining responsibility for the correction. A vague warning to be more careful may leave the same conditions in place.

A brief incident record can identify the affected material, the correction, the action taken and the change needed before reuse. It need not preserve every conversation or become a public account of an individual teacher’s mistake. The amount of documentation should fit the decision it supports and the institution’s requirements.

Where an error changes the validity of an assessment or another consequential decision, the relevant school process must address that consequence. The classroom example here does not establish a universal incident procedure. It illustrates why an AI activity needs a workable correction route before the first mistake occurs, including a person who has both the authority and time to act.

Contents · Previous chapter · Next chapter

CHAPTER 16 OF 19 · Sustain and extend practice

16. Explain the trial to learners and families in useful language

Back to contents

A clear explanation should describe what learners will do and why. For example: the class will examine candidate explanations against a short source, practise identifying unsupported claims, and complete an individual task so teachers can plan the next lesson. This gives families an educational account they can understand without needing a technical description of the system.

Explain the permitted support and the alternative route. If no home account or subscription is required for this trial, say so when communicating the actual arrangements. If a later proposal changes those conditions, review the implications before describing it as a continuation of the same activity. Families should be able to distinguish a school-supported lesson from an expectation that they provide additional tools or coaching.

Learners can contribute observations about unclear instructions, inaccessible presentation or advice they could not evaluate. Give them a practical way to raise those issues. Their reports can guide improvement without asking them to diagnose a technical fault or prove that an algorithm behaved unfairly.

Report findings with their limits. A parent summary might say that the class practised source-based reasoning, that individual checks revealed further teaching needs, and that the team is revising the activity. Avoid presenting the twenty-one supported successes as proof that AI has improved every learner’s understanding. Honest communication helps the school explain a revision or pause without appearing to reverse an inflated earlier promise.

Contents · Previous chapter · Next chapter

CHAPTER 17 OF 19 · Sustain and extend practice

17. Build continuity that survives a change of staff

Back to contents

Institutional learning requires something a new colleague can use. At the end of the trial, retain a concise teaching record: the purpose, current materials, permitted support, examples of learner reasoning, access arrangements, unresolved questions and the latest review decision. Include who maintains it and when it should next be examined. A large folder of unlabelled outputs is difficult to interpret.

Distinguish the shared lesson record from an individual learner’s record. The former explains the teaching approach. The latter should carry only the relevant, contextualised evidence needed for the next educational decision, through the school’s established processes. A learner who needed a hint on one task should not acquire a permanent label suggesting general dependence on AI.

A receiving teacher should be able to check the meaning of a previous judgement. What task was used? What support was present? What could the learner explain? What remains uncertain? Fresh classroom evidence can confirm or change the interpretation. Continuity should make teaching more responsive while leaving learners room to demonstrate growth.

The article on organisations that learn examines the wider problem of turning individual experience into usable shared knowledge. In a school, the same principle becomes concrete when another teacher can reproduce the intended learning conditions, understand why revisions were made and improve the approach without depending on the memory of its original author.

Contents · Previous chapter · Next chapter

CHAPTER 18 OF 19 · Sustain and extend practice

18. Review changes before extending the approach

Back to contents

Expansion changes the problem. Moving from one class to several classes introduces different learners, teachers, timetables and support conditions. A teaching sequence that worked with two closely collaborating teachers may require more preparation when colleagues have less time to meet. A wider trial should therefore state what has changed and what evidence is needed under the new conditions.

Keep a small number of review questions consistent: are learners developing the intended capability, can they demonstrate it under the relevant conditions, are access arrangements workable, and can staff sustain the necessary teaching and checking? Add questions when the new context creates a concrete reason. Avoid collecting data simply because a platform makes it available.

A review should also occur when a material condition changes, such as a new source pack, a different age group, altered product behaviour or a change in who maintains the activity. The extent of the review can match the change. Rechecking a revised instruction does not require repeating every previous discussion, while a new high-stakes use needs a separate decision.

The AI and Learning Transfer article explains why successful performance in one setting leaves questions about another. School leadership faces a related institutional problem: an approach must be understood well enough to adapt, and its success should be examined again when the conditions that supported it no longer hold.

Contents · Previous chapter · Next chapter

CHAPTER 19 OF 19 · Sustain and extend practice

19. Judge progress by the capability the school can sustain

Back to contents

The educational promise of AI becomes meaningful when learners can ask better questions, use evidence, explain their reasoning and act on corrections. Leadership contributes by creating the conditions in which those capabilities can be taught and checked. The tool is one part of that arrangement, alongside subject knowledge, teacher judgement, accessible materials and time for review.

In the fictional case, the most useful result is a better-informed next decision. The school knows that many supported products meet the criteria, that individual understanding remains uneven, that access requires deliberate design and that checking consumes real staff time. It has identified changes that can be tried and evidence that remains missing. Those findings are more actionable than a general statement that the trial was innovative.

A school develops institutional capability when colleagues can explain the purpose of an approach, recognise its limits, correct its errors and hand it on responsibly. That capability can support future trials with different tools. It also allows the school to stop using a particular feature while keeping the educational knowledge gained from the work.

The broader AI, education and civilisation argument remains conditional: stronger tools can contribute to stronger human institutions when those institutions preserve understanding, responsibility and the ability to revise. For a school leader, the next practical step is to choose one worthwhile learning problem, organise a manageable investigation and make the resulting decision clear enough for others to examine and improve.

Contents · Previous chapter · Continue the series

Sources and evidence notes

All classroom counts, staff minutes, incidents and proposed decisions in this article are fictional teaching examples. The cited frameworks provide context and organising ideas; they do not validate the invented trial or establish a causal learning benefit from AI.

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.