AI and Digital Learning Approval Gate

Before a school adopts an AI tutor, learning assistant, automated feedback system or digital education platform, it must decide whether the tool improves learning without surrendering student agency, safety, privacy or the ability to change course. An impressive demonstration is not enough. Educators need a practical approval method that checks the educational purpose, the data route, how outputs are verified, the human reviewer, the learner’s independent capability and what happens when the provider or platform fails.

Ministry of Education V3.0 approaches AI adoption as a sequence of gates: diagnose the real learning need, confirm authority and age suitability, evaluate the product, test its claims, protect information, preserve human judgement, pilot with realistic learners, compare non-AI alternatives and maintain an exit plan. This reference is an operational manual for teachers, school leaders, curriculum teams, technology procurement officers and adults deciding which learning tools deserve trust.

This is an independent eduKateSG educational design framework, not an official Singapore MOE procurement standard or compliance certification. Laws and institutional rules governing child safety, privacy, intellectual property, procurement, cybersecurity and assessment must be applied by the people responsible for those duties. A product does not become approved merely because it is described favourably here.

1. First Ask What the Learner Must Be Able to Do

The starting question is not ‘Which AI model is most advanced?’ It is ‘Which capability is missing, and what will the student be able to do independently after support?’ If a learner cannot translate a word problem into an algebraic equation, a useful tool should help diagnose the barrier, present an appropriate representation and prompt the learner to attempt a similar new problem. A tool that instantly supplies polished solutions can look efficient while concealing the original gap.

Decide in advance how success will be measured without the tool. An intervention whose only observed gain is faster output should not automatically be described as an improvement in understanding. This principle applies equally to automated essay feedback, science explanations, language conversations and adaptive question banks.

2. Classify the Intended Use Before Approving the Product

Use case	Potential educational purpose	Distinct risk	Essential control
Teacher planning	Draft examples, questions and alternative explanations	Errors, misalignment and unsuitable content	Teacher verifies against current syllabus and learners
Student practice	Hints, feedback and graduated tasks	Answer dependence and inaccurate explanations	Independent follow-up tasks and age safeguards
Translation and accessibility	Alternative representations or language support	Loss of meaning or confidential disclosures	Qualified check where material; approved data route
Assessment support	Analyse anonymised class misconceptions	Opaque scoring or unfair conclusions	Human marking authority and validity review
Administrative workflow	Reduce repetitive tasks and improve communication	Unauthorised data access or erroneous actions	Permission control, audit trail and correction
High-impact decisions	Recommendations about placement, support or discipline	Bias, mistaken inference and serious rights consequences	Formal authority and independent human review; restrict automation

Do not treat all these uses as one risk category simply because they rely on machine learning. The stakes are different when a tool proposes a worksheet versus when it influences an entitlement, student welfare decision or educational pathway.

3. The Twelve-Gate Adoption and Approval Protocol

Gate 1 — Document the problem and a non-AI alternative

State the teaching or administrative problem in plain language. Compare the AI route with a realistic alternative such as a structured teacher-designed worksheet, an existing accessible platform or a smaller change in feedback practice. Adoption is not justified by novelty alone.

Gate 2 — Establish the lawful and institutional decision-maker

Identify who can approve classroom use, purchases, digital services, data processing and assessment changes. Verify institutional policies and relevant contracts before student interaction begins. A teacher’s enthusiasm cannot substitute for approvals required by the organisation.

Gate 3 — Confirm age suitability, supervision and access

Examine provider terms, permitted ages, required supervision, accessibility support, languages and available non-digital routes. A learner should not be pushed into independent use contrary to age restrictions or safeguarding policies. Adults must consider developmental readiness, not only technical ability.

Gate 4 — Trace every category of information

Map inputs, prompts, uploaded documents, generated outputs, logs, analytics, exports and third-party integrations. Identify where information travels, why it is necessary, who can access it and whether it may be used for model improvement. Do not assume that removing a child’s name makes all remaining educational records anonymous.

Gate 5 — Review privacy, security and retention

Apply the rules relevant to the institution. Require appropriate access management, deletion provisions, security measures and incident handling. For Singapore organisations within its scope, the Personal Data Protection Commission’s education-sector guidance is a relevant official reference. Public agencies and other bodies may have different governing requirements; the guidance must not be presented as universally applicable.

Gate 6 — Check content rights and provenance

Determine whether materials, student submissions or third-party resources may lawfully be supplied to the service. A resource that can be viewed online is not automatically licensed for ingestion into a commercial AI system. Check terms governing copying, redistribution, attribution, generated outputs and retention. Keep a source register for materials that teachers reuse.

Gate 7 — Test truthfulness against subject knowledge

Use representative examples, including misleading prompts, difficult vocabulary, numerical tasks and errors the learner is likely to make. Verify answers against trusted curriculum or disciplinary sources. Generated references may be inaccurate or fabricated. Do not accept a confident explanation as evidence of correctness.

Gate 8 — Evaluate teaching behaviour and learner agency

Inspect whether the tool encourages attempts, asks useful questions, admits uncertainty and supports conceptual understanding. Watch for excessive praise, overdependence, shortcut answers or instructions that bypass the learning objective. Students should remain able to question and correct the system rather than treating it as an unquestionable authority.

Gate 9 — Stress-test fairness and accessibility

Try the tool under ordinary rather than demonstration conditions: different language ability, unreliable connectivity, assistive technologies, older devices and varying prerequisite knowledge. Document who receives weaker support. Do not infer personal traits or protected characteristics from behaviour without a legitimate and authorised purpose.

Gate 10 — Run a bounded pilot with human oversight

Establish a baseline, comparison where feasible, a predefined learning task and an incident route. Preserve accessible alternatives and responsible adult supervision. A pilot should measure learner benefit, teacher workload, cost, exclusions and failure handling. Do not upload identifiable student records merely to evaluate the tool.

Gate 11 — Verify independent transfer

After a suitable period, ask students to complete an unfamiliar task without AI assistance. Can they explain, check and recover from an error? If output quality improved only while the tool remained open, the pilot may have demonstrated productivity rather than durable learning. Record the distinction.

Gate 12 — Approve conditionally, monitor and preserve exit

An approved tool should have a named owner, permitted purposes, user access rules, support process, review date, incident escalation and cancellation route. Vendor model updates or policy changes should trigger renewed verification. If the tool fails a critical safeguard, the institution must be able to suspend it and maintain essential learning.

4. The Educational AI Risk Register

Risk family	Warning sign	Protective response
Accuracy and hallucination	Unverifiable claims, wrong calculations or invented citations	Require source checking and teacher escalation
Student dependence	Learners complete tasks but cannot work independently	Fade prompts and test transfer
Age and safety	Inappropriate unmoderated interactions or content	Restrict access and use established safeguarding processes
Privacy	Unnecessary identifiable data supplied to external services	Minimise data and obtain proper approval
Fairness	Different groups receive lower-quality or inaccessible feedback	Use subgroup and accessibility testing
Vendor dependency	Learning records cannot be exported or tool cannot be replaced	Secure usable export and offline alternatives
Hidden workload	Teachers spend more time repairing outputs than before	Measure full operating burden
Rights and licensing	Third-party content processed beyond permitted terms	Respect licence conditions; use approved sources
Governance opacity	No one can explain a high-impact automated recommendation	Retain authorised accountable human decisions

A risk register is not a list designed to make the product fail. It is a map of conditions that responsible implementation must satisfy. Each material risk needs an owner, plausible consequence, control, evidence of the control working, and a trigger for escalation.

5. A Procurement Checklist Beyond the Demo

What specific learning problem is the product designed to address?
Which ages, languages, devices and assistive technologies are supported?
What claims are supported by independent evidence rather than testimonials?
Which data categories are collected and where are they transferred?
Which third-party services or model providers are involved?
Can records be exported in an intelligible, usable form?
What happens when the vendor changes its model, pricing, policies or product?
Can the organisation set permissions, review usage and delete information appropriately?
What training, support and remediation are included?
Which obligations survive contract termination and how will the exit be tested?

Technical performance reports alone cannot establish classroom value. Procurement should involve learning specialists alongside privacy, security, accessibility and commercial reviewers. A vendor’s answer to a contractual question is a representation to be checked and recorded, not a substitute for independent assurance where stakes demand it.

6. Worked Example: A Generative AI Mathematics Helper

Illustrative scenario. A school proposes a generative AI mathematics helper for lower-secondary students. The demonstration solves algebra questions quickly and offers several methods. School leaders initially describe it as personalised tuition for everyone.

The diagnostic gate shows that the actual need is narrower: some students cannot connect fraction operations with algebraic expressions. A better pilot begins with short teacher-selected questions that identify the missing link. The tool is configured or guided to provide hints and explanations rather than merely final answers, with teacher review of the prompts and outputs.

Students first attempt a task independently, receive appropriate support, explain the method and then complete a new problem without the tool. Teachers compare the result with the baseline and record errors, dependence and access issues. If the AI’s explanations are occasionally wrong or unsuitable, the school documents how those errors are detected and corrected. A small-group teacher-led support programme may also be used as a comparator rather than assuming the software must be superior.

The related Secondary 1 Mathematics small-group tutoring reference illustrates the importance of precise starting-position diagnosis. It is not evidence endorsing a particular AI vendor.

7. Human Control Cannot Be Merely Ceremonial

A policy may claim ‘human in the loop’ while the human reviewer has insufficient time, information or authority to challenge the output. Genuine oversight means the educator or authorised professional understands the tool’s role, can inspect relevant evidence, may override its recommendation, has a non-AI pathway and knows how to escalate an error.

Where student placement, protection, discipline, evaluation or entitlements are at stake, an opaque suggestion must not quietly become the deciding authority. The institution should record which parts of the decision were assisted by software and who remained accountable.

8. Teaching Students to Be Competent AI Users

AI literacy is not memorising a list of prompts. It includes understanding that generated text may be incorrect, identifying when a request contains private information, comparing an explanation with reliable material, recognizing the cost of delegating thinking, and knowing when to seek a human expert. Younger learners need simpler supervised tasks; older learners can discuss limitations, bias, provenance and system design.

UNESCO’s AI Competency Framework for Students describes human-centred mindset, AI ethics, AI techniques and applications, and system design. The AI Competency Framework for Teachers sets out corresponding professional-development dimensions. These are external frameworks for reference, not endorsements of the independent V3.0 method.

9. The Approval Decision Matrix

Finding	Institutional response	Required follow-up
Clear benefit and safeguards verified	Conditional deployment within defined scope	Scheduled learning and risk review
Promising tool, limited evidence	Bounded pilot with appropriate approval	Independent tasks and comparison evidence
Unresolved access or reliability problem	Adapt design or restrict use	Retest under realistic conditions
Unclear data processing or age protection	Do not deploy to affected learners	Obtain proper assurance or change product
High-impact automated outcome without accountable review	Do not use for that decision	Establish lawful human governance
Material vendor or service failure	Pause and activate exit route	Restore learning, investigate and document repair

Decisions should be proportionate to risk and evidence. A programme need not be rejected because AI is imperfect; ordinary textbooks and human explanations can also contain errors. But the presence of a new technical system can make errors scale rapidly or remain less visible, so effective controls are necessary.

10. The Exit Drill

Before scaling, simulate an interruption: the vendor is unavailable for a week, the price changes, a serious error is detected or data access must be revoked. Can teachers continue essential activities? Can students retrieve necessary materials? Can records be transferred or deleted under the applicable rules? Who communicates with families and who pays for recovery?

An exit plan written in a contract but never tested is a hypothesis. Run a small recovery exercise under supervision, document the time and missing resources, and revise the plan before dependence grows.

11. A Termly AI Governance Review

Are teachers and students using only tools authorised for their intended purposes?
Did any new model, policy or vendor change affect safety or reliability?
Did the educational gain appear in independent transfer checks?
Which learners were unable to access the system or received inferior support?
How many reported output errors were investigated and corrected?
Has the tool reduced or increased the full burden on educators and families?
Are content rights, data deletion and access permissions still valid?
Can an approved alternative operate if the service is withdrawn?

12. What Gold-Standard AI Education Looks Like

The gold standard is not a school with the largest number of AI subscriptions. It is a school where digital support is selected for a precise educational reason, learners become more capable rather than dependent, educators retain real judgement, rights are protected, claims are tested and the system remains recoverable when technology changes.

Education should produce people able to understand and improve the world, not simply generate work faster. AI earns its place in the classroom when it serves that purpose under accountable human control.

Ministry of Education V3.0 Reference 12 | The AI and Digital Learning Approval Gate: Safety, Agency and Exit