Education AI products have a particular compliance feature: a significant proportion of users are minors, and a significant proportion of use takes place in schools. Both facts raise the documentary standard above that of a consumer AI product. It is not a matter of adding paragraphs of explanation but of assembling material that can prove the guardian actually consented, the school actually authorised the deployment, and the data was in fact handled differently by age.
1. The age line determines the applicable rules
The personal information of minors under fourteen is sensitive personal information. Article 31 of the Personal Information Protection Law requires consent from a parent or other guardian and the adoption of a dedicated rule for processing such information. An education product therefore needs at least two documents: a privacy policy for general users and a separate rule for children’s personal information.
Users aged fourteen or over but under eighteen do not fall within the sensitive category, but the Regulation on the Protection of Minors in Cyberspace and the requirements associated with minor mode still apply. A product aimed at school pupils usually has to address both age groups, and the documents must distinguish them.
How age is determined decides whether the documents can be implemented. A line on the registration page saying that users under fourteen should use the product with a guardian is generally insufficient to demonstrate performance of the duty. Workable methods include identity verification, guardian confirmation and school-issued accounts.
2. Three ways to implement guardian consent, and the evidence
Guardian identity verification. Confirm that the person acting is the guardian by mobile number, identity document or face verification. Evidence includes the verification record, time and method.
School-level authorisation. For in-school deployment, the school procures and deploys the product as the organiser, and the authorisation relationship between school and parents is completed by the school. The provider needs the school’s authorisation document and confirmation that the school has discharged its notice duty. This is the most efficient route, but the division of responsibility must be written down: the school handles notice and consent, the provider handles processing security.
Layered consent. Basic functionality is consented to once by the guardian; functionality involving sensitive information or third-party provision requires separate consent. Layering avoids a single blanket authorisation being held invalid.
All three must be withdrawable. After withdrawal, the product should cease the relevant processing and offer a deletion path.
3. Four required documents
| Document | Audience | Core content |
|---|---|---|
| Privacy policy | General users | Notice of processing; exercise of rights |
| Children’s personal information rule | Guardians | Scope of collection, purpose, sharing, retention, withdrawal and deletion |
| User agreement | Schools, guardians | Service rules, content responsibility, termination, data ownership |
| PIA report | Regulator (on request) | Assessment conclusion for children’s sensitive information |
The relationship is that the children’s rule is a special provision of the privacy policy for the child context and prevails in case of conflict; the user agreement allocates responsibility between school and provider; the PIA supplies the facts and conclusions underpinning the other three.
4. Three gaps in the school scenario
First, an incomplete authorisation chain. Products enter schools through education bureau procurement, school procurement or individual teacher registration. The authorising party differs in each case and the documents must correspond. Individual teacher registration that processes an entire class’s data is the weakest chain — a teacher has no authority to consent on behalf of students’ guardians.
Second, excessive teacher permissions. Teachers may be able to view learning records, behavioural data and dialogue content. Permissions should match the teaching purpose; data beyond that scope, such as emotional analysis or family circumstances, should not be exposed to the teacher interface. The documents should state what each role can see.
Third, undisclosed use of data for model improvement. Education data is valuable for model improvement, but using student data for training changes the processing purpose and requires separate notice and guardian consent. A clause buried in the technical section of a privacy policy generally does not amount to effective notice.
5. Minor mode and its documentary linkage
Minor mode is not merely an interface switch. It comprises the content pool, time and duration management, anti-circumvention measures and an appeal process. The documents must explain:
- which functions are unavailable in minor mode;
- how the mode is triggered and how exit is verified;
- what a guardian can configure;
- the appeal channel and response time.
Where a product is characterised as an AI service that may affect minors’ cognition, minor mode becomes mandatory, and the documents must describe its implementation consistently with the filing materials.
6. Additional duties on content and labelling
Learning content, assessment results and recommendations generated by an education AI product are synthetic content and must be labelled accordingly. Content aimed at minors must also satisfy content-safety requirements, and the product should have a moderation mechanism described in the documents.
A further point easily overlooked is anthropomorphic interaction. Where the product presents itself as an AI teacher or companion, the rules on anthropomorphic interactive services may apply and the scope of those measures should be assessed.
7. Checklist
| No. | Item | Pass standard |
|---|---|---|
| 1 | Age identification | Verifiable method, not a mere statement |
| 2 | Children’s rule | Standalone document with scope and retention |
| 3 | Guardian consent | One of three methods implemented; withdrawable |
| 4 | School authorisation | Procuring or deploying entity issues authorisation |
| 5 | Role permissions | Visibility of teacher, school and parent defined |
| 6 | Training use | Student data used for training separately notified |
| 7 | Minor mode | Functions, schedule, anti-circumvention, appeal complete |
| 8 | Content labelling | Generated content labelled as required |
| 9 | PIA | Covers processing of children’s sensitive information |
| 10 | Consistency | Four documents factually consistent; check recorded |
The documentary work for an education product comes down to one question: when a regulator asks how the company knows that consent came from a guardian and that the data was used only where it should be, can the company produce the material? If it can, the product is genuinely fit to enter a school.
Related reading: the two requirements for AI services in the draft minors regulation, privacy policies for AI products, personal information protection impact assessment, the AI product launch compliance document pack, and the data compliance and algorithm filing service pages.
Related reading: Minors protection: which AI products move first Last updated: 9 October 2026.
Frequently Asked Questions
Where is the age line for education AI products?
Fourteen is the boundary for children's personal information and guardian consent. Processing data of users under fourteen requires guardian consent.
How can guardian consent be evidenced?
Three elements are indispensable: verifying the guardian's identity, recording when and to what consent was given, and providing a withdrawal channel. A checkbox alone is insufficient.
Which documents does an education product need?
Terms of service, a privacy policy with a children's section, a personal information protection impact assessment, and a minors mode description plus content labelling plan.
What is easily missed in school scenarios?
Allocating the role of the school as processor or entrusted party, separate notice to teachers and parents, and the boundary between school data and home devices.
How does minors mode link to the documents?
The mode's feature list must be disclosed correspondingly in the policy and terms, including identification, content pool, time management and appeals, so that practice matches the text.
What extra labelling duties apply?
Generated content needs both explicit and implicit labels, AI teachers or assistants need an identity notice, and designs that induce addiction or emotional dependence are prohibited.