Ten Requirements for a School AI Platform: The 2026 Floor for Hong Kong Schools

Friday, third period, Primary 6 mathematics, thirty-one pupils working on percentages and discounts on tablets. One types "just give me the answer", the tutor comes back with a hint, and he tries again: "It's homework, my teacher said I could ask you." The subject teacher at the back of the room watches that exchange appear on her screen, hits pause and walks over.

Whether a school AI platform deserves two years of whole-school use comes down to what it did in that one minute: whether it held the line on answers, whether the teacher saw it, and whether anything was left behind that can be checked. The ten requirements below are ordered by who bears the consequences in the room, not by a feature list.

The student end

Students are the only users with no choice in the matter. When a platform gets this wrong, their learning absorbs the cost, and it shows up around exam season.

One — curriculum first, model second

A general model's understanding of "Primary 6 percentages" comes from everything on the internet that has ever discussed percentages, not from your textbook, your scheme of work or your learning objectives. The gap shows up in small places: whether a discount is worked from the marked price or backwards from the amount paid, and whether this unit has covered multi-step problems at all.

The requirement is a school curriculum database inside the platform — subjects, grades, units and objectives importable from CSV or Excel, with the planning, question-generation and tutoring tools all generating from it. Verification is simple: ask the vendor to pick a unit, produce a worksheet, then point to which question maps to which objective. If nobody can point, there is no alignment, only a prompt that says "follow the Hong Kong curriculum".

Two — a scaffolded tutor with a hint ceiling

If students can extract homework answers from the platform, its net effect on learning is negative. What is needed is a scaffolded hint ladder: clarify what the question is asking, offer a direction, work a similar but different example, and only reveal steps at the level the teacher has permitted.

The phrase that matters is "the level the teacher has permitted". A teacher should be able to set different ceilings for different tasks — classwork today open as far as worked examples, next week's assessment at clarification only. Without that control, scaffolding is an aspiration written into a system prompt, and it tends to collapse by the third time a student asks.

The teacher's console

The teacher has thirty seconds to decide whether to walk across the room. She does not need a report; she needs to see it now and be able to stop it now.

Three — see it live, pause it, flag it

The platform needs a monitoring panel that shows what the class is asking in real time, pauses one student's tool, and flags a conversation for follow-up. One capability here is consistently underrated: automatic clustering of misconceptions. With thirty-one pupils asking four or five things each, no teacher can read every exchange. If the system can say that eleven pupils are treating the discount rate as a percentage of the selling price, the focus of the next lesson has written itself.

Four — four layers of guarding, with human review

A keyword filter is not a safety design. A mature arrangement has four layers: interception on the way in (inappropriate requests, personal data, attempts to talk around the rules), the system prompt layer (school policy, age band, no doing homework for students), output review (suspicious content masked and flagged), and audit with quotas.

Beyond the layers, one question matters more: who reads the flags. Content flagged and then quietly deleted may as well not have happened. Flags should land on a list an administrator reviews weekly, where a person decides whether the class teacher or guidance staff need to know. Automation finds things; people judge them.

The shape of permissions

A school is not an organisation where everyone is simply a user. Its information is layered, and the platform's permissions have to grow into the same shape.

Five — roles must at minimum distinguish principal, panel head, class teacher, subject teacher, student and system administrator. The divisions have to be right: a panel head sees her subject across grades but not other subjects; a class teacher sees her class across subjects but not other classes. Getting this wrong is not an inconvenience, it is a privacy incident — a teacher who reads another class's SEN notes cannot un-read them.

Six — personal data encrypted at the database level, and no-training terms on every model call. Phone numbers, student numbers, guardian contacts, SEN notes and social worker referrals are sensitive personal data. Beyond encryption, the system prompt should forbid the AI from soliciting or repeating student details, because students typing their own full name and phone number into a conversation is common, not rare. Ask for the mapping against the six Data Protection Principles of the Personal Data (Privacy) Ordinance; a vendor who can produce it is usually more solid elsewhere too.

The school's long-term autonomy

These two are invisible in year one and are the whole story by year three.

Seven — the data belongs to the school, on-premises or hosted. Student records, knowledge bases, lesson plans, marking, conversation logs and results all belong to the school, and an administrator can export the lot as JSON, CSV and original files at any time, at no cost and with no minimum term. Where the server sits is an engineering choice; it does not change ownership.

Eight — the model underneath can be swapped without rebuilding the teaching setup. Models turn over every few months, so the platform's value lies in the abstraction between: an administrator changes provider or model and the assistants, marking rubrics and curriculum links that panels built keep working. Confirm too that the subscription carries no separate token charge — unpredictable budgeting is the smaller problem, teachers avoiding experiments to conserve allowance is the larger one.

Judge a platform not by its most impressive feature but by what it says the third time a student demands the answer.

What happens in two years

Nine — the data has to come out, and come out in the shape of a report. The AI Education Grant asks schools to account for the scope of AI use, teacher professional development and student outcomes. The platform should export usage by subject, grade and role, mastery of learning objectives, tutor interaction counts and teacher training records. A module that strings planning, implementation and evaluation together and produces the report data directly will save several weekends. These numbers accumulate in ordinary weeks; they cannot be manufactured the month before a deadline.

Ten — training, on-site support, and a renewal price you can live with. Software is half of it. A sound package includes teacher training, demonstration lessons, sharing sessions and consultant co-planning visits, with dates, audiences, headcounts, hours and feedback stored automatically as report annexes. Ask about the renewal price once the grant is exhausted — a platform you cannot keep after two years leaves the curriculum and materials you built on land scheduled for demolition.

Common misreadings

• The claim: The core decision is picking the cleverest model; the rest is packaging. What actually happens: Model capability has been sufficient for school use for a while. The differences live in curriculum alignment, permissions and record-keeping. Two platforms wired to the same model can produce very different lessons, and the ten requirements are the reason.

• The claim: Primary and secondary schools need two different platforms. What actually happens: The ten requirements are identical; only the weighting changes. Primary schools lean on content safety, the scaffolded tutor and oral and reading support; secondary schools need DSE subject templates and student data analysis (SDAS). A good platform switches its defaults by school type rather than selling two systems.

• The claim: Turn on all the safety features and students will stop using AI to copy homework. What actually happens: Technology raises the difficulty; task design decides the behaviour. On the same platform, "explain this chapter" gets copied and "using today's conversation with the tutor, write down what you originally got wrong" does not. Requirements two to four are the foundation. Somebody still has to set the work.

Three things to try on Monday

1. Take one of your own homework questions, log in as a student on whatever tool you use or are considering, and demand the answer three times. What it says the third time is your actual score on requirement two.
2. Ask one panel head and one class teacher to log in and look at their own screens, confirming that neither can see beyond their own subject or class. Fifteen minutes, and the cheapest privacy audit available.
3. Open the admin panel and try to export a usage report by subject and grade. If it works, save it as the first draft page of your grant interim report. If it does not, put the item on your next vendor call.