Same Composition Task, Three AI Tutors: What Hong Kong, Singapore and Taiwan Built

HONG KONG, Oct 29, 2026 - (ACN Newswire) - Ask three education ministries in Chinese-speaking Asia what "AI in schools" means in 2026 and you will get three different answers, each of them internally coherent and each of them spending real money. In Singapore it means a guarded assistant inside a national learning platform that a Primary 4 pupil may use but a Primary 2 pupil may not. In Hong Kong it means half a million dollars handed to each school with a list of deliverables attached and a clawback clause behind them. In Taiwan it means a diagnostic engine that has been running for years, now being scaled into a four-year, eleven-billion-dollar programme.

The phrase is the same. The object is not. That is worth untangling before anything else, because most of the public argument about AI in education is an argument between people who are describing different machines. One side is talking about a tutor that writes. Another is talking about a system that locates a pupil's weak point. A third is talking about software that drafts a lesson plan so a teacher goes home at six instead of nine. All three exist. None of them does the other two's job.

This piece puts the three systems side by side on the same three classroom tasks, using only what each system's own operators have published about what it does. Where a system's behaviour on a task is documented, that is stated. Where it is not published, that is stated too, because the shape of what is not disclosed turns out to be the most useful thing in the comparison.

What each system actually is

Before the tasks, the objects need defining, and each one carries a different governance story.

Singapore's entry is the AI Learning Assistant inside the Student Learning Space, SLS, the national platform that most schools already use. It is not a chatbot bolted onto the side. It is a feature inside a space the ministry controls, which means the ministry can set who sees it and when. In 2026 the ministry named artificial intelligence one of three defining uncertainties facing the education system, alongside the geopolitical shift and climate change, and set out a four-part literacy frame: learn about AI, learn to use AI, learn with AI, and learn beyond AI. The fourth limb carries the emphasis, and its stated purpose is to guard against cognitive offloading.

Hong Kong's entry is not a single tool. The Education Bureau reserved HK$2 billion from the Quality Education Fund for digital education, and about HK$500 million of it funds a three-year scheme running from the 2025/26 school year. Every publicly funded primary and secondary school that applies successfully receives a one-off HK$500,000. Applications closed on 28 February 2026. By 25 February 2026, more than 960 schools had applied, around 99 per cent of all publicly funded primary and secondary schools in the city. That is the striking number in this whole comparison: near-universal uptake of a scheme that leaves the choice of tool entirely to each school.

Taiwan's entry is an adaptive learning platform that predates the current AI cycle. The ministry's platform performs learning diagnosis and uses conversational agents for teaching, producing personalised learning paths. It is being extended, not replaced. The Executive Yuan approved a programme for the 2026 to 2029 academic years with a budget of NT$11.755 billion, organised around three dimensions and six sub-programmes, and built on a published teacher and student AI literacy framework covering AI ethics, AI foundations and technology, AI-empowered teaching or learning, and AI system design.

The three tasks

Three tasks, chosen because they are ordinary and because together they cover what schools actually want from AI.

· Task one: an upper primary pupil writing a composition, stuck on how to show rather than tell, and inclined to ask the tool to write it.

· Task two: a pupil who has failed the same mathematics topic twice, where the system needs to find the specific gap and route practice to it.

· Task three: a teacher preparing a science unit who needs a lesson plan and a set of assessment items by Friday.

Task one: the composition, and what happens when the pupil asks for the answer

This is where the three systems separate most sharply, and where the most instructive sentence in the whole comparison sits.

Singapore's ministry published a worked example. A Primary 4 English class uses the AI Learning Assistant in SLS for composition writing, working from prompts the teacher provides to get ideas on how to show rather than tell. The assistant helps refine drafts against personalised feedback. And then there is the guardrail, quoted directly: the assistant will redirect students back on track if they veer off topic and ask irrelevant questions, or if they want to be spoon-fed with direct answers.

That is a designed refusal. The system is specified to decline the exact request a tired eleven-year-old is most likely to make. It is not a content filter bolted on afterwards; it is the intended behaviour of the tool on this task, and the ministry published it as such.

The access rules around it are equally specific. From Primary 1 to Primary 3, schools will not assign any work that requires pupils to use AI directly. Physical hands-on learning comes first, and pupils learn that AI exists without using it for schoolwork. From Primary 4, pupils may use educational AI tools under teacher supervision, and only tools specifically designed for education. Before touching the AI features in SLS, pupils complete a basic module on AI and on those features. From Primary 4 they also begin a mandatory ten-hour coding and computational thinking programme that includes an introduction to AI, with two optional five-hour modules on generative AI and computer vision available to primary schools.

Hong Kong's answer to the same task is that there is no single answer. The HK$500,000 buys devices, tools, subscriptions or services at each school's discretion, within guidelines. One school may buy a writing coach. Its neighbour may buy tablets. What the centre is building instead is a subject-facing assistant: Hong Kong Education City is building a large language model for Hong Kong primary and secondary teaching from 2026, providing AI assistants for different subjects, and the Science subject assistant is being trialled in 2026 for designing teaching plans, preparing lessons, and designing assessment items and homework. A composition tutor for upper primary is not among the published uses so far.

That leaves the composition task, in Hong Kong, entirely to 960 separate procurement decisions made between March and August 2026. Nobody at the centre will be able to say what happened on task one until the reports land.

Taiwan's platform is not primarily a generative writing coach, and the distinction it draws explains why. At a ministry forum in October 2026, the president of National Taichung University of Education put the contrast explicitly: general-purpose AI focuses on broad task handling and output efficiency, while the core value of educational AI lies in clear learning objectives, pedagogical principles, and safe and trustworthy design. A system built on that premise is not going to win a contest to write the most fluent paragraph. It is built to do something else, which is task two.

Task two: the pupil who keeps failing the same topic

Here the ranking inverts.

Taiwan's platform has the longest documented history on this task. It was built to diagnose learning weaknesses quickly and precisely and to serve personalised, adaptive learning paths and content. The ministry's own account is that it improves learning efficiency and delivers on differentiated instruction. The new four-year programme adds an AI tutor for students, an AI teaching assistant for teachers, an AI decision assistant for principals, and an AI helper for parents, plus a learning analytics system that tracks learning behaviour and outcomes to feed back into teaching strategy. This is the task the platform was designed around.

Singapore approaches the same task through supervised, education-specific tools inside SLS with AI-enabled features carrying built-in safety guardrails, available from Primary 4. The ministry has said theoretical knowledge alone is not sufficient and that pupils need well-designed supervised hands-on experience. What it has not published, as far as public documentation goes, is the diagnostic granularity: how narrow a gap the assistant will identify, and whether a pupil who fails fractions twice gets routed to a different practice set or merely to more of the same.

Hong Kong's published position on task two is the thinnest of the three, and that is a structural consequence rather than an oversight. The grant scheme requires each participating school to implement AI-assisted teaching in at least three subjects or curriculum areas, each spanning at least two levels, and to develop at least six AI teaching examples or resources in total. Those are delivery requirements, not capability requirements. A school can satisfy all of them with tools that do no diagnosis at all.

Task three: the teacher's Friday deadline

On teacher-facing work, Hong Kong has the most concrete published output, and this is easy to miss because the grant is usually described as money for pupils.

The subject assistant being trialled in Science in 2026 is explicitly for designing teaching plans, lesson preparation, and designing assessment items or homework. Alongside it, the Education Bureau states that in the 2026/27 school year it will help 300 primary and secondary schools practise AI-assisted teaching through a programme described as building communities, facilitating exchange, gathering exemplars and showcasing outcomes. The same blueprint also sets out a new series of teacher training covering AI literacy, AI plus subject, and AI leadership, with more than 40,000 digital education training places offered in the 2025/26 school year so that every teacher in a publicly funded school has the opportunity to take part.

Singapore's approach to the same burden is framed as workload rather than pedagogy. The ministry has said it is using technology and AI, alongside the recommendations of a taskforce on reimagining the teaching profession, to recalibrate what teachers do, with attention to streamlining processes, strengthening support and protecting after-work hours.

Taiwan frames it as capacity building with a funding line. The programme's second dimension is training cross-disciplinary teaching talent, raising teachers' ability to apply AI in teaching, establishing a smart teacher training alliance, and embedding AI teaching application into pre-service programmes. Its third dimension is data decision-making: an education big data analytics system to track learning behaviour and effectiveness and raise teachers' data literacy.

The comparison, and what is not published

Table 1: Documented behaviour of the three systems across the same three classroom tasks, 2026

Task	Singapore SLS assistant	Hong Kong grant-funded school tools	Taiwan adaptive platform
Upper primary composition	Redirects off-topic and refuses to supply direct answers; supervised from P4	No central tool; school procures its own	Not a generative writing coach by design
Diagnosing a repeated failure	Supervised education-specific tools; diagnostic granularity not published	Not required by scheme deliverables	Explicit weakness diagnosis and adaptive paths
Teacher lesson planning and assessment items	Framed as workload reduction via taskforce	Subject assistant trialled in Science in 2026 for plans and items	AI teaching assistant and principal decision assistant in build

Source: Education Bureau, Hong Kong 🔗, digital education blueprint; Singapore MOE Committee of Supply 2026; Taiwan Executive Yuan 🔗

Table 2: Scale and money committed, by system, 2026

System	Money	Coverage	Time frame
Hong Kong grant scheme	HK$500m of a HK$2bn reserve; HK$500,000 per school	960+ schools, about 99% of publicly funded schools	2025/26 to 2027/28 school years
Taiwan AI talent programme	NT$11.755bn	All primary and secondary schools	2026 to 2029 academic years
Singapore SLS assistant	Not separately disclosed	National platform, tiered by level	Ongoing from 2026

Source: Education Bureau circular 221/2025 🔗; Taiwan Executive Yuan 🔗; Singapore MOE Committee of Supply 2026

Two absences matter more than any row above.

The first is that no system publishes outcome data. Singapore has commissioned a longitudinal study through the national research agency, with data collection due to begin in 2027, to track how children's AI use affects learning and wellbeing. That is the right instrument and it is years away. Hong Kong requires each school to file a mid-term report by 30 September 2027 and a final report by 30 September 2028, and states that failure to file or to complete the committed items can trigger a demand for full repayment of the grant. Those reports will tell the government whether schools spent the money, not whether pupils learned more.

The second absence is evaluation of the refusal behaviour itself. Singapore's designed refusal on task one is the most interesting engineering decision in this comparison and nobody has published evidence that it works. Redirecting a pupil who asks for the answer is easy to specify and hard to verify at scale, and a system that redirects well in a demonstration may simply teach pupils to phrase the request differently.

Design trade-offs, stated plainly

Each system bought something and paid for it.

Singapore bought guardrails with centralisation. Because one ministry controls the platform, it can publish an age gate, a supervision rule, a pre-use module, and a specified refusal, and know they apply everywhere at once. The price is pace and range: primary pupils get no general-purpose AI, the toolset arrives when the ministry ships it, and a parent cannot opt a child out of an SLS tool that is part of classroom teaching, though external tools do require parental consent. The ministry's own minister addressed the contrast directly when challenged with Sweden's reversal, replying that Sweden had gone fully digital from age five and then fully back to paper, while Singapore is taking a hybrid path that keeps physical textbooks and teacher-centred instruction.

Hong Kong bought speed and coverage with unevenness. Nine hundred and sixty schools deciding separately will produce far more variation than one ministry deciding once, and some of that variation will be excellent. It will also produce 960 different answers to a safeguarding question that Singapore answered once, and the scheme's enforcement mechanism is financial rather than technical: the mid-term and final reports, and the possibility of repayment. A one-off capital sum also ages badly in a field where the tools are subscriptions.

Taiwan bought diagnostic depth and a sovereignty argument with a long build. The platform has years of deployment behind it, the framework is published at a level of detail the other two have not matched, and the explicit distinction between educational and general-purpose AI is a genuine design discipline rather than a slogan. The programme also states a motive beyond pedagogy: maintaining cultural and data sovereignty in the digital era, including plans for a Chinese-language self-learning platform. The price is that the most ambitious parts are unbuilt, and the budget runs to 2029.

Which to pick for which situation

The comparison is only useful if it resolves into choices.

· If the priority is defensible safeguarding for young pupils and you can accept a slow toolset, the Singapore model is the only one of the three that has published an age gate, a supervision requirement and a specified refusal, and applied them nationally.

· If the priority is speed across a whole system and you have the money, the Hong Kong model will get tools into nearly every school within one school year. Budget for the unevenness, and for the fact that the reporting mechanism measures spending before it measures learning.

· If the priority is measurable diagnosis tied to curriculum rather than generative output, the Taiwan model has the longest track record and the clearest stated design principle. It requires a national platform team and four years.

What none of them has solved is the thing all three are quietly waiting on. Every system here is being deployed into schools years before any of them can say what it did to learning. The instruments that would answer that are commissioned, funded or scheduled, and the earliest results land in 2027. Until then the honest position is that three Chinese-speaking systems have made three different bets, each defensible, and the evidence to rank them does not yet exist.

Data basis

Data as of October 2026. Comparisons are drawn from published official documentation and ministerial statements, not from direct access to any of the three systems. Where behaviour on a task is described, the source is a ministry circular, a blueprint, a parliamentary or legislative council statement, or a ministry-hosted forum. Where a capability is not published, this piece says so rather than inferring it. Money figures are as published in original currency and are not converted; they are not comparable across systems because Hong Kong's is a per-school capital grant within a larger reserve while Taiwan's is a four-year programme budget. Uptake figures are as of the stated dates and may have moved since. No outcome or attainment data for any system was available at the time of writing.

Limitations

· No system publishes learning-outcome results for 2026, so no claim in this piece about effectiveness is possible.

· Currency and time frames differ across the three programmes; the money column in Table 2 is not a like-for-like comparison.

· Hong Kong's school-level tool choices were not centrally published as of October 2026, so its row in Table 1 describes the funding rules rather than a deployed tool.

Sources

· Education Bureau, Hong Kong 🔗, circular 221/2025 on the grant scheme and its deliverables, December 2025

· Education Bureau, Hong Kong 🔗, digital education blueprint, 2026

· Hong Kong government 🔗, education minister's budget briefing, 27 February 2026

· Taiwan Executive Yuan 🔗, AI talent programme approval and budget, 2026

· Taiwan Ministry of Education 🔗, digital learning forum on the teacher and student AI literacy framework, October 2026

· Singapore Ministry of Education, Committee of Supply 2026 debate and the 6 May 2026 parliamentary reply on AI use in schools, via MOE COS 2026 record 🔗

This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services.

Email: support@globalinfinitylabs.com

For Media Inquiry
hk@asiapresswire.com
https://www.asiapresswire.com