We write to the people who make the decisions

Written with the 2025 Policy Address commitments on digital education and the Education Bureau’s AI for Empowering Learning and Teaching Funding Programme in mind, ahead of the 2026–27 Budget.

The issue

The money is moving faster than the method. The 2025 Policy Address committed to an AI literacy learning framework, AI in the core curriculum and more AI training for teachers, backed by HK$2 billion from the Quality Education Fund. On 16 December 2025, the Education Bureau launched the AI for Empowering Learning and Teaching Funding Programme, giving each public sector and Direct Subsidy Scheme school HK$500,000 to use between 2025/26 and 2027/28. That’s a serious commitment. But nothing yet tells a primary teacher how a seven-year-old should use AI, or when they shouldn’t.

How AI is used matters more than whether it’s used. The strongest evidence so far is a randomised trial of nearly a thousand high-school maths students, published in PNAS in June 2025 (Bastani et al.). Students given a standard ChatGPT-style assistant scored 48% higher during practice, but when the AI was taken away for the exam they scored 17% lower than students who never had it. A second group used an AI tutor that gave hints instead of answers, and that design largely removed the harm. If teenagers can be set back like this, six- to twelve-year-olds, who are still building the basic skills AI most easily replaces, are at greater risk. UNESCO’s 2023 guidance recommends 13 as the minimum age for using generative AI independently.

Two gaps fall on the same children.

• Access. Legislative Council research found that in 2019, 94% of Hong Kong households had internet access, but only 71% of households earning under HK$10,000 a month did, and up to 13% of primary pupils had no computer or tablet of their own.
• Guidance. A child from a better-off family usually has a parent or tutor who can show them how to use AI well: ask for a hint, check the answer, explain it back. A grassroots child is more likely to have no AI at all, or AI with nobody guiding them.

A flat HK$500,000 per school closes neither gap. A school where most families pass the means test gets exactly the same as a school where few do.

Where this comes from. This proposal grew out of our own tutoring. The children we teach have never been short of interest. What they lack is access, and someone to show them how to use these tools in a way that builds their thinking instead of replacing it.

What we propose

1. An age-staged “Think First” ladder in the AI literacy framework

Stage	Principle	In class
P1–P3	Think without AI	No individual pupil accounts on generative AI. Teachers can show AI to the whole class, always framed as “Is the computer right?” Lessons build the habits AI can’t supply: explaining how you know, estimating before calculating, finding the mistake in a worked example.
P4–P6	Think, then check with AI	Supervised use only, in “hint mode”, following the four steps below.
S1–S3	Think with AI	Guided independent use, with assessment that rewards the student’s own reasoning, not just the final answer.

2. A four-step protocol for every AI-assisted primary lesson

1. Attempt. The child works on the problem alone first, for at least the first third of the time.
2. Ask. The child asks the AI for a hint, a question or a check, never for the answer.
3. Check. The child compares the AI’s reply with their own attempt and finds at least one difference, or explains why the two agree.
4. Explain. With the AI off the screen, the child explains the answer in their own words.

The last step is the safeguard. If a child can explain the answer without the AI, they learned it. If they can’t, the AI did the thinking.

3. “Hint mode” as a condition on public money. Any AI tool bought for primary pupils with public funds, including the HK$500,000 grant, should default to hints and guiding questions rather than full answers, let teachers see pupils’ prompts and the AI’s replies, and never use children’s data to train AI models.

4. Think First Labs: a three-year pilot in 100 primary schools. The 100 public primary schools with the highest share of pupils on full-grant School Textbook Assistance, a means-tested measure the government already collects, would run two 90-minute after-school sessions a week for P4–P6 pupils. One teacher, supported by trained secondary-school and university tutors at about one tutor to six pupils, would use school-owned devices on site, so no child needs a device or broadband at home to take part.

5. Teacher training. A six-hour “Teaching thinking with AI” module through the Education Bureau’s existing professional development programme, with at least two teachers trained in every public primary school by the end of the 2026/27 school year.

Cost

Per school, per year	HK$
Teacher and tutor stipends	130,000
Devices, connectivity and materials	35,000
Independent evaluation	15,000
Total	180,000

100 schools × HK$180,000 × 3 years = HK$54 million, about 2.7% of the HK$2 billion already set aside in the Quality Education Fund for digital education. The training module and the hint-mode condition fit within existing budgets.

The case for

• It protects learning. The best evidence we have shows that AI built to give hints instead of answers avoids the learning loss that unrestricted AI causes. The protocol builds that into how lessons run, not just into the software.
• It reaches the children who stand to gain most. Better-off children already get guided AI use at home. The Labs give grassroots children the same, at a fraction of the cost of private tutoring.
• It uses data the government already has. School Textbook Assistance records identify the schools in most need without a new means test.
• It’s cheap and testable. At under 3% of the existing allocation, it can be evaluated properly before anyone scales it up.

The case against

• It holds younger children back from AI. Keeping individual accounts out of P1–P3 means younger children miss some benefits, and some parents and schools will feel Hong Kong is falling behind more permissive systems.
• It takes class time. Trying first and explaining afterwards makes AI lessons slower in an already crowded curriculum.
• Hint mode can be got around. A determined child can rephrase a question until the AI gives the answer, and a buying condition can’t fully stop that.
• The evidence for young children is thin. The main trial was with high-school students. Effects on primary pupils could differ either way.
• The money might do more elsewhere. HK$54 million could pay for smaller classes or more teaching assistants, which have a longer track record.

Our recommendation

On balance, we recommend all five measures. The case against is strongest on the evidence, which is exactly why the Labs should start as an independently evaluated three-year pilot rather than a territory-wide rollout. The cost of getting this wrong, a generation who can use AI but can’t think without it, is far higher than the cost of moving carefully.

How we’d measure it

• The key test. After each AI-assisted unit, pupils sit a short test without AI. Lab pupils should match or beat similar pupils who learned the same unit without AI. If they score lower, AI is replacing their thinking and the approach has to change.
• Reasoning. Teachers score each pupil’s “Explain” step against a simple four-level rubric.
• Reach. At least 70% attendance by eligible P4–P6 pupils in Lab schools.
• Results. Internal assessment results and class banding, compared with similar schools outside the pilot.

What’s happened since

In June 2026, the Education Bureau published its Blueprint for Digital Education Development in Primary and Secondary Schools. It includes an AI Literacy Learning Framework and a Guide to Using AI in Teaching, and requires teachers to complete at least 30 hours of digital education training in every three-year professional development cycle, starting from the 2026/27 school year.

References

• Bastani, H. et al. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” PNAS 122(26)
• UNESCO (2023). Guidance for generative AI in education and research
• Education Bureau, “EDB launches AI for Empowering Learning and Teaching Funding Programme”, 16 December 2025
• Legislative Council Research Office, “Measures to alleviate the digital divide among students”
• Working Family and Student Financial Assistance Agency, student financial assistance schemes
• Education Bureau, Blueprint for Digital Education Development
• The Standard, “New digital education blueprint to introduce AI literacy framework and mandate teacher training”, 17 June 2026

Written in light of the Employees Retraining Board reforms in the 2024 Policy Address, the government’s Report on Manpower Projection to 2028, and the 2025–26 Budget.

The issue

A shortage and a surplus at the same time. The government’s Report on Manpower Projection to 2028, published in November 2024, projects a shortfall of about 180,000 workers by 2028, with 10 of the 17 industries studied each short by more than 10,000. Unemployment is low: 3.2% (seasonally adjusted) in December 2024 to February 2025. Yet the IMF estimates that almost 40% of jobs worldwide are exposed to AI, rising to around 60% in advanced economies. Hong Kong leans heavily on finance, trade and professional services, where much of the routine clerical work is exactly what generative AI does well.

Why this turns into structural unemployment. Cyclical unemployment ends when the economy recovers. Structural unemployment doesn’t, because the jobs that come back are different from the ones that were lost. If AI cuts routine office roles while the shortages are in hands-on, person-facing work like elderly care, healthcare support, building services and construction, a displaced clerk and an unfilled care-home vacancy can sit side by side for years. The longer someone is out of work, the more their skills and confidence wear down, and the harder they are to re-employ. Economists call this hysteresis: it’s how a temporary shock becomes a permanent loss.

The system pays people to retrain after they lose their job, not before. The 2024 Policy Address accepted the recommendations to reform the Employees Retraining Board (ERB), moving it from training low-skilled workers towards skills-based training for the whole workforce. Since 1 January 2025 the ERB has dropped its education cap, opened its courses to everyone in the labour force aged 15 or above, and committed to at least 15,000 more training places in 2025–26. That’s welcome. But the ERB’s retraining allowance, up to HK$8,000 a month (HK$333 a day, or HK$166.50 a half-day), is paid for placement-tied courses aimed mainly at people who are unemployed or between jobs. A clerical worker who can see automation coming but still has a job gets no income support to retrain in the evenings. If they use the Continuing Education Fund instead, they have to pay the course fees upfront and claim the money back later. For a low-paid worker, both are real barriers.

Investment in AI has outpaced investment in the people it affects. The 2025–26 Budget earmarked HK$1 billion for a Hong Kong Artificial Intelligence Research and Development Institute. Supporting AI itself makes sense, but nothing comparable has been committed to helping workers adjust to it.

Now is the cheapest time to act. With unemployment at 3.2%, workers can retrain while they’re still employed and employers are still hiring. Waiting until displacement shows up in the unemployment figures means paying for longer spells out of work, lost tax revenue and higher welfare costs.

What we propose

1. Publish an AI Exposure Index for Hong Kong occupations. The Labour and Welfare Bureau and the Census and Statistics Department should map Hong Kong’s occupations against international measures of AI exposure, publish the results, and include them in every future Manpower Projection. Occupations in the top quarter of exposure would count as high-exposure occupations for the measures below.

2. A “Retrain Before Redundancy” allowance. Employed workers in high-exposure occupations who earn below the median monthly wage would get the ERB’s existing half-day rate of HK$166.50 per session for evening and weekend upskilling, up to 60 sessions (HK$9,990) in any 12 months. This extends an allowance that already exists to people who are still in work, rather than creating a new scheme.

3. Pay course fees upfront for at-risk workers. For the same group, the Continuing Education Fund should pay approved course fees straight to the training provider instead of reimbursing the worker afterwards.

4. Career Switch Pathways into shortage sectors. The ERB should design placement-tied courses with employers in the shortage industries named in the Manpower Projection, such as elderly care, healthcare support, building and city operations, and construction. Employers taking part would commit to interviewing every graduate of the course. Trainees would receive the existing retraining allowance of up to HK$8,000 a month.

5. A “Train-to-Keep” incentive for employers. Firms that use government technology subsidies, such as the Technology Voucher Programme, to automate work with AI could claim a top-up of up to HK$20,000 per worker they retrain into a new role instead of dismissing. That turns AI adoption from a reason to cut staff into a reason to redeploy them.

Cost

Item	Basis	Up to (HK$, 3 years)
Retrain Before Redundancy allowance	20,000 workers × HK$9,990	200 million
Train-to-Keep top-up	2,000 workers × HK$20,000	40 million
AI Exposure Index	Within existing LWB and C&SD resources	—
Upfront CEF payment	A change in payment timing, not new spending	—
Total		≈ 240 million

Career Switch Pathways use the ERB’s existing course and allowance budgets. Since not every eligible worker will claim all 60 sessions, actual spending would likely be well below the ceiling.

Funding source: the Employees Retraining Fund, which is paid for by the Employees Retraining Levy that employers pay when importing workers. As labour importation grows to fill shortages, it’s fair for that levy to fund retraining for local workers whose jobs are changing.

The case for

• Prevention is cheaper than cure. Paying a worker HK$9,990 to retrain while employed costs far less than months of unemployment, lost tax revenue and welfare support after they’ve been let go.
• It matches supply to demand. Steering displaced workers into shortage sectors tackles both problems at once and reduces how many workers need to be imported.
• It builds on what exists. Every measure extends something already in place: the ERB allowance, the Continuing Education Fund, the Technology Voucher Programme and the Manpower Projection. That keeps the admin light.
• It shares the cost fairly. Employers who gain from AI and imported labour help pay for local workers to adjust.

The case against

• The forecasts may be wrong. Nobody knows exactly which jobs AI will replace. An exposure index might flag the wrong occupations and train the wrong people.
• Some of it pays for training that would have happened anyway. Some workers would have retrained regardless.
• Labels can backfire. Calling an occupation “high-exposure” could unsettle workers, weaken their bargaining power, or make employers less willing to hire into those roles.
• Not every switch is realistic. A middle-aged office worker may not want to move into elderly care or construction, and the pay may be lower. Training can’t fully fix a mismatch in preferences and wages.
• It may slow AI adoption. Any condition on technology subsidies, however light, adds friction for firms.

Our recommendation

On balance, we recommend all five measures, starting with the AI Exposure Index, since the others depend on it. The strongest objection is that forecasts are uncertain, so the index should be reviewed every year and the high-exposure list updated as the evidence changes. The Train-to-Keep incentive is deliberately a reward rather than a requirement, so it encourages firms to redeploy staff without penalising them for adopting AI.

How we’d measure it

• Targeting. The share of allowance recipients who come from high-exposure occupations.
• Outcomes. Participants’ employment and earnings 12 months after training, compared with similar workers who didn’t take part.
• Mismatch. A quarterly figure from the Census and Statistics Department on job vacancies per unemployed person, by occupation group. A widening gap between groups is the early warning sign of structural unemployment.
• Time out of work. The share of unemployed former clerical workers who have been out of work for more than six months.

What’s happened since

• The 2026–27 Budget (February 2026) announced that the ERB will be upgraded as Upskill Hong Kong, offering skills-based courses including AI applications.
• Hong Kong’s seasonally adjusted unemployment rate rose to 3.8% in December 2025 to February 2026, up from 3.2% a year earlier.
• In September 2026, the Labour and Welfare Bureau’s Mid-term Update of Manpower Projection projected that, as companies use more AI for routine admin, clerical support workers will see a surplus of about 25,000 to 30,000 by 2028, while the overall shortage narrows to around 130,000.

References

• Census and Statistics Department, unemployment and underemployment statistics for December 2024 – February 2025, 18 March 2025
• China Daily HK, “Report: HK may face huge manpower shortage by 2028”, 14 November 2024
• The Chief Executive’s 2024 Policy Address, paragraph 208
• Employees Retraining Board, skills-based training strategy from 1 January 2025
• Employees Retraining Board, maximum monthly retraining allowance raised to HK$8,000
• Legislative Council Question 20, “Development of artificial intelligence”, 19 March 2025
• IMF, “AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity”, January 2024
• ERB, on the 2026–27 Budget’s plan to upgrade the ERB as Upskill Hong Kong
• news.gov.hk, March 2026
• Bastille Post, on the Mid-term Update of Manpower Projection