Reframing Welfare and Poverty Policies in Hong Kong amidst the Surging AI Era

I · Introduction

I spent twelve years inside the machine. Seven bureaux. Enough submissions to fill a small room.

You learn a particular kind of humility there. You learn that a policy is not what the paper says. A policy is what the system can actually do on a Tuesday morning in a district office, with the staff it has, under the ordinance it was given.

So when I read that Hong Kong will deploy artificial intelligence across two hundred public service procedures by the end of 2027, I do not read it as a technology story. I read it as a welfare story that nobody has written yet.

Here is the shape of the problem. The Government is spending more on social welfare than at any point in its history. Recurrent expenditure on social welfare is estimated at $135.9 billion in 2026–27 — 22.7 per cent of all recurrent government expenditure, and 9.9 per cent above the revised estimate for 2025–26. At the same time, the same Government is cutting recurrent expenditure by two per cent in each of the next two financial years and deleting more than ten thousand civil service posts within this term. And at the same time again, it is standing up a billion-dollar artificial intelligence institute, a three-billion-dollar AI subsidy scheme, and a data facility cluster at Sandy Ridge.

Three trains. Three timetables. One set of tracks.

Money spent is not the same as poverty reduced. The two are related. They are not the same sentence.

This commentary argues something uncomfortable. Hong Kong's welfare architecture is well-built and well-funded for a world that is ending. It is designed to catch people who lose their jobs. It is not designed to catch people who keep their jobs while the value of their work quietly drains away.

That is not a criticism of the officials who built it. I was one of them. It is a statement about timing. Every welfare system in the world was designed against the last transition. Ours was designed against deindustrialisation — against the closure of the factories, the shift of manufacturing north, the retraining of the machine operator into the security guard. It worked. It is still working. The Comprehensive Social Security Assistance Scheme remains one of the most reliable last-resort floors in Asia.

But the next transition does not look like the last one. The last transition took the job away and left the person standing in the street. You could see it. You could count it. The next transition leaves the person at the desk, with the title, with the payslip — and takes the task. Then the next task. Then the reason to hire the person behind them.

You cannot see that in an unemployment rate. Hong Kong's seasonally adjusted unemployment rate stood at 3.7 per cent in April to June 2026, unchanged from the preceding three-month period, with underemployment at 1.6 per cent. By any historical standard, that is a healthy labour market. It tells us almost nothing about what is coming.

Unemployment is a photograph. Obsolescence is a film.

This commentary applies the 1134211 discipline to the question. Three policy problems. Four policy options, one of them a balanced recommendation. Two implementation challenges with mitigation. A stakeholders' assessment. Then metrics, and one thing to do on Monday.

II · Background: The Architecture We Actually Have

Hong Kong does not have a welfare state. It has a welfare architecture — a set of separately legislated, separately administered instruments, each built for a specific contingency, assembled over five decades without a unifying statute.

The load-bearing pillars are these. The Comprehensive Social Security Assistance Scheme provides a means-tested cash floor to bring household income up to a prescribed level. The Social Security Allowance Scheme delivers non-means-tested and means-tested allowances to the elderly and to persons with disabilities. The Working Family Allowance supplements low-income working households. The Mandatory Provident Fund provides a defined-contribution retirement pillar. Public rental housing, subsidised healthcare and free basic education deliver the largest transfers of all, in kind rather than in cash.

Note what is absent. There is no contributory unemployment insurance. There is no statutory wage-replacement scheme. A worker who loses employment in Hong Kong falls, with no intermediate landing, from a wage to a means-tested last resort assessed on household assets.

That gap was tolerable when job loss was episodic and re-employment was quick. It is the single most important structural fact in this commentary.

In 2021, the Government published the Hong Kong Poverty Situation Report for 2020. It recorded a pre-intervention poverty rate of 23.6 per cent, some 1.653 million people. Then the series stopped. The current-term Government replaced poverty-line reporting with a strategy of targeted poverty alleviation. On 18 June 2026, the Commission on Poverty released the Report on Impact of Targeted Poverty Alleviation Strategy in Hong Kong. It is a serious document. It is also, methodologically, a report about inputs.

We stopped counting the poor in 2021. The poor did not stop counting.

Meanwhile the arithmetic of ageing runs underneath everything. Persons aged 65 and above are projected to rise from 1.45 million in 2021 to 2.74 million by 2046 — more than one in three. The total fertility rate stood at 841 live births per 1,000 females in 2024. The labour force is projected to peak at 3.66 million in 2038, then decline.

Against that slow demographic clock, a fast technology clock: the Hong Kong Artificial Intelligence Research and Development Institute commences operation this year; AI across one hundred public procedures within 2026, two hundred by end-2027; a $3 billion AI Subsidy Scheme; the civil service establishment falling to about 188,000 posts.

III · Policy Problems

Three problems. Each stated at Layer 3, because Layer 1 statements of the obvious are what fill most commentary on this subject.

**Problem 1 — The state measures relief, not exposure.**

The administration has a high-resolution picture of what it delivers and a low-resolution picture of what households face. Artificial intelligence does not produce its labour effect through visible dismissal. It produces it through the quiet withdrawal of the junior rung — fewer graduate openings, smaller teams, higher output expectation per head, a slow re-specification of what a clerk or a paralegal or a customer service officer is for. None of that registers as unemployment until years later. By then the cohort that should have been caught is thirty-four years old with eight years of narrowing experience.

A government that can only see the fall cannot govern the slope.

**Problem 2 — The safety net is calibrated to the absence of work, not the obsolescence of skill.**

A displaced mid-career worker with a modest flat, a spouse in work and some savings will typically fail the household means test, receive nothing, and burn savings during transition. The architecture encodes an assumption from the last century: that joblessness is temporary and skill is durable. Under AI diffusion, both invert.

The International Monetary Fund's staff analysis estimates that around 40 per cent of global employment is exposed to artificial intelligence, rising to roughly 60 per cent in advanced economies. Hong Kong, with a services share of output above 90 per cent and one of the most cognitively concentrated occupational structures in Asia, sits at the exposed end of that distribution.

A safety net catches people who fall. It does not catch people whose ladder is quietly removed.

**Problem 3 — The state is simultaneously the accelerant, the largest routine-cognitive employer, and the payer of last resort.**

The Government occupies three roles at once, and no institutional mechanism reconciles them: accelerant — funding, procuring and deploying AI across the economy and its own operations; employer — one of the largest single employers of routine cognitive labour in the territory; and insurer of last resort — the sole payer when displacement lands. The accelerant sits with the Innovation, Technology and Industry Bureau; the employer function sits with the Civil Service Bureau; the insurance function sits with the Labour and Welfare Bureau. Each performs competently against its own objective.

Nobody owns the sum.

The Government is at once the largest employer of routine cognitive labour, the largest funder of the safety net, and the accelerator of the technology that erodes the first. Three roles. One balance sheet. No reconciliation mechanism.

IV · Policy Options

Four options. I set out costs honestly, because a commentary that only lists benefits is advocacy, not analysis.

**Option A — Incremental Reinforcement.** Continue the present course. Expand targeted poverty alleviation. Index CSSA and Social Security Allowance rates. Keep publishing the Impact Report. In favour: no legislation required, politically frictionless, protects the surplus. Against: it answers none of the three problems. It is a good policy for 2015.

**Option B — A Skills-Contingent Transition Allowance.** Create a time-limited, individually assessed Transition Income: a partial wage-replacement payment conditional on enrolment in a recognised reskilling pathway under Upskill Hong Kong, assessed on the individual's earnings trajectory rather than household assets. In favour: attacks Problem 2 directly, uses the open Employees Retraining Ordinance amendment window. Against: deadweight loss, moral hazard, and the eligibility trigger is genuinely difficult.

**Option C — Statutory Unemployment Insurance.** Legislate a contributory unemployment insurance scheme layered on MPF's existing collection infrastructure. In favour: the textbook answer, it pools risk, removes the cliff between wage and last resort. Against: legislative runway beyond this term, employer resistance high, and it insures against unemployment — which, per Problem 2, is not the primary AI-era risk. It is the right answer to the wrong century's question.

**Option D — Recommended: The Reframed Three-Layer Compact.** Not a scheme. A sequence, in which each layer earns the right to the next.

• Layer One — Measure exposure, not just relief. Commission an Occupational Task Exposure Baseline for Hong Kong's employed population. A task-level mapping against generative-AI exposure and complementarity, disaggregated by age, education, sector and household income decile. Publish it. Update it annually. Cost: a research commission.
• Layer Two — Convert the retraining allowance into a Transition Income. A pilot restricted to occupations in the top exposure quintile, time-limited to nine months, training-contingent, incubated via the Community Care Fund with a hard sunset and published evaluation.
• Layer Three — Fund it from the transition that is causing it. Ring-fence a fixed, published proportion of the existing AI programme envelope as a standing Labour Transition Reserve. Not a new tax on artificial intelligence. A hypothecation discipline inside the same budget cycle.

Two roads diverged. One is paved with transfers. The other is paved with transitions. We can afford to walk part of both — but not if we refuse to survey either.

Why recommend Option D? Because it is the only one of the four that acts on all three problems, requires no primary legislation to begin, and produces evidence that improves the next decision. It buys optionality. In a domain this uncertain, optionality is the product.

V · Implementation Challenges

Two, and both are severe.

**Challenge 1 — You cannot administratively prove that a machine took the job.** No employer will record "replaced by artificial intelligence" on a termination. Attribution is contested, often genuinely unknowable. Mitigation: abandon causation, adopt task-based triggering. Eligibility should never ask why the job ended. It should ask two verifiable questions: what was the last occupation, and where does that occupation sit in the published Exposure Baseline? If it falls in the top exposure quintile, a rebuttable presumption of transition risk applies.

You cannot prove a job was taken by a machine. You can measure the task that was.

**Challenge 2 — The fiscal and institutional collision.** Social welfare already absorbs $135.9 billion, 22.7 per cent of recurrent expenditure, growing at 9.9 per cent. The Productivity Enhancement Programme demands two per cent recurrent cuts. There is no space. That is not obstruction; it is arithmetic. Mitigation: fund the pilot from non-recurrent provision and the Community Care Fund; legislate a hard expiry with mandatory published evaluation; sequence the ask — Baseline first.

The institutional collision is harder. Three bureaux, three objectives, no owner. The only realistic mitigation is to place the Baseline under the Commission on Poverty — a body that already reports to the Chief Secretary, already publishes on poverty impact, and already convenes across bureaux and the business sector. It does not require new machinery. It requires an added term of reference.

VI · Stakeholders' Assessment

The Labour and Welfare Bureau is the natural policy owner and the most likely quiet supporter. Its power play is the Impact Report itself: whoever writes the evidence base frames the next Policy Address.

The Financial Secretary's Office will resist, correctly and predictably. Their compromise is a costed, sunset pilot funded outside the recurrent baseline. Their power play is the Estimates: nothing exists until it appears in a Head, and they control the Head.

The Innovation, Technology and Industry Bureau and the Digital Policy Office occupy the most interesting position. Publicly they will welcome anything that builds social licence for AI adoption. Privately they will resist any framing that attaches a social cost to their programme envelope. Their power play is speed: technology bureaux move faster than welfare bureaux, and the file that moves first sets the terms.

Employers will resist Option C absolutely and Option D conditionally. Labour unions will treat Option D as insufficient — a pilot is not a right. The welfare sector will support the Baseline enthusiastically, because their case has been evidence-starved since 2021.

Civil service staff associations are the stakeholder nobody lists, and they matter most. Ten thousand posts are being deleted while AI enters two hundred procedures. Their power play is institutional memory and operational cooperation — the quiet, decisive currency of implementation.

The synthesis: no stakeholder opposes measurement. Almost every stakeholder opposes some part of the programme. That asymmetry is the entire strategic case for sequencing the Baseline first: it is the only move on the board that nobody has a principled reason to block.

VII · Conclusion and Synthesis

Hong Kong is not failing its poor. That story is lazy and I will not write it. A city that commits $135.9 billion of recurrent expenditure to social welfare, that legislated the Basic Housing Units Ordinance to end substandard subdivided units for some 220,000 residents, that runs cross-boundary payment arrangements so that elderly recipients in Guangdong and Fujian are not stranded by a border — that city is trying hard, with competence, at scale.

The risk is narrower and more dangerous than failure. The risk is succeeding at the last problem while the next one forms.

Ninety-one per cent of Hong Kong professionals report already using AI tools at work. Only sixty-six per cent believe it will benefit their careers — the lowest optimism of the ten Asian markets surveyed, against eighty-eight per cent in the Mainland.

Ninety-one in a hundred already use it. Sixty-six in a hundred believe it will be good for them. That gap is not a mood. It is a policy problem that has not yet found a policy.

The Commission on Poverty should add a single line to its terms of reference and commission, for publication alongside the next Impact Report, an occupational task-exposure baseline for Hong Kong's employed population.

Not a scheme. Not a levy. Not a statute. A number, published, updated annually, that tells this city where its work is thinning before the caseload tells us afterwards.

We built a net for people who fall. The next transition will not push people off the ledge. It will lower the ledge, slowly, while they are standing on it, and they will not know until they look down.

**Measure the ledge.**

*Written under the 1134211 Framework · August 2026 · Sources: 2026–27 Budget and Estimates, 2025 Policy Address, Census and Statistics Department, Legislative Council papers.*