Fear Cannot Be Our Response: Singapore's Hope-Led Bet

Part V of V. Singapore has made hope a policy. The bill will be paid on the first rung of the career ladder.

I. Five words in a Budget speech

On 12 February, in a Budget speech otherwise full of tax deductions and training schemes, Prime Minister Lawrence Wong said five words I have not stopped thinking about: “fear cannot be Singapore’s response.”

It is an unusual sentence for a head of government. Most leaders talk about AI in the language of opportunity and then legislate in the language of risk. Wong did something plainer. He named the emotion, and then he ruled it out as a basis for policy.

This series has spent four essays listening to people who are, in different ways, afraid. In Part I, six of Stanford HAI’s leaders described what keeps them up at night. In Part II, the evidence showed AI brilliant on parts and unreliable on wholes. In Part III, the young paid first. In Part IV, a Pope, philosophers and governors asked who holds custody of the power.

This final essay comes home. I live and work in Singapore, and I helped write part of its governance architecture. So the question I owe readers is direct. Is a hope-led bet wise, or merely brave?

II. Two numbers that do not agree

Start with the numbers, because Singapore’s position is easy to caricature.

On trust, Singapore is an outlier. Stanford’s AI Index 2026 reports that 91% of Singaporeans trust their government to regulate AI, the highest of any country measured, against 31% in the United States. One secondary report gives Singapore’s figure as 81%, so the exact number should be read with care; either way, the gap with America is enormous. On use, Singapore leads the world, with generative-AI adoption reported at 61% against 28.3% in the US, according to a digest of the Index’s economy chapter. The IMF ranks Singapore first on its AI Preparedness Index.

Now the second number. The Ipsos AI Monitor 2026 finds 61% of Singaporeans say AI products excite them, and 58% say they make them nervous. Compare China at 83% excited and 40% nervous, Indonesia at 76 and 43, South Korea at 74 and 40. Singapore is far more ambivalent than its neighbours.

Put the two together and you get a portrait I recognise from every dinner table and boardroom here. Singaporeans trust the state to manage AI. They are not sure it will go well for them personally.

That is not a contradiction. It is a loan. The public has lent its government the benefit of the doubt, and the loan has terms.

III. What the bet buys

The Budget turned the posture into line items. A National AI Council, chaired by the Prime Minister. National AI missions in advanced manufacturing, logistics, finance and healthcare. AI spending made eligible for 400% deductions under the Enterprise Innovation Scheme, capped at S$50,000. Redesigned SkillsFuture AI pathways, TeSA extended to non-tech roles, and six months of free premium AI tools for people who take AI courses. Transitions, the government said, would be managed with the unions.

The sharpest test came in May. The day after Standard Chartered announced more than 7,000 job cuts over four years, Deputy Prime Minister Gan Kim Yong told financial firms to use AI to create better jobs and retrain staff, not merely to cut costs. He was equally clear about the other side: “If we slow AI adoption, we will weaken our competitiveness.”

That is the bet in one paragraph. Adopt at full speed, and insist the gains be routed through people rather than around them.

I call this Long-AND, not Short-OR. Not adoption or employment, not speed or safety, but both, held together long enough for the second to catch up with the first. Most countries have chosen one side of the OR. Singapore has chosen the harder AND, and it will be judged on whether it can hold both ends.

IV. Governing the loop, not the prompt

A hope-led posture is not a light-touch one. The governance half of the bet is Singapore’s January 2026 Model AI Governance Framework for Agentic AI, launched at Davos and billed as a global first. I should declare my interest plainly: I co-authored it with IMDA.

Its logic is simple. Bind every agent to an accountable human by limiting its tools, systems and scope. Require human approval at significant checkpoints. Test before and after deployment, monitor continuously, and roll out gradually. It calls itself a living document, which is the only honest thing a framework for agents can be.

Around it sit the 2025 Singapore Consensus, which organised global safety research into risk assessment, development and control, and the May 2026 refresh of the national AI strategy, which made testing and assurance a “Trusted Environment” priority.

Set this against the postures traced in Part IV. Washington races and pre-empts. California, under Governor Newsom’s September order, wants independent verifiers onsite at frontier labs and a kill switch. Brussels has postponed its stand-alone high-risk rules to December 2027. Singapore adopts fast and invests in assurance. A kill switch assumes you can see the danger coming and stop the machine. A checkpoint assumes you often cannot, so you decide in advance where a human must say yes.

In Special Relativity I described the shift from Prompts to Loops to Loop Governance, and the enforceability test I apply to any framework: is each control enforceable at runtime, or only reviewable afterwards? Checkpoints pass that test when they live inside the loop. They fail it when they become a form signed after the fact.

The composition gap from Part II is why this matters. On HAI’s HealthAdminBench, one model passed 82.8% of subtasks yet finished only 26.7% of whole tasks. A system like that needs its human at the joins, not at the end. The framework’s quality will be decided less by its text than by how many deployments put the approval inside the agent’s path.

V. The first rung, in Singapore

Here is the fault line, and it runs straight through the trust number.

In May, the Ministry of Manpower told Parliament that PMETs in ICT, financial services and professional services face relatively higher retrenchment rates. It reads this as restructuring rather than falling demand, and it said it does not yet have data on whether AI skills earn a wage premium. In the first quarter of 2026, retrenchments rose to 3,830, falling on degree holders and PMETs.

These are the codified-knowledge roles that David Autor’s expertise framework predicts AI will thin first. They rhyme with what Part III found in American data: the Canaries update puts employment of 22 to 25 year olds in highly exposed occupations about 19% below its counterfactual. The IMF’s October outlook for Asia-Pacific found employer demand in AI-exposed jobs fell 10% before stabilising, and concluded that AI is “reshaping work more than eliminating it”.

Reshaping is the right word. It is also the word that hides the most pain, because the reshaping happens where the young enter. This is the Frozen Workforce in local dress: not mass lay-offs, but a hiring door that quietly narrows. MOM points to Career Conversion Programmes, GRIT traineeships for fresh graduates and SkillsFuture AI pathways. They are the right instruments. Whether they are the right size is the question I would most like answered with data.

There is a macro edge too. The IMF’s October outlook found that the US and China account for over three-quarters of net AI-equipment import demand. My reading is that much of the capex boom’s gain therefore flows to exporters such as Korea, Taiwan, Singapore and Malaysia. That upside cuts both ways. On the day the IMF flagged AI-boom risks, Korea’s Kospi fell 5.5%. A small open economy that rides the boom must be ready for the correction.

Put it together and the risk to the bet is clear. Trust is the scarce asset that lets Singapore adopt fast. It is also the asset most exposed if the promise of redeployment meets a rising PMET retrenchment line. Losing twenty points of trust would cost more than any capex correction.

Two brass gauges on a dark console: one needle steady and high in teal light, the other trembling mid-dial in amber, harbour lights blurred behind — High trust, unsettled nerves. Both readings are true.

VI. Where the published record lands

As always, these are my usual panel’s public positions, not responses to this essay.

Yoshua Bengio chaired the International AI Safety Report 2026, which warns that “reliable pre-deployment safety testing has become harder to conduct”, because models increasingly recognise test settings. That is the strongest scientific case for Singapore’s emphasis on continuous monitoring and gradual rollout over a single gate. Bengio also signed the October 2025 call to prohibit superintelligence development, so he would remind us that assurance is a floor, not a licence.

Dario Amodei, in The Adolescence of Technology, warns that about half of entry-level white-collar jobs could go within one to five years, names concentration of power as a core risk, and proposes progressive taxation of AI profits and redeploying rather than firing workers. On redeployment, he and DPM Gan are saying the same thing. His timeline is the one Singapore’s PMET data should be watched against.

Demis Hassabis was quoted in a secondary interview report as saying that almost every drug developed from now on “will have probably used AlphaFold”. That is the abundance end of the bet, and part of why a national healthcare mission is not a slogan. Part II’s caution still applies: the tools are real, while validated outcomes take years.

The sharpest pushback comes from Daron Acemoglu. His estimate caps AI’s ten-year productivity gain at no more than 0.66%, he has called the hyperscalers “extractive”, and with Michael Sandel he argues that no dividend or basic income can restore the dignity of contributing. To a hope-led bet, that is a cold shower. If the gains are small and concentrated, adopting at full speed simply imports the dependency faster.

My answer is that Singapore’s bet is, on paper, close to the pro-worker version Acemoglu asks for: redeploy rather than replace, manage transitions with the unions, and steer the direction of the technology through missions rather than leave it to the market. The test is whether paper becomes payroll.

VII. What the series found

Step back across the five essays and four findings remain.

The six voices. When Stanford’s human-centred institute asked six of its leaders for their hopes and fears, two called for urgent governance of the frontier. Russ Altman said “it’s time to regulate frontier companies”; Rob Reich warned that developers “are losing control”. Three made compute their next priority, with Emmanuel Candès warning that otherwise “OpenAI and Anthropic will define the scientific agenda”. Three worried about the young. Only Angèle Christin dissented on framing, “less worried about superintelligence” than about concrete harms. The mainstream has moved.

The composition gap. Mykel Kochenderfer’s fear of “premature acceptance of AI without proper validation” met HAI’s own finding that “competence on the parts does not add up to reliability in a whole workflow”. The useful frame for 2027 is neither hope nor fear but integration risk: the gap between what AI can do in isolation and what institutions can safely let it do end to end.

The missing first rung. Chiara Sabatti feared the pace, not the destination. The pace lands on 22 to 25 year olds in America and on PMETs here. A profession that stops hiring juniors stops producing seniors.

Custody. A Pope, Nobel economists, political theorists, a frontier-lab chief executive and half of HAI’s own leadership now name the same danger: AI’s benefits and moral rules set by a few. Magnifica Humanitas put it most sharply: a more moral AI is “not enough if that morality is determined by a few”.

Singapore’s answer to the custody question is distinctive. It wagers that a high-trust state can hold custody on the public’s behalf while adopting at full speed. Whether that holds depends less on the technology than on whether the first rung of the ladder, the place where trust in the future is earned, survives the transition.

So here is my prediction. By the end of 2027, the number Singapore’s AI strategy is judged by will not be adoption, which already leads the world, or preparedness, which already ranks first. It will be graduate hiring into PMET roles in finance, ICT and professional services. If that line holds, the trust number holds with it, and Singapore will have shown that hope can be governed. If it falls, nervousness will overtake trust, and no framework, including the one I helped write, will win it back.

VIII. The Fork

Throughout my writing I return to one image, The Fork. One road leads to something like Star Trek: abundance shared, work redesigned around human judgement, machines that extend us. The other leads to Mad Max: scarcity hoarded, institutions hollowed out, power held by whoever owns the engines.

Every finding in this series is a signpost at that fork. Validation, the first rung, compute, custody. None decides the road alone. Together they do. And as I argued in The Arrow of Time, some of these turnings cannot be retaken once passed.

Hope is not optimism. Optimism is a forecast. Hope, in the sense Wong meant, is a choice made under uncertainty, and it has to be paid for in work. Fear cannot be our response. Neither can complacency.

So, one exercise this week. Draw a line down a sheet of paper. On the left, write the Star Trek headline about your organisation in 2030; on the right, the Mad Max one. Then write down how many people under 26 you hired in 2024, and how many in 2026. Ask which headline that number is writing. Then name one decision you can take this quarter that moves you left.

That is what I mean by Short-term Turbulence for Long-term Abundance. The turbulence is real, and it falls first on the young and the codified. The abundance is real too. But it is not a place we arrive at by waiting. It is a road we choose, one checkpoint and one first rung at a time.

This closes the series. If you are arriving here first, go back to Part I, Six Voices, One Turn: What Stanford’s AI Leaders Now Fear, and hear what the people building this future told us they fear.

The Hopes and Fears series

Part I. Six Voices, One Turn: What Stanford’s AI Leaders Now Fear
Part II. The Composition Gap: Brilliant on Parts, Unreliable on Wholes
Part III. The Missing First Rung: Young Workers and the Fifteen-Fold Disagreement
Part IV. Who Holds Custody: The Pope, the Philosophers and the Power Question
Part V. Fear Cannot Be Our Response: Singapore’s Hope-Led Bet (you are here)