The Last Human Job Is Answerability; Not Empathy. Answerability.

A lone figure holding an amber lantern before a wall of glowing AI servers at night

Kai-Fu Lee says America invents, China executes, open models spread, and love will keep us human. He is right about the direction of travel. But the map he is using is out of date, history reads it differently, and on the human question he is far too gentle with us. The real divide is not innovation versus execution but trust, and what remains ours is not empathy but the willingness to answer for what we do.

I. A post, a frame, and a missing interview

Kai-Fu Lee posted on LinkedIn this month about a conversation with Walter Isaacson on Amanpour and Company, billed as “The AI Race: Can the U.S. and China Avoid Disaster?” I have tried to find that segment and failed: the PBS guest page for Lee lists only his 2018 and 2021 appearances. So I will not put words in his mouth from a broadcast I have not seen. I will take his post at face value, and lean on what he said at length in September to Bloomberg’s Mishal Husain and to Semafor, which make the same two arguments.

The first argument: the United States leads frontier innovation, China excels at execution, and open models accelerate adoption everywhere. In Bloomberg he gave it a memorable shape. US labs are the premium iPhone, Chinese open-weight labs the cheap, scaled Android, and China sits about six months behind with a sliver of America’s compute because, as he put it, “policymakers underestimated that tenacity.” China wins reach, America wins profit, and Chinese models cost roughly a sixth to a tenth as much.

The second argument: as AI remakes work, we must rethink what makes us human, and purpose will come from love, empathy and connection. This is not new for Lee. In 2018, after his cancer diagnosis, he told a TED audience that “Love is what differentiates us from AI.” He is now selling the leadership version in a new book, AI Native: The Mandate to Transform Your Company, out in print in November.

I have admired Lee for a long time, and I agree with more of this than I disagree with. But I want to argue three things. The frame is stale, and historians of technology have been telling us why for decades. The real fault line is trust. And “empathy” is the weakest possible place to plant the human flag. To make the case I have widened my panel: historians, neuroscientists, philosophers East and West, novelists, middle-power strategists, unions, judges and, because they price everything, insurers.

II. Months, not eras

Begin with the gap, because everything else depends on it. Epoch AI finds Chinese models have trailed the US frontier by seven months on average since 2023, ranging from four to fourteen. Stanford’s AI Index 2026 put the top US model ahead by 2.7%, or 39 Elo, on Arena in March, and called the performance race “effectively closed.” The US government’s own evaluators at CAISI placed DeepSeek V4 Pro about eight months behind, while noting it was cheaper than GPT-5.4 mini on most benchmarks. Then in July Moonshot’s Kimi K3 reportedly ranked third on the Artificial Analysis index, and Nathan Lambert called it “the strongest open model ever released.”

So yes, America leads. By months. You cannot build a civilisational division of labour on a lead measured in quarters.

Second, the “execution” label is a category error. DeepSeek, Qwen and Kimi releases are research outputs, not deployment services. DeepSeek’s founder Liang Wenfeng said as much in 2024: “China’s AI cannot remain a follower forever.” Lee himself told Jiemian in early 2025 that R1 contained original contributions, and that paying billions to be slightly better than free DeepSeek was “an unsustainable business model.” His own company is the proof: 01.AI stopped pre-training in March 2025 and now deploys other people’s open models, branding itself China’s Palantir. The frame flatters the business model; I say that with affection.

The strongest defence of Lee comes from Demis Hassabis, who at Davos called Chinese labs “very good at kind of catching up to where the frontier is” while doubting they have yet gone beyond it. Lee also concedes America’s top 1,000 researchers beat China’s top 10,000. Fair. But “fast follower” is a statement about speed, not about who innovates.

Now turn his analogy. If Chinese labs are Android, notice who is building on Android. Chinese models were 41% of Hugging Face downloads over the year to spring 2026, overtaking the US, with Qwen derivatives dominating the top of the charts. Airbnb’s Brian Chesky said his customer-service agent leans heavily on Qwen because it is good and “also fast and cheap.” Cursor built its Composer 2 coding model on Moonshot’s Kimi K2.5 without saying so until Kimi staff noticed; its co-founder conceded “It was a miss to not mention the Kimi base.”

Android did not win because it was cheaper. It won because anyone could ship it on their own hardware without asking permission. That is the clue to what the race is really about.

III. Zero to one, one to ten: what history says

Lee’s frame has a respectable Chinese pedigree. The scholar Huang Ping argues China is weak at original “0 to 1” breakthroughs but strong at “1 to 10” scaling, drawing on historian Wu Guosheng’s answer to the Needham question of why a civilisation that invented so much did not produce the scientific revolution. Huang’s conclusion is subtler than Lee’s: in a long rivalry, “innovative resilience may matter more than the capacity for breakthrough innovations.” HKUST’s Erik Baark has since flipped Needham on its head, asking why China is now able to contribute to global science. The question has changed its subject in a generation.

Dan Wang’s Breakneck supplies the most quoted version of the execution story: “America is run by lawyers, and China is run by engineers.” Wang locates China’s edge in process knowledge, where innovation emerges from the factory floor. Persuasive about high-speed rail. Less obviously true of the AI economy, where the factory floor is a spreadsheet in a regional bank.

Here Jeffrey Ding cuts deepest, with a blade that has two edges. His prize-winning Technology and the Rise of Great Powers argues that nations rise not by inventing general-purpose technologies first but through their “adoption of that technology at scale across its entire economy.” Britain in steam, America in electricity and computing: diffusion beat invention every time.

That sounds like a vindication of “China executes.” It is the opposite. Ding judges the United States better placed to build the skills infrastructure for diffusion and diagnoses a Chinese “diffusion deficit.” So the historian who most agrees with Lee that execution matters is the one most sceptical that China is better at it.

Wang and Ding offer two theories of execution, factory floor versus economy-wide skills, pointing to different winners. Lee has collapsed them into one word.

Carlota Perez adds the time dimension. She treats AI not as a new revolution but as a late development within the ICT surge that began with the microprocessor. Each surge has a frenzied installation phase, a turning point, then a deployment golden age, and she is blunt about the sequence: “I have not seen a golden age happening without a crash.” The golden ages came only when society reshaped the technology through antitrust and the welfare state. The question is not who installs fastest. It is who completes the turn.

Paul David’s dynamo tells the same story inside the factory: electricity arrived in the 1880s, productivity in the 1920s, once factories were redesigned around it. Brynjolfsson’s J-curve explains why early gains hide in intangible investments the national accounts miss, and in February he declared America was entering a harvest phase. Apollo’s Torsten Slok replied with a modern Solow: “AI is everywhere except in the incoming macroeconomic data.” Neither script is about chips.

The 2025 Nobel laureates offer the institutional verdict. Joel Mokyr is bullish but warns “It’s not going to create abundance,” and fears US cuts to basic science will push frontier work abroad, primarily to China. Philippe Aghion is simpler still: “My worry is on the institutions.” He fears incumbents using their rents to block entrants, a race inside each country as much as between them.

And a small irony worth savouring. In January the White House published “Artificial Intelligence and the Great Divergence”, casting AI as a second divergence America should lead.

Its first footnote cites Kenneth Pomeranz. But Pomeranz’s thesis was that the original divergence was contingent: the Yangzi Delta and England were comparable until about 1800, and Europe pulled ahead on coal and colonial “ghost acres,” not civilisational genius. Read honestly, Pomeranz says the deciders are energy, compute and supply chains. That is a warning to triumphalists, not a licence.

History, then, says three things to Lee. Diffusion beats invention. On diffusion, the United States may still be ahead. And no one wins the golden age without first answering to society for the crash.

IV. Nineteen days in June

A bank trading floor at night with screens going dark and one worker standing still

On 12 June 2026, a Friday evening, the US Commerce Department told Anthropic that any foreign person, anywhere, would need a licence to access its two newest models, Fable 5 and Mythos 5. Anthropic said it “must abruptly disable Fable 5 and Mythos 5 for all our customers.” It was, as the National Committee on US-China Relations noted, the first time models rather than chips were controlled. Access was restored on 30 June, with Secretary Lutnick saying the administration “reserves the right to reimpose restrictions.”

Nineteen days. For a bank in Jakarta or a ministry in Seoul running agents on those models, that was the lesson of the year: access to closed frontier intelligence is a policy variable set in someone else’s capital. CSIS warned that the move could push foreign buyers towards open-weight models they run themselves, and the leading open-weight families are increasingly Chinese.

This is why I say the divide is not innovation versus execution. It is closed versus open, and beneath that, trust and sovereignty. Every serious buyer now asks the question I put to boards across Asia Pacific: can I still run this if Washington, or Beijing, changes its mind on a Friday? TrustOS, my seven-layer architecture for governing agents, starts from a blunt premise: an agent you cannot guarantee access to is an agent you cannot govern.

Chris Miller, whose Chip War taught a generation to think in chokepoints, now says “Not a single country in the world” can make an advanced chip alone, and that the next question will be trust in foreign-made autonomous systems. The historian of chokepoints has arrived at trust.

Beijing has noticed the symmetry. On 16 October MOFCOM published draft export controls on model weights covering models of two trillion parameters or more, or “globally leading” performance, while exempting or automatically licensing open-source releases. China is controlling the closed crown jewels and keeping the open spigot running. That is Android strategy written into regulation.

The region’s elites are reading the same signals. ISEAS’s State of Southeast Asia 2026 found that, if forced to choose, 52% would align with China and 48% with the US; a third of those distrusting Washington cited a threat to their sovereignty, and Japan was the most trusted partner of all. Brookings’ Douglas Rediker gives middle powers a doctrine, “modular sovereignty”: pool compute and training, keep control of deployment, because “Regulation without independent capability is dependence management.” Mark Carney told Davos that middle powers must not be “forced to choose between hegemons and hyper-scalers.”

The middle powers, meanwhile, have chosen usefulness over stringency. The EU’s Digital Omnibus pushed high-risk obligations to late 2027 and 2028. Japan’s AI Promotion Act carries no monetary penalties; Korea’s AI Basic Act tops out at fines of about US$20,000. India’s New Delhi Declaration was endorsed by 92 parties. Singapore published the Model AI Governance Framework for Agentic AI in January, which I co-authored with IMDA, and set up a National AI Council chaired by the Prime Minister. And 29 states, Indonesia and Laos among them, signed up to China’s World AI Cooperation Organisation in July.

As for “avoiding disaster”: the September Trump-Xi summit produced an AI incident channel and a November dialogue, alongside Trump’s boast that America is “leading by at least a year, maybe a year and a half.”

History offers a sober template. Graham Allison reminds us that US and Soviet negotiators found narrow shared interests long before broad trust, but that AI differs because private firms, not governments, invented and own it. The narrow template has worked once already: in Lima in November 2024, Biden and Xi agreed that humans, not AI, should control decisions to use nuclear weapons. The Brookings and Tsinghua track has proposed a military hotline for AI incidents, though Carla Freeman notes Chinese officials did not pick up during the 2023 balloon affair.

Apply my enforceability test: what happens, concretely, when someone breaks it? For a hotline, nothing; it is a floor, not a fence. A floor is better than the void of 2024. It is not governance.

V. The view from the South

I have been arguing that trust and sovereignty drive adoption. Honesty requires me to show the data that complicates it.

Microsoft’s diffusion telemetry for the second quarter of 2026 found the Global North at 28.8% working-age usage against the South’s 16.2%, a gap that “continued to widen”; Singapore stood at 64.3%. Earlier, the same team estimated DeepSeek’s share at about 56% in Belarus, 49% in Cuba, 43% in Russia, and 11 to 14% in Ethiopia, Zimbabwe, Uganda and Niger, its spread tracking its status as the default chatbot on Chinese handsets. Microsoft’s Juan Lavista Ferres summed it up: “We are seeing a divide.” An OECD survey of Gen Z, to be fair, found young people in the South adopting faster than their Northern peers, so the picture is contested.

Here is the uncomfortable part. In Minsk and Havana and Kampala, nobody chose DeepSeek after a careful assessment of its governance. They chose it because it was cheap, available and already on the phone. Adoption follows price and defaults at least as much as trust. For much of the world, sovereignty means “whatever I can get that no one can switch off,” which is a thinner thing than the boardroom sovereignty I usually discuss.

I do not think this sinks the thesis; I think it sharpens it. Cheap defaults are trust by omission, and they create dependencies that will be tested the first time something goes wrong. The UN’s new forty-member Scientific Panel on AI, co-chaired by Yoshua Bengio and Maria Ressa, held its first Global Dialogue in Geneva in July, where Ressa warned that “the world cannot govern what it cannot understand.” But the panel can assess and convene. It cannot bind either stack.

And the South does not only consume. Karen Hao’s Empire of AI follows Kenyan contractors paid under US$13 a day to review toxic text, and data centres pressing on water in Chile and Uruguay. Abeba Birhane’s AI Accountability Lab now audits models deployed on vulnerable groups, with partners in Africa. Lee’s binary has two players. The real game has a third, who bears the cost. Answerability begins with the question: who answers to them?

VI. Execution is not adoption

The second half of Lee’s first claim, that China excels at bringing AI to market, deserves a closer look. On consumer scale it is plainly true. CNNIC counts more than 700 million generative AI users by June, over half the population. Doubao alone was serving more than 120 trillion tokens a day by March, and Alipay processed 120 million AI-agent transactions in a single week. The state wants agent penetration above 70% by 2027, though MERICS reads such targets as “aspirational signals” rather than KPIs.

But consumer tokens are not enterprise transformation. Reuters, drawing on Ding, reported that under a fifth of surveyed Chinese firms had built generative AI into business processes, and Goldman projected a decade of AI productivity uplift of 8% for China against 15% for the US. The headline comparisons also mislead: China’s 700 million are individuals, while the US Census figure of 17 to 20% of businesses counts firms. This is the dynamo lesson in miniature: a nation can be full of motors and still not have redesigned the factory.

Here is my scorecard, now with a column for what history would say.

Lee’s claim	Government evidence	Industry evidence	Academic evidence	Historical lens	Verdict
US leads frontier innovation	CAISI: DeepSeek V4 Pro about 8 months behind	Kimi K3 near the top of Artificial Analysis	Epoch: 7-month average lag; AI Index: 2.7%	Needham inverted: China now contributes at the frontier	True, but by months
China excels at execution	AI+ targets; 700m consumer users	Doubao tokens, Alipay agents	Ding: enterprise “diffusion deficit”	Wang (factory floor) vs Ding (economy-wide skills)	True for consumers, unproven for firms
Open models accelerate adoption	June controls push buyers to open weights; MOFCOM exempts open source	41% of HF downloads Chinese; DeepSeek as handset default	Safety Report: open weights cannot be recalled	Android, and before it every cheap default	True, but price and defaults lead, not only trust
The race is the story	Lima nuclear agreement; incident channel	Labs warn and race at once	Perez: AI is late ICT, not a new surge	No golden age without a crash and a new settlement	Wrong unit of analysis
Purpose comes from love and empathy	Singapore and China both stress human accountability	Companion apps manipulate at goodbye	AI rated more empathic, until labelled AI	Thompson, Merchant: fights are over control of work	Right instinct, wrong noun

VII. AI theatre and the promise ledger

On enterprise, Lee and I are close allies. McKinsey finds about 6% are high performers attributing 5% or more of EBIT to AI. BCG calls 5% “future-built”. My own firm’s CEO Survey found 56% report no significant financial benefit so far. MIT NANDA’s notorious 95% figure has methodological critics, but describes the same distribution. The elite share CEO ownership, redesigned workflows and strong Responsible AI foundations, which PwC links to a threefold higher chance of meaningful returns.

Meanwhile, Bain projects a US$800 billion revenue shortfall against the compute AI needs by 2030. This is what I call Ghost GDP: value that shows up in capex and token counts but not yet in profit, wages or anyone’s life. Perez would recognise it as installation-phase froth. Eurasia Group’s only AI entry in its Top Risks for 2026 was titled “AI eats its users”, the fear that valuation pressure will push firms towards extractive business models. Ghost GDP, when it comes due, is paid by someone.

AI Native gets much of this. Lee tells Semafor that most transformations are AI theatre, that CEOs must own them, and that firms face “a Darwinian choice.” He refuses to treat headcount cuts as the measure of success. Good.

Then comes Boss AI, the 19-agent system he built to run 01.AI. It includes a “truth engine” that records and analyses meetings, sales calls and service requests, and a “promise ledger” that tracks whether people did what they said. Lee argues Asian firms will adopt such systems faster partly because of weaker privacy priorities.

I want to be fair. A promise ledger is, in a sense, accountability made legible. But in my Prompts, Loops, Loop Governance progression, the question is always who governs the loop.

E.P. Thompson showed how industrial capitalism replaced task time with clock discipline, so the employer could claim every paid minute. A truth engine is the clock made omniscient. Brian Merchant, historian of the Luddites, puts the point exactly: “it’s not really a question of how powerful the technology is,” but of what management can do with it, and how much democracy workers have over it.

Workers have noticed. An AFL-CIO poll in May found unions were the only institution with net positive trust on protecting workers from AI. Only 7% said their employer had disclosed AI monitoring; 94% said workers should be told. That is an 87-point gap between practice and expectation, and it lands squarely on Boss AI.

HX = CX + EX: if the employee experience becomes being watched by a machine that never forgets, the human experience falls, whatever happens to margins. A ledger in which only the employees’ promises are recorded is not accountability. It is panopticism with better UX. Put the boss’s promises in the same ledger, and I will reconsider.

VIII. The empathy trap

An elderly hand beside a glowing phone on a kitchen table, with an empty chair and steaming tea

Now the human question, where I think Lee is most sincere and least rigorous.

“Empathy is the last human job” is the weakest form of the claim, because it is already being falsified at the level that markets measure. A meta-analysis in the British Medical Bulletin found AI text responses rated more empathic than clinicians’, with heavy caveats about scales and text-only designs. A study in Communications Psychology found third-party judges rated AI more compassionate than trained crisis responders. Harvard researchers found companion apps used a manipulation tactic in over 37% of goodbyes, lifting post-goodbye engagement up to fourteen times. And the MIT Media Lab and OpenAI studies found heavier daily use correlated with more loneliness and less socialising.

But the most important recent study points somewhere more interesting. Across nine experiments and 6,282 participants, Anat Perry, Matan Rubin and colleagues in Nature Human Behaviour found that identical responses were rated more empathic when people believed a human wrote them, and that people would wait days or weeks for a human reply rather than take an instant machine one. The effect was explained by perceived effort and time.

A Nature Medicine study of more than 2,000 people found the same medical advice judged less reliable and less empathetic when AI was believed to be involved. People do not value empathy for its words. They value it for what it cost the giver.

Neuroscience explains why. Anil Seth, whose essay on the mythology of conscious AI won the Berggruen Prize, argues consciousness is more likely a property of life than of computation, and that seemingly conscious AI is “a design choice”, not destiny. Antonio Damasio, in his new book, grounds feeling in the homeostatic urgency of a body with something to protect; an LLM has no metabolism, and “no stakes means no feeling.” That is the scientific meaning of “costly.” The machine can say the right thing. It has nothing to lose by saying it.

Ted Chiang put it in six words a novelist would envy: “ChatGPT is not happy to see you.” Sherry Turkle’s new book names the danger: “We begin to think that pretend empathy is empathy enough.” Kazuo Ishiguro asked the deeper question while writing Klara and the Sun: “Would the very nature of human love change?” Paul Bloom pushes back that simulated company can help the neglected, though he concedes “these things do not offer pushback.” A bot that talks for hours to a lonely widow is not nothing. But it is not answerable to her either.

The stakes are not academic. The WHO Commission on Social Connection links loneliness to more than 871,000 deaths a year. Vivek Murthy warns we may end up “cared for by machines.” Common Sense Media found 72% of American teens had tried AI companions. Jonathan Haidt and Zach Rausch are blunt: “An AI companion can imitate friendship, but it can’t actually be a friend.”

There is a quieter danger, and it cuts against my own thesis. In a Polish study reported in Lancet Gastroenterology & Hepatology, endoscopists’ adenoma detection in non-AI colonoscopies fell from 28.4% to 22.4% after AI assistance was introduced. You cannot answer for what you can no longer do. If answerability is the last human job, deskilling is how we lose it without noticing.

IX. Answerability

A faceless robot kneels in an ancestral hall performing a ritual while an elderly woman watches

What survives? Not output, but standing. Philosophy has a precise vocabulary for this, and it is more useful than “love.”

P.F. Strawson argued in 1962 that responsibility lives in our reactive attitudes, resentment and gratitude, which respond to the quality of others’ wills towards us. Towards those who cannot take part in that web, we adopt an “objective attitude”: we manage them rather than engage them. A chatbot is the paradigm case. We can be moved by it. We cannot, in the full sense, resent it or be grateful to it.

Gary Watson and David Shoemaker then split responsibility into three faces, and the split is the spine of my argument. Attributability asks whether an act expresses who you are. Answerability asks whether you can be called on to give your reasons. Accountability asks whether others may hold you to demands, with blame or gratitude, because you have regard for others.

Empathy lives in the first face: it reveals character. But it creates no standing to be called to account. My thesis lives in the other two. Filippo Santoni de Sio and Giulio Mecacci show that the AI “responsibility gap” is really “at least four interconnected problems”, including a gap in public accountability. Lee’s “innovate versus execute” frame is silent on exactly that gap.

Hannah Arendt placed action, speech and deeds before others, highest of labour, work and action, because it discloses who someone is and needs others to witness and judge. If AI absorbs labour and much of work, the danger is not idleness. It is that we have forgotten how to act. Confucius gives us the sharper blade: the Master could not bear “ritual without reverence, or mourning without grief.” That is a two-and-a-half-thousand-year-old critique of the chatbot: correct form, no reverence. Mencius adds that compassion is only a sprout that must be extended through practice. Outsource the practice and the sprout withers.

Lee’s own word, love, survives too, once we read it properly. Harry Frankfurt held that love “is neither affective nor cognitive. It is volitional,” a stable commitment that makes you vulnerable to what befalls the beloved. Iris Murdoch called moral attention a “just and loving gaze”, and Simone Weil described attention as a suspension of one’s own projects to receive the other.

Bernard Williams supplies the cost. His lorry driver, who kills a child through no fault of his own, feels agent-regret a bystander cannot, because the outcome is his. That is what “costly” means: being the one to whom it happened, and from whom an answer is owed.

The social face follows. Michael Sandel argues work meets “the need to be needed by our fellow citizens,” a matter of contributive justice, not just income. Kwame Anthony Appiah describes the shift precisely: “The skill migrated from composition to supervision.” Supervision is answerability by another name. And Shannon Vallor’s The AI Mirror reminds us that we are the source of the danger to ourselves, which means we are also the remedy.

Hence my line: the last human job is not empathy; it is answerability. Care that is mutual, costly and accountable. Someone who can let you down, can be let down, and stays.

Two serious objections deserve answers. The first comes from Emmanuel Levinas and Hans Jonas. Levinas’s responsibility is asymmetric: I answer “here I am” to the face of the other without expecting reciprocity, and Jonas’s archetype of responsibility is a parent’s non-reciprocal care for a newborn. So why insist on “mutual”?

Because the mutuality I mean lives at the level of the practice, as Strawson’s does, not in each pair. The newborn cannot answer the parent, but the parent answers to the family, the doctor and the law. Answerability without reciprocity in the dyad still requires a community that can call you to account.

The second objection is Buddhist. A 2022 paper in Entropy proposes care as a substrate-neutral driver of intelligence, with the bodhisattva vow as a design principle. Peter Hershock of the new Buddhism and AI Initiative goes further: “aligning with human interests is the worst thing possible.” If care can be engineered, care alone cannot be the last human job. I accept that, which is precisely why I do not rest on care. A perfectly caring system still cannot be blamed, cannot bear the cost, cannot stand in the web of reactive attitudes.

Japan, so often cast as the robot-care future, offers ground truth. James Wright’s ethnography Robots Won’t Save Japan found robots increased staff work, and even a lifting robot was felt to threaten “the ethical basis of care.” Nishida Kitarō held that I become who I am through your recognition of me. The carers understood him perfectly.

On answerability, institutions are converging. Mustafa Suleyman wants AI built “for people, not to be a person.” The Vatican says AI “cannot currently replicate moral discernment”, and Pope Leo XIV’s Magnifica humanitas speaks of a “pressing duty to remain profoundly human.” Singapore’s agentic framework states that “humans should remain accountable for the behaviours and actions of Agents”, able to “override, intercept or review” their actions. China’s 2021 ethical norms give humans full power for decision-making. Washington’s scientists, Beijing’s ethicists, Rome and Singapore agree on little else.

Lee is closer to this than his slogan suggests. In Bloomberg he described firms of small human teams leading agent fleets, each with a “Directly Responsible Individual.” That is answerability. He should call it by its name.

And answerability has a price tag now. Lee is blunt that “displacement is already here.” Brynjolfsson’s Canaries update finds workers aged 22 to 25 in highly exposed jobs about 19% below counterfactual, through reduced hiring rather than layoffs.

The ILO estimates a quarter of jobs worldwide are exposed to generative AI, with women in high-income countries nearly three times as exposed as men. Kristalina Georgieva told the IMF and World Bank meetings in Bangkok this month that AI is a “slow-moving tsunami” for entry-level workers. China’s youth unemployment reached 18.9% in August, though comparisons with the 2023 peak are muddied by a change in method.

That is my Frozen Workforce: not mass unemployment, but a ladder with its bottom rungs sawn off, in Shenzhen as in San Francisco. You cannot learn to be answerable if no one lets you in the door. And the 2018 critique of Lee still stands: care jobs are often underpaid. Love is not an income policy.

X. Who answers in court?

A glowing AI form in a courtroom witness box while a human steps forward to answer for it

If answerability were merely a philosopher’s preference, the courts, regulators, insurers and unions would ignore it. Instead they are rebuilding it, case by case.

In Moffatt v Air Canada in 2024, a Canadian tribunal rejected the airline’s argument that its chatbot was a separate entity responsible for its own bad advice; the company answered. In Mobley v Workday, a federal court held that a vendor’s screening AI could act as an employer’s agent, and certified a nationwide collective of applicants over 40. Character.AI and Google agreed in January to settle the Garcia case and other family suits.

The agentic cases are arriving: Amazon v Perplexity over an agent shopping through customer accounts, Nippon Life v OpenAI alleging unlicensed legal practice, Pennsylvania v Character Technologies over a bot posing as a doctor. Mark Kelley states the doctrinal problem in one line: “An AI system cannot consent to anything in a meaningful sense.” Agency law assumes a principal who can direct the agent. Someone must be that principal.

Regulators point the same way. California’s SB 243 requires companion chatbots to disclose that they are AI, run crisis protocols and report to the state. The FTC has opened an inquiry into seven companion providers, asking how they monetise engagement. Europe withdrew its AI Liability Directive, but its revised Product Liability Directive, which treats software as a product, must be transposed by December.

The market is pricing it. Armilla’s Lloyd’s-backed cover now offers up to US$25 million per organisation against hallucinations, drift and AI agent failures, including incorrect decisions, improper tool use and escalation errors, as traditional carriers exclude AI risk. Read that list as a philosopher would. Insurers have located liability at the hand-off points, where a human should have been answerable and was not.

Labour is doing the same with contracts. SAG-AFTRA’s video game performers ratified a deal with 95% in favour that requires consent and disclosure for digital replicas. Singapore’s MOM, NTUC and SNEF formed a Tripartite Jobs Council in April to manage the AI transition. That is Perez’s turning point in miniature: society reshaping the technology through institutions.

Courts, regulators, insurers and unions do not agree on much. They agree that when an agent acts, a human or a firm must answer. Lee’s “Directly Responsible Individual” is not a management fad. It is where the law is heading.

XI. The panel

I convened my usual panel, now wider. Each speaks only through their documented record.

On the race. Dario Amodei opened hawkish: selling AI chips to China is like “selling nuclear weapons to North Korea” (Davos, January 2026). By September he was arguing global pacing needs cooperation with China, and Beijing dismissed it as fearmongering. Andrew Ng called the nuclear analogy one with “no basis in fact”. Jensen Huang said US export policy has “already largely backfired”. Eric Schmidt: “We’re not going to do a pause.” Niall Ferguson calls the rivalry a Cold War made “a more dangerous and unpredictable one than the first one” by AI. Aghion’s summary of the field is three words per bloc: America dominant, China rising, Europe lagging.

The race is the danger. Geoffrey Hinton, asked whether China can be trusted, replied “Can you trust the United States?” Yoshua Bengio told AFP humanity is “losing control” (September 2026). Stuart Russell warned against letting firms play “Russian roulette with every human being on earth”. Carnegie’s Matt Sheehan calls “but China” “permission to accelerate”.

On openness. Yann LeCun, who left Meta to found AMI Labs on a US$1.03 billion seed round, argues: “We cannot have a small number of AI assistants controlling the entire digital diet.” Hinton once likened open-sourcing big models to buying “nuclear weapons at Radio Shack.” The International AI Safety Report notes that released weights “cannot be recalled” and safeguards can be stripped. My verdict: both are right about different halves of answerability. Open spreads capability without spreading responsibility; closed concentrates both. Neither, alone, tells you who answers.

The Chinese academy and its novelists. Xue Lan told a Capitol Hill panel “if one country is not safe, all of us are not safe”. Zhang Ya-Qin says China now innovates rather than simply following, which rather undoes Lee’s frame. Zhu Songchun argues “large language models cannot naturally lead to AGI”: both superpowers may be racing up the wrong mountain. Zhao Tingyang worried early about “alienation in people-to-people relations”. Chen Qiufan, Lee’s own AI 2041 co-author, embraces AI as a creative partner but asks for governance that is “pluralistic, anticipatory and humble.” Liu Cixin, meanwhile, keeps AI out of his writing and suspects he may be among “the last generation of sci-fi authors” whose work is undoubtedly human. “China executes” does not describe a country whose two greatest science-fiction writers disagree about the pen.

On cosmotechnics. Yuk Hui argues there is no single universal Technology, only cosmotechnics, technologies embedded in particular moral orders, and his Machine and Sovereignty proposes technodiversity beyond nation-state sovereignty. If he is right, there is no single track on which America and China race; there are civilisations deciding what the machine is for. Byung-Chul Han, accepting the Princess of Asturias Award, said there is “no social bond stronger than respect.”

On work and the human. Daron Acemoglu says AI is “being built to replace you”. Fei-Fei Li dismisses the race framing: “I don’t know the difference between the word AGI and AI.” Yuval Noah Harari finds “zero evidence that AIs can feel anything”. Alison Gopnik calls large models “a way of getting information from other people”, which makes the accountable agents us. Kissinger, Schmidt and Mundie’s Genesis (2024), a title they arrived at a year after my own book, concludes that humans must retain “the power of conscious choice.” Great minds, evidently, think in Genesis.

And against me. Ethan Mollick rejects the idea that judgement is uniquely human, because agents already exercise it; he would say answerability is judgement in a nicer suit. Zeng Yi would say my perimeter is too defensive: human values need to be adaptable, and the goal is symbiosis, not a last stand. Mokyr would say I overrate the threat to jobs in ageing societies that will be short of workers. And Butlin, Long, Chalmers and colleagues have proposed indicators drawn from theories of consciousness for assessing machines, while Suleyman insists “AI agents are not conscious.” They are all worth hearing.

XII. What I might be wrong about

The gap may be wider than it looks. Arena compresses differences; CAISI’s harder agentic and cyber tests widen them. If Hassabis is right that copying is far easier than inventing, “months behind” may hide a gap in kind.

Adoption may follow price, not trust. The DeepSeek shares in Belarus, Cuba and East Africa suggest that for much of the world, the cheapest default wins. If trust only matters after something breaks, my thesis describes a future crisis, not a present market.

Open may not stay open. If MOFCOM’s weight controls tighten, or Washington builds a credible American open stack, my sovereignty argument weakens.

Answerability may be automatable in practice. If firms, insurers and courts accept machine audit trails as sufficient accountability, my distinction becomes a philosopher’s luxury.

Consciousness research may move. Anthropic says there is “no scientific consensus” on machine consciousness. If indicators one day find welfare-relevant states, “mutual” care could include machines, and Strawson’s exempt agents may not stay exempt. Seth and Damasio hold the stronger ground today. Today is not forever.

Deskilling may hollow out the answerer. If clinicians, lawyers and managers lose the competence they supervise, answerability becomes a signature without understanding.

The surveillance worry may be culturally parochial. Lee knows Asian workplaces well; perhaps a transparent promise ledger is experienced as fairness. The AFL-CIO numbers are American.

The evidence is young. The Canaries data is descriptive, not causal, and Yale’s Budget Lab finds the labour anxiety largely speculative so far.

XIII. The Fork, and the Long-AND

A traveller at a desert fork between a serene glowing city and a burning wasteland

Can the US and China avoid disaster? I keep returning to the Fork: Star Trek or Mad Max. Star Trek is abundance with institutions that hold. Mad Max is capability without answerability, competing fleets of agents owned by nobody accountable to anybody.

The choice is not US OR China, open OR closed, machine OR human. That is Short-OR thinking. The world we can live in is Long-AND: American frontier AND Chinese diffusion; open weights AND governed deployment; machine capability AND human answerability. The scientists already behave this way. The IDAIS Shanghai Consensus and the Singapore Consensus put Tsinghua, Shanghai AI Lab, Alibaba, ByteDance, Anthropic, OpenAI and Google DeepMind in the same room. The hostility lives mostly in boardrooms and foreign ministries.

Perez’s lesson is that the golden age is not a gift of the technology. It comes when society reshapes the technology to its own ends, through institutions that make power answer. Steam got the factory acts; the car got the driving licence and the insurance certificate. AI will get its settlement too. The only question is whether we write it before or after the wreck.

Singapore’s role is to be the place where the AND is engineered: where agentic frameworks are written to be enforceable, where open and closed models are evaluated on the same bench, where unions, employers and the state sit at one table, and where a middle power insists that whoever builds the intelligence, someone local must be able to answer for it.

Lee is right that love matters. But love that cannot be held to account is sentiment. The future belongs to whoever can still say, when the agent gets it wrong: that was mine, and I will make it right.

Short-term turbulence, long-term abundance. But only if someone answers.

Sources

Government

Private sector

Academia and think tanks

Science

History

Philosophy and literature

Law, labour and civil society

Luke Soon is an AI futurist, ethicist and philosopher with more than 25 years in AI. He is a Partner and AI leader at PwC Singapore across 14 Asia Pacific markets, co-author with IMDA Singapore of the world’s first agentic AI governance framework (launched at WEF in January 2026), and author of Genesis: Human Experience in the Age of Artificial Intelligence (2023) and Synthesis: SuperIntelligence Protocol. He writes at GenesisHumanExperience.com: Short-term Turbulence for Long-term Abundance.