Jack Clark on AI’s Uneven Impact

Anthropic co-founder Jack Clark talks with Tyler Cowen about why AI’s economic effects will land unevenly rather than everywhere at once — hitting coding hard and gardening, healthcare, and much of government last — and lays out a deliberately conservative 3–5% growth forecast against colleagues who talk of 20–30%.

Key ideas

1. Growth will be bimodal, not uniform. Clark expects a fast-growing, high-productivity sliver of the economy (coding, and other digitisable work) alongside a much larger, slow-moving majority (healthcare, most government functions, most jobs) that adopts AI late. His ten-year US growth range is 3% (bear case) to 5% (bull case) — deliberately below the 20–30% figures he hears from optimists, because every attempt to cross from the digital world into the physical one has run into far more friction than expected.
2. A political movement to freeze jobs ‘in bureaucratic amber’ is a live risk. If AI companies and their customers fail to generate enough visible examples of good transitions, Clark expects rising pressure to legally protect existing jobs from displacement — a response he calls understandable but likely driven by ‘chaotic winds of political forces’ rather than reasoned policy.
3. Media’s ad-funded cross-subsidy model cannot survive cheap intelligence. The old system — high-traffic writers subsidising specialist ones — is already broken; Clark expects the field to split into a Substack/Patreon-style economy of individually subsidised creators and a separate ‘universe’ economy of large franchise fiction extended by AI. What replaces the funding for real-time news reporting, he says plainly, is unclear.
4. Autonomous AI agents raise a legal-accountability problem with no answer yet. Clark repeatedly ‘dodges’ Cowen’s questions on how law should treat unowned, independent agents, invoking IBM’s old maxim that a computer cannot be held accountable — only people can — while noting that any move to disincentivise agents by controlling their resources runs into unresolved questions of AI moral patienthood.
5. Silicon Valley shifts from the nerd’s era to the ‘manager nerd’s.’ Clark expects the high-status skill to become orchestrating fleets of AI coding agents rather than writing code directly, and expects humanities graduates — his own background — to rise in status because they get handed the era’s genuinely unsolved problems: technologically driven unemployment, AI moral status, and what values get built into these systems.

Content

Which parts of the economy AGI reaches last

Cowen opens by asking where strong AI’s effects arrive last. Clark’s first answer is the high-status, artisanal end of skilled trades — gardening above all — where clients pay for a specific person’s taste, the way collectors pay for an artist who orchestrates a team rather than for the artist’s own hands. In desk work, he points to anything requiring two people to reach alignment: Claude could write excellent sales copy, but ‘we don’t send Claude to sell Claude’ because people still want to transact with other people, especially when deploying large pools of capital through human proxies. On legal obstacles, his flippant answer is the legal profession’s own resistance to being undercut; his real answer is healthcare, which is bound up in personal-data standards that have proven extraordinarily hard to change even without AI in the picture. He already uses Claude off the books — reassuring himself about his own toddler’s bumped head before calling an advice nurse — but cannot pass that assessment on to a formal healthcare provider such as Kaiser Permanente. He expects a growing practice of ‘laundering’ AI-derived judgement into human systems not built to accept it directly.

Government, political will, and the case for protecting jobs

Cowen expects the US federal government — still running decades-old software in places — to be the slowest adopter. Clark takes the other side: national-security-adjacent AI use will move fast, and he suspects the ‘boring’ parts of government could follow faster than expected too, because governments consistently say they want growth and efficiency, and because voters want more change from government than they are getting. The harder problem is transition mechanics — Cowen presses on who gets hired, and at what wage, to build a system capable of laying off the rest of a department’s staff — and Clark concedes this only works once an AI system is powerful enough to help design its own transition, after which the outcome depends on ‘political will.’ Asked whether some future society might legally protect half of today’s jobs against AGI the way medicine and law protect professional licensing today, Clark rates the odds high: a political movement to freeze jobs ‘in bureaucratic amber’ is his central worry, arising less from reasoned policy than from the absence of enough visible good-transition evidence to counter it. He resists Cowen’s more sanguine framing — that guaranteed, unnecessary jobs might function as a kind of welfare state for meaning — on the grounds that people need work that plausibly generates its own meaning, and he is not confident any deliberately preserved category of jobs would do that.

AI teddy bears and the ethics of care

The conversation’s most personal stretch concerns AI companions for children — Cowen’s ‘AI teddy bears.’ Clark, father to a child not yet two, admits he already wants one: something to occupy a toddler while a parent cooks dinner, the way television has always been rationed for exactly that purpose. He does not think this makes him an outlier, and predicts most parents would take the trade if it meant a genuinely ‘well-meaning friend’ for occasional use. Pressed on the harder edge — a child preferring the bunny to human friends, or surveillance of a child’s private, often nonsensical chatter — Clark draws an analogy to a live debate inside Anthropic about whether to monitor an AI model’s chain of thought: over-monitoring risks teaching the system (or the child) to produce only ‘safe to be seen’ output, corrupting the very thing being observed. Some things, he argues, need protected space for unmonitored creativity, whether the subject is a seven-year-old or a reasoning model.

The economics of media, LLM competition, and growth forecasts

On journalism, Clark — a former Bloomberg reporter — describes the online-advertising shift that already broke value-through-quality in favour of value-through-attention, cross-subsidising specialist reporters like himself with viral ones. He expects intelligence becoming cheap to finish the job: what survives is a Substack/Patreon economy of individually subsidised creators, plus a separate ‘universe’ economy — large franchise fiction such as Warhammer 40k, extended and monetised by AI systems. What happens to real-time news with real institutional context, subsidised historically by business models that no longer work, he says he ‘genuinely does not know.’ On market structure, Clark expects several very large foundation-model providers persisting for years, each ringed by ‘concentric circles’ of specialised wrappers, with differentiation happening at the edges — in domains like coding or scientific-experiment design where taste still matters. Pressed on Vernon Smith’s argument that six-or-more competitors behave like perfect competition and squeeze out room for above-market ethical behaviour, Clark reframes the industry as still selling to ‘teenagers’ asking about speed and colour, with a newer class of enterprise buyers starting to ask about safety and accident rates the way fleet operators ask about a car’s seatbelts — a shift he expects to reward safety investment as markets mature. On growth, his forecast is 3% (bear) to 5% (bull) over ten years, well below headline figures of 20–30%, because the digitally-native, fast-growing slice of the economy is starting from a small base, and because AI’s repeated attempts to cross from digital to physical work — self-driving cars, robot manipulation — keep running into far more friction than anticipated. He cites a robotics paper he had read two days earlier reporting a 60% success rate for reinforcement-learning-trained robot hands, well short of what any parent would accept from a robot butler.

Governing autonomous AI agents

Cowen poses a genuinely open legal question: how should law treat an AI agent that is unowned, anonymous, or deliberately disavowed by its creator, deployed somewhere with weak enforcement, doing overwhelming good but occasionally causing harm? Clark’s honest answer is that he does not have one. He reaches for IBM’s old warning that a computer cannot be held accountable for a decision — only people can — and notes that fully independent agents making consequential decisions therefore create a real gap in policy and legal thinking. He is sympathetic to Cowen’s instinct that liability should not trace back indefinitely to the original model-builder, and floats controlling agents’ access to resources as a possible disincentive mechanism, akin to a kill switch — but flags that this collides with unresolved questions of AI moral patienthood: if an agent is a moral patient, turning it off is not obviously a neutral act.

Manager nerds, AI consciousness, and status

Clark predicts the era of the engineer-nerd gives way to the ‘manager nerd’ — someone whose value lies in orchestrating fleets of AI coding agents rather than writing code by hand, already visible in start-ups running lean staff counts on top of large agent fleets. He expects humanities graduates, his own background, to rise in status precisely because they are handed the era’s unsolved problems: the economic policy of technologically driven unemployment, and the moral status of AI systems. On Geoffrey Hinton’s claim that current AI systems are already conscious, Clark’s answer is careful rather than dismissive: there is a meaningful difference between running experiments on potatoes and on monkeys, and he thinks today’s systems are still in the ‘potato regime’ — responsive to stimuli in complex ways, without an apparent sense of self, and without persistent memory across a conversation. He expects a trajectory toward consciousness rather than a bright line, and if these systems are conscious today, he expects it to be an ‘alien’ consciousness, not a human-like one.

National sovereignty, AI hubs, and closing reflections

Asked what becomes of a smaller national government whose education, welfare, and defence systems are progressively run on American AI, Clark points to Anthropic’s own ‘collective constitutional AI’ work — surveying Americans for additional principles to fold into Claude’s constitution — as a model for how governments retain a role: encoding the normative preferences of their own populations even as the underlying systems are foreign-built. He expects the values of the most widely used AI systems to carry real cultural-export power (as Hollywood once did), but doubts this meaningfully erodes the Chinese Communist Party’s control, pointing to DeepSeek’s conspicuously careful safety training around CCP-sensitive topics as evidence the state is alert to AI as a media technology, not only a computational one — part of a broader Sovereign AI question that Clark expects most countries, bar a small minority, to resolve by integration rather than refusal. On where AI hubs might emerge, he is optimistic about the UK, citing its already-digitised gov.uk infrastructure and Anthropic’s own memorandum of understanding with the UK government, and expects Singapore to succeed less by building its own frontier models than by fine-tuning others’ into a ‘sovereign’ system suited to local governance. Closing on lighter ground, Clark names two books he returns to — qntm’s There Is No Antimemetics Division and Fernand Braudel’s Capitalism and Material Life, the latter shaping his scepticism that AI will change ordinary material life as fast as enthusiasts expect — and says what he most wants to learn next is how to test for theory of mind in AI systems, alongside, more mundanely, how to juggle.

See also

• Jack Clark — speaker
• Tyler Cowen — host
• Constitutional AI — Anthropic’s alignment technique, discussed here in its ‘collective’ population-survey form
• Sovereign AI — national AI-independence thesis, discussed here via Peru, the UK, and Singapore
• Consciousness — philosophical background to Clark’s ‘potato regime’ answer on AI consciousness