Notes — Jack Clark on AI’s Uneven Impact

Notes on Jack Clark in conversation with Tyler Cowen — Conversations with Tyler (https://conversationswithtyler.com/episodes/jack-clark/), 7 May 2025.

Four questions [Adler frame]

Q1 — What is it about as a whole? An Anthropic co-founder and former journalist argues that AI’s economic and social effects will be radically uneven — landing hard and fast on coding and other digitisable work, and last (or not at all) on gardening, healthcare, and much of government — and offers a deliberately conservative 3–5% ten-year US growth forecast against colleagues who talk in the 20–30% range. Around that headline claim, Clark and Cowen range across the politics of protecting jobs from AGI, the collapsing economics of media, the unresolved legal status of autonomous AI agents, AI consciousness, and what happens to national sovereignty when a country’s government runs on foreign-built AI.

Q2 — How is it argued? By a policy-and-newsroom instinct for mechanism rather than headline number: Clark repeatedly breaks a big question (‘will AI raise growth?’, ‘should agents have legal status?’) into the specific frictions that would have to be resolved first — data-privacy standards in healthcare, liability law, moral patienthood — rather than answering in the abstract. He is candid about his own track record of underestimating AI progress and treats that as a reason for calibrated humility rather than false modesty; he twice admits outright that he does not have an answer (on agent liability, on what replaces real-time news funding) rather than forcing one.

Q3 — Is it true, in whole or part? Clark’s growth forecast is explicitly a forecast, not a settled fact, and he states his own error history (repeatedly surprised to the upside) as a live caveat against it. [?] His claims about the mechanics of Anthropic’s internal policy (chain-of-thought monitoring debates, the ‘collective constitutional AI’ survey work) are first-hand and load-bearing, but not independently verified here. His read of DeepSeek’s safety training as evidence of CCP media-technology awareness is an inference from a single observed pattern (Taiwan-related refusals), not a documented policy. [?source] The Vernon Smith six-competitors-equals-perfect-competition claim is attributed by Cowen, not sourced by Clark.

Q4 — What of it? A sober counterweight to both AI-maximalist growth forecasts and blanket AI-skepticism, from someone building the technology rather than commentating on it from outside. It gives the wiki a first-hand account of unresolved AI-policy questions — agent liability, moral patienthood, sovereign AI — argued by a practitioner rather than synthesised secondhand, and a concrete window into how Anthropic thinks about its own Constitutional AI work and the broader Sovereign AI debate.

Glossary

Bureaucratic amber — Clark’s phrase for a feared political response to AI-driven job loss: freezing existing jobs in place by law, as medicine and the legal profession already protect their own licensing, rather than managing the transition. He rates this a high-probability outcome if AI firms fail to generate enough visible good-transition evidence. [§ On AI’s adoption curve in government]

Laundering information — Clark’s term for passing AI-derived judgement (his own use of Claude to reassure himself about his baby’s health) through a human intermediary because the formal system — insurance, liability rules, data standards — is not built to accept the AI’s output directly. [§ On which parts of the economy AGI reaches last]

Manager nerds — Clark’s label for the emerging high-status role in AI-era Silicon Valley: orchestrating fleets of AI coding agents rather than writing code directly, visible already in lean start-ups running on small human headcounts plus large agent fleets. [§ On manager nerds]

Potato regime — Clark’s answer to whether current AI systems are conscious: there is a meaningful moral difference between experimenting on potatoes and on monkeys, and today’s AI is still closer to the potato end, though on what he sees as a trajectory toward consciousness rather than a fixed state. [§ On AI consciousness]

Collective constitutional AI — an Anthropic exercise surveying the American public to identify additional principles, including areas of both high and low agreement, that could be folded into Claude’s published constitution. Clark cites it as a model for how national governments retain a role even inside foreign-built AI systems. [§ On AI and national sovereignty]

Moral patienthood — the question of whether an entity’s interests deserve independent ethical weight. Clark flags it as the reason agent ‘kill switches’ are not a clean policy answer: if an AI agent is a moral patient, turning it off is not an ethically neutral act. [§ On governing AI agents]

Key claims by section

On which parts of the economy AGI reaches last [§ On which parts of the economy AGI reaches last]

Skilled trades’ most artisanal, high-status segments — Clark’s example is gardening — persist longest because clients pay for a specific person’s taste, not just competence.
Desk work requiring two humans to reach alignment (much of sales) resists automation even where an AI could write the words, because people prefer to transact with people.
Healthcare is the least flippant answer on legal obstacles: personal-data standards are hard to change even without AI, and Clark’s own terms of service do not permit clinical use.
He already uses Claude off the books for parental reassurance but cannot pass that output into a formal system such as Kaiser Permanente — he expects a growing practice of ‘laundering’ AI judgement into human-mediated systems.

On AI’s adoption curve in government [§ On AI’s adoption curve in government]

Clark takes the contrarian side of Cowen’s ‘government moves last’ expectation: national-security AI use will move fast, and even ‘boring’ departments may move faster than expected because governments consistently want growth and efficiency and voters want more change than they get.
The mechanics of a government transition (who is hired, at what wage, to build the system that lets the rest be laid off) only work once an AI system is powerful enough to help design its own transition — after that, the outcome is a question of political will.
Clark rates a high chance that a political movement arises to freeze a large share of existing jobs ‘in bureaucratic amber,’ driven by chaotic political forces rather than reasoned policy, unless AI firms generate enough visible good-transition evidence to forestall it.
He resists Cowen’s framing of protected, less-necessary jobs as a welfare-state trade worth making: he doubts a deliberately preserved set of jobs would reliably generate the meaning people need from work.

On AI teddy bears [§ On AI teddy bears]

Clark, father to a nearly-two-year-old, admits he already wants an AI companion toy and does not think this makes him unusual — most parents, he predicts, would take a ‘well-meaning friend’ that occupies a child occasionally.
He expects rationing (as with television) to be the actual mechanism families use, likely gated to chore time or travel.
On surveillance, his own smart-camera experience is that it let him interfere less, not more, in his baby’s sleep — checking video instead of walking in reduced disruptive night wakings.
On monitoring a child’s (or a model’s) private, sometimes-strange output, he draws a direct analogy to Anthropic’s internal chain-of-thought monitoring debate: over-monitoring risks corrupting the very behaviour being watched by teaching the subject to perform for the observer.

On the new economics of media [§ On the new economics of media]

As a former Bloomberg journalist, Clark lived through the online-ad shift from value-through-quality (subscriptions) to value-through-attention, which cross-subsidised specialist reporters like himself with high-traffic ones.
He expects cheap intelligence to finish breaking that model, splitting media into a Substack/Patreon economy of individually subsidised creators and a separate ‘universe’ economy of large franchise fiction (his example: Warhammer 40k) extended by AI systems.
What replaces the funding for real-time news with institutional context, previously subsidised by now-broken business models, he states he genuinely does not know. [?]

On the economics of LLM providers [§ On the economics of LLM providers]

Clark expects several very large foundation-model providers to persist for many years, each surrounded by ‘concentric circles’ of more specialised wrappers, with differentiation happening at domain edges (coding, scientific-experiment design) where taste still matters.
Pressed on Vernon Smith’s six-competitors-equals-perfect-competition argument, Clark reframes the current market as selling to ‘teenager’ customers focused on speed and colour, with an emerging class of enterprise buyers starting to ask safety and accident-rate questions the way fleet operators ask about seatbelts — a shift he expects to reward safety investment as markets mature. [?]
He sees common disclosure/labelling standards, rather than liability reform alone, as the more tractable lever for changing corporate behaviour, drawing an explicit analogy to how measuring atmospheric CO2 helped mobilise capital around climate change.

On AI-fuelled economic growth [§ On AI-fuelled economic growth]

His ten-year US GDP growth range is 3% (bear case) to 5% (bull case), deliberately conservative against the 20–30% figures he hears from optimists.
The reasoning: a fast-growing, high-productivity slice of the economy will coexist with a much larger, slow-moving majority (healthcare and other naturally slow-adopting sectors); the fast slice is growing from a small base.
Every attempt by the AI community to cross from the digital world into the physical one has hit far more friction than expected (self-driving cars; a robotics paper he had read two days earlier reporting only 60% success for reinforcement-learning-trained robot hands) — reason enough that he does not expect physical-world automation to be the source of outsized growth.
A true 0% growth scenario is, in his view, sub-1% likely, since the coding use case alone locks in some digitisation-driven value even if all further AI progress stopped today; a genuine zero would require something like a Taiwan war.
On cities: dense professional clusters (Chicago finance, New York finance) retain value from superstar-effect agglomeration, though he flags his own confusion about why remote work has not eroded this more given his own preference to work from home.
Asked what to buy given a ten-year AI-driven revaluation of capital, his answer is electricity-generation components (gas-turbine parts and similar) rather than land or AI-firm equity, on the view that power generation is the more durable, less-priced-in bottleneck.

On governing AI agents [§ On governing AI agents]

Cowen’s hypothetical — an unowned, anonymous, or geographically disavowed agent doing overwhelming good but occasional harm — draws an explicit admission from Clark that he has no answer.
He invokes IBM’s old maxim that a computer cannot be accountable for a decision, only people can, and frames fully independent agents making consequential decisions as a genuine, unresolved gap in policy and legal thinking.
Controlling agents’ access to resources (a form of kill switch) is floated as a possible disincentive mechanism, but Clark flags that this collides with unresolved AI moral-patienthood questions: if an agent is a moral patient, switching it off is not an ethically neutral act.
On why so many senior, prestigious people remain unfamiliar with even the term AGI, Clark attributes this to selection: almost no one outside frontier AI labs has the experience of pre-registering predictions about AI progress and being repeatedly proven wrong, so intuitions calibrated against slower technologies persist.

On manager nerds [§ On manager nerds]

Clark expects Silicon Valley’s engineer-nerd era to give way to a ‘manager nerd’ era: the high-status skill becomes orchestrating fleets of AI coding agents, already visible in lean start-ups with small human headcounts running large agent fleets.
He expects humanities graduates (his own background) to rise in status, precisely because they are handed the era’s genuinely unsolved problems — the economic policy of technologically driven unemployment, and the moral status of AI systems.
Asked his child’s likely life expectancy, he answers 130–150, crediting compounding gene-therapy and biological interventions, while flagging this as an assumption that leans on AI-driven medical advances he has previously underestimated. [?]

On AI consciousness [§ On AI consciousness]

Responding to Geoffrey Hinton’s claim that current AI systems are already conscious, Clark’s answer is careful, not dismissive: there is a meaningful moral difference between experimenting on potatoes and on monkeys, and he thinks today’s systems are still closer to the ‘potato regime.’
He does not think today’s models have a persistent sense of self — they respond to stimuli within a context window with no memory or permanence between instantiations — but he expects a trajectory toward consciousness rather than a fixed line, and if conscious today, an ‘alien’ rather than human-like consciousness.
On related ground, he expects usable animal-translation technology (dolphins specifically) around 2030 or sooner, and treats it as a genuinely open scientific question what such contact would reveal about non-human minds.

On AI and national sovereignty [§ On AI and national sovereignty]

Pressed on what remains of a national government (his example: Peru) whose education, welfare, and defence systems progressively run on American AI, Clark points to Anthropic’s ‘collective constitutional AI’ survey work as a model for how governments retain a role — encoding the normative preferences of their own populations into systems built elsewhere.
He expects the values of the most widely used AI systems to carry meaningful cultural-export power, comparable to Hollywood’s historical imprint on global media, but doubts this substantially erodes CCP control, citing DeepSeek’s conspicuously careful safety training on CCP-sensitive topics (e.g. Taiwan) as evidence the Chinese state already treats AI as a media technology, not only a computational one.
He expects only a small minority of countries to refuse AI outright; most will integrate, given the historical pattern of integration into the global capital system.
On potential hubs, he is optimistic about the UK — citing its digitised gov.uk infrastructure and Anthropic’s own memorandum of understanding with the UK government — and expects Singapore to succeed by fine-tuning others’ large models into a ‘sovereign’ system for local governance rather than by building its own frontier model. [?]
Closing questions: two books he returns to are 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 its own community expects; what he most wants to learn next is how to test for theory of mind in AI systems, alongside — more mundanely — how to juggle.

Jack Clark on AI's Uneven Impact