I have spent some time listening to the two sides of the AI (and the future of jobs) forecasting debate. From where I sit, Tyler Cowen has the more optimistic take. On the other hand, you have the AI Futures folks. I read their AI 2027 last summer, and from my recollection, its forecast was directionally strong, but too fast, I think. They recently came out with this new report, AI 2040, which I have not yet read. But I did listen to what I believe is the gist of it in a two-hour podcast the other day.
As I listened, I wondered what Tyler Cowen might say in response to Daniel Kokotajlo’s arguments. So, I assembled a prompt for GPT-5.6 Sol, as one does these days, asking it to stage a debate between Kokotajlo and Cowen using their latest available materials and content, something light I could read after a hard day of work. The result, which you can read below is quite good. And I am deliberately presenting what GPT 5.6 sol gave me back, largely untouched, because 1) there is not much I could add given the time constraint of writing this essay 2) I believe that the “medium” itself is the message: The synthesis itself makes a case about the subject matter.
The Crux of it is, or at least part of the crux of whole debate, is that Kokotajlo assumes that AI capability growth may outrun institutional adaptation, while Cowen assumes that institutional friction, expanding demand and human preference will remain powerful even after capability becomes extraordinary. My reading of it is that Cowen offers the stronger model of the next several years; Kokotajlo identifies the regime change that becomes decisive if autonomous AI research and general-purpose robotics advance rapidly.
Here are the References synthesized into the debate:
So you can read (or watch) the source materials for yourself.
Cowen, T. (2026, July 8). Human life in a post-AGI world. Google DeepMind. https://tylercowen.com/human-life-in-a-post-agi-world-talk/
Cowen, T. (2026, June 28). My ARC talk on AI and jobs. Marginal Revolution. https://marginalrevolution.com/marginalrevolution/2026/06/my-arc-talk-on-ai-and-jobs.html
Cowen, T. (2026, May 13). Some non-obvious reasons why AI will create some transitional problems in employment. Marginal Revolution. https://marginalrevolution.com/marginalrevolution/2026/05/some-non-obvious-reasons-why-ai-will-create-some-transitional-problems-in-employment.html
Roberts, R. (Host). (2026, March 30). AI, employment, and education (with Tyler Cowen). In EconTalk. Econlib. https://www.econtalk.org/ai-employment-and-education-with-tyler-cowen/
Pangambam, S. (2026, July 13). Transcript of Daniel Kokotajlo interview: Diary of a CEO podcast. The Singju Post. https://singjupost.com/transcript-of-daniel-kokotajlo-interview-diary-of-a-ceo-podcast/
The Diary Of A CEO. (2026, July 13). He risked everything to warn you: No one is ready for what’s coming, and the AI companies know it! YouTube.
Veld, S. (2026, January 11). What happens when superhuman AIs compete for control? AI Futures Project.
*** GPT 5.6 Sol Synthesis** *
Moderator: By the 2030s, will most people still have jobs?
Daniel Kokotajlo: Not if companies successfully build superintelligence. The usual response—“technology destroys some jobs but creates new ones”—works only while humans retain capabilities machines lack. Once AI is better, faster, and cheaper than humans across essentially every cognitive task, it can perform the newly created jobs too. Add sufficiently capable robotics, and eventually almost all economically useful human labour becomes technically replaceable.
Tyler Cowen: That moves too quickly from technically possible to economically and institutionally implemented. Economies are not benchmark suites. Firms, governments, laws, buildings, energy systems, professional norms and consumer preferences change slowly. Electricity took decades to diffuse fully. Even very capable AI must pass through human organisations, and those organisations are full of bottlenecks. I expect substantial disruption and more leisure, but not the disappearance of employment.
Round 1: Does comparative advantage save human labour?
Cowen: Even when one party is better at everything, comparative advantage can preserve trade and specialisation. Humans will remain useful where our participation is legally required, socially preferred or tied to trust, legitimacy and personal meaning. People will pay to see a human musician, talk to a human political representative, receive human reassurance or experience something created by a particular human life.
There will also be enormous new sectors: gathering real-world data, conducting physical and biomedical experiments, integrating AI into institutions and helping different societies adopt it on their own terms. Humans may use AI throughout these jobs, but that does not mean the humans vanish.
Kokotajlo: Comparative advantage does not guarantee that wages remain high enough to sustain a mass labour market. Horses technically retained comparative advantages after automobiles arrived; their economic role nevertheless collapsed.
Your examples are plausible transitional jobs. But an AI that can design experiments, operate robots, persuade people, administer organisations and rapidly improve the relevant technology will progressively absorb them. Eventually the remaining human jobs exist because society politically reserves them for humans—not because humans are economically necessary.
Cowen: But “politically reserved” is not a trivial exception. Politics, liability, status and preference are part of economic demand. We do not consume only the cheapest technically adequate output. We consume stories, identities, authenticity and relationships.
Kokotajlo: Agreed—but that supports a boutique market for human authenticity, not necessarily employment for billions of people.
Round 2: Why has mass unemployment not appeared already?
Cowen: The current evidence should reduce confidence in imminent labour-market collapse. Even young-worker unemployment has not clearly broken from its pre-AI trend. AI-exposed work is changing, but aggregate employment is not falling off a cliff.
The more realistic near-term problem is transition. New jobs may be concentrated in slow, regulated sectors such as energy and healthcare. Employers may struggle to identify who is actually good at supervising AI. Governments may spend money poorly because they cannot distinguish competent AI-enabled workers from people producing polished-looking nonsense. That can create painful mismatch without creating permanent mass unemployment.
Kokotajlo: Current employment data are a lagging indicator of the scenario I am worried about. Frontier companies are not trying to automate the whole economy first. Their sequence is more like:
Automate parts of AI research.
Use that acceleration to develop much more capable AI.
Deploy the resulting systems across the economy.
Therefore, widespread unemployment may arrive only after systems have become strategically powerful. Waiting for unemployment statistics to prove the danger could mean waiting until the crucial decisions have already been made.
Cowen: That depends on a fast intelligence explosion. I assign more probability to bottlenecks, diminishing returns and imperfect deployment. Humans will slow the process even when the underlying models improve quickly.
Kokotajlo: Yes. The disagreement is largely about whether AI research automation produces an unusually fast takeoff. If progress stalls, your account becomes much more plausible.
Round 3: Will AI create more work—or make work optional?
Cowen: AI is already making some highly motivated people work more, not less. It raises the return to initiative: one ambitious individual with agents may accomplish what previously required an organisation. The immediate response for young people should therefore be to learn these systems intensely rather than assume there will be nothing left to do.
The winners may not be conventional rule-following professionals. They may be unusually curious, energetic people who repeatedly experiment with the technology. Intelligence remains valuable, but initiative, judgement, charisma and the ability to discover worthwhile problems become relatively more important.
Kokotajlo: That is good advice during the transition. In the late 2020s or early 2030s, many jobs may indeed consist of managing teams of agents. But “learn to manage AI” is not a permanent refuge if the AI eventually becomes better than you at managing AI.
I expect a period in which AI greatly increases individual productivity, followed—conditional on superintelligence—by a world where human employment is largely a political choice.
Cowen: You are describing the technological endpoint. I am describing the lived economic path. That path might last decades and contain enormous quantities of work, entrepreneurship, relocation and institutional rebuilding.
Kokotajlo: And I worry that the path could compress into several years.
Round 4: Who loses first?
Cowen: Not necessarily the poorest workers. A particularly vulnerable group is the highly paid, upper-middle-class knowledge worker whose advantage consists of processing information within established systems: consultants, lawyers, administrators, junior analysts and some academics.
They may remain employed but suffer a large status and income reduction. A partner earning $1.5 million might become a manager in an energy company earning $300,000. That is not destitution, but it could produce intense political backlash because these people currently possess disproportionate cultural and institutional power.
Kokotajlo: I agree that white-collar status disruption comes early. But I would emphasise entry-level positions. If AI performs the basic writing, coding and analysis through which juniors learn, firms may hire fewer beginners. Eventually, however, the distinction between junior and senior matters less: sufficiently capable agents threaten both.
Cowen: The apprenticeship problem is real. We will have to redesign education and employment so that people learn by supervising, testing and challenging AI rather than by spending five years producing first drafts. I have argued that a substantial share of education should involve learning to use AI properly.
Round 5: Is income the main problem?
Kokotajlo: No. People need both income and power.
A citizen’s dividend, AI tax or broad ownership mechanism can distribute some of the enormous output produced by AI. But employment currently gives ordinary people leverage. Workers can strike, withhold expertise and make themselves economically costly to ignore.
If almost all production comes from AI companies, data centres and robot fleets, governments may become less dependent on the population. People could receive comfortable payments while losing political influence. A post-work settlement therefore needs durable democratic control, not merely UBI.
Cowen: That is one of Daniel’s strongest arguments. A wealthy post-work society is not automatically a liberal society. Concentrated AI capability could become a power magnet for corporations or the state.
But I resist designing the entire economy around the assumption that labour will become worthless. Prematurely freezing a particular redistribution system could entrench incumbents and slow the very growth that makes abundance possible. Build energy, housing, healthcare, research capacity and broad access to AI first.
Kokotajlo: The danger is that “wait until displacement is obvious” becomes “wait until the owners of AI possess overwhelming bargaining power.”
Closing statements
Cowen: AI will radically alter work, destroy particular career ladders and reduce the returns to some forms of intelligence. Yet human wants are expansive, institutions are slow and economic activity is not a fixed quantity. Expect turbulent reallocation, more experimentation, new human-facing work and eventually more leisure—not a permanent society in which virtually nobody works.
Start learning AI now. Develop initiative, embodied skills, judgement, persuasion and the ability to work on problems the models cannot cleanly specify.
Kokotajlo: Historical reassurance is unreliable when the technology being developed is a general substitute for human cognition and may accelerate its own development. If superintelligence arrives, the question will not be which jobs AI can perform. It will be which roles society deliberately keeps human.
Prepare before unemployment becomes the headline: slow dangerous races, solve alignment and control, distribute ownership, and protect the political power of citizens in a world that may no longer economically need their labour.