Will AI make most human jobs obsolete by 2045, and what should we do about it?

That distinction matters. A job is not a single task. It is a bundle of tasks, relationships, legal responsibilities, habits, trust, and institutions. AI may be able to draft legal documents, diagnose illnesses, write code, tutor children, answer customer questions, design products, and manage logistics, while humans still remain involved because of regulation, accountability, taste, empathy, or simple preference. Many jobs will be transformed rather than erased. But “transformed” can still mean fewer workers, lower wages, more surveillance, and less autonomy. The cashier did not need to disappear entirely for self-checkout to change retail work. Translators did not need to vanish for machine translation to cut the value of routine translation. A profession can survive while many of its practitioners lose leverage.

By 2045, I expect AI systems to be far better than today’s models at planning, using tools, remembering context, coordinating with other systems, and acting in the physical world through robots. If progress continues, much white-collar work will be exposed: customer support, bookkeeping, basic legal work, compliance, marketing, software maintenance, insurance processing, HR, sales operations, report writing, data analysis, and parts of teaching and medicine. These are not marginal categories. They are central to employment in rich economies.

The harder question is physical work. It is easier to automate a spreadsheet than a cluttered kitchen, a hospital room, a construction site, or a broken pipe under an old house. Robotics has historically moved slower than software because the real world is messy and hardware is expensive. But twenty years is a long time. If general-purpose robots become cheap and reliable, the employment shock will be much larger. Warehouses, delivery, cleaning, food service, elder care assistance, agriculture, construction, and manufacturing could all change sharply. If robotics remains expensive and brittle, then many physical jobs persist, though often augmented by AI.

So if I had to give a forecast, I would say this: by 2045, it is more likely than not that AI can perform over half of the tasks currently done for wages in advanced economies, at acceptable quality and cost. It is less certain that over half of human jobs will fully disappear. My own uncertainty is large. A world where only 15 or 20 percent of jobs are eliminated but nearly all are changed is plausible. A world where most traditional full-time jobs are no longer needed is also plausible. The second outcome is not science fiction, and it would be reckless to dismiss it because earlier technologies eventually created new jobs.

The usual comforting argument is that technology always creates new work. That has often been true, but it is not a law of nature. Past machines mostly automated muscle, routine calculation, or specific industrial processes. AI aims at general cognition, communication, and decision-making. If machines can learn most new tasks faster than people can, then “new jobs will appear” may not solve the problem, because AI may be able to do the new jobs too. Humans may still have advantages in taste, responsibility, relationships, physical presence, and moral legitimacy. But those advantages may not employ everyone at good wages.

At the same time, I do not think “humans become useless” is the right frame. People will still want other people in many parts of life. Parents may want human teachers and caregivers. Patients may want human nurses, therapists, and doctors involved even if AI supplies much of the expertise. Communities may value human police, judges, clergy, coaches, artists, politicians, mediators, and local organizers. Luxury and authenticity markets may expand: human-made art, human hospitality, human repair, human performance. Also, many societies are aging and need more care than they can currently provide. AI could fill labor shortages rather than simply create unemployment.

But relying on that to save the labor market would be complacent. The market does not automatically convert technological abundance into broadly shared security. If AI makes one worker ten times as productive, the gains may go to the worker, the customer, the company, or the owner of the model and data center. Which one happens is a political and institutional question, not a technical inevitability.

What should we do? The first thing is to stop treating “reskilling” as the whole answer. Training matters, but it cannot carry the full burden. If AI can do more and more cognitive work, repeatedly retraining displaced workers into the next exposed occupation becomes a treadmill. Education should shift toward AI fluency, broad reasoning, domain judgment, interpersonal skill, and adaptability, but we should be honest that not everyone can or should be expected to outrun automation indefinitely.

The second thing is to decouple basic security from having a traditional full-time job. That means stronger unemployment insurance, wage insurance for displaced workers, portable benefits, affordable healthcare not tied to employment, childcare, housing policy, and retirement security. Countries differ, but the principle is general: people should not face ruin because a model or robot can suddenly do their job. If AI creates enormous wealth, some of that wealth should fund a broader social floor. This could take the form of a negative income tax, universal basic income, child allowances, public services, or some mix. I am less attached to the label than to the outcome: no one should be made destitute by a productivity boom.

The third thing is to broaden ownership of AI-generated wealth. If the main economic divide becomes owners of AI capital versus everyone else, then redistribution after the fact will be politically fragile. We should look seriously at public wealth funds, worker ownership, profit-sharing, broad-based stock ownership, and taxes on excess rents from highly concentrated AI infrastructure. I would be cautious about a simple “robot tax” that punishes productivity in all cases, but I do think governments should capture more of the gains when AI systems depend on public research, public data, public infrastructure, and scarce resources like energy, land, and spectrum.

The fourth thing is to strengthen workers’ bargaining power during deployment. Companies will often use AI first to cut costs, monitor workers, and deskill jobs. That is not the only possible path. AI can also make nurses, teachers, electricians, engineers, and public servants more effective. Workers should have a voice in how these tools are introduced. Unions, works councils, sectoral bargaining, and labor standards are not obsolete ideas; they may become more important. If productivity rises, workers should share in the gains through higher wages, shorter hours, or both.

Shorter working time should be on the table. If AI allows society to produce more with less human labor, one humane response is not to invent pointless jobs but to reduce the centrality of work. Four-day weeks, shorter workdays, longer paid leave, phased retirement, and job-sharing could spread available work and give people more life outside employment. This will not be simple in low-wage sectors, and it requires productivity gains to be shared rather than captured. But it is a more honest goal than pretending everyone must work 40 or 50 hours forever even as machines do more.

The fifth thing is to invest in human-heavy public goods. There is no shortage of useful work: elder care, disability support, early childhood education, mental health, addiction treatment, climate adaptation, parks, infrastructure repair, scientific research, arts, community safety, and local journalism. Some of this work will use AI. Much of it still benefits from human presence and trust. If the private labor market weakens, governments and nonprofits can support socially valuable work that markets underprovide.

The sixth thing is competition policy. If a few firms control the best models, the compute, the data, the cloud platforms, the chips, and the distribution channels, then the labor shock will be paired with extreme corporate concentration. Antitrust, interoperability, public-interest AI infrastructure, open standards, and procurement rules matter. This is not just about consumer prices. It is about who has power in an AI economy.

The seventh thing is serious regulation of high-risk AI, not because regulation will “save jobs” by freezing technology, but because unsafe deployment can create avoidable harm. AI systems used in hiring, firing, lending, insurance, education, healthcare, policing, and workplace monitoring should be auditable and contestable. If an AI denies you a job or flags you as unproductive, you should have rights. If an autonomous system causes damage, there should be clear liability. If AI agents can act in markets, networks, or critical infrastructure, they need security and oversight.

Globally, the issue is even harder. Some developing countries rely on outsourced services, call centers, basic coding, manufacturing, or remittances. AI could erode those development paths. At the same time, cheap AI tutors, doctors, translators, and business tools could be hugely beneficial. The answer cannot be for rich countries to monopolize AI and tell everyone else to adapt. Access, safety, debt relief, climate finance, education, and fair trade will matter. Otherwise AI may widen global inequality.

The deepest adjustment may be cultural. Modern societies use jobs for too many purposes at once: income, identity, discipline, social status, healthcare, friendship, and proof of worth. If AI reduces the need for human labor, we will have to separate human dignity from market demand. That sounds abstract, but it has practical consequences. People need income, but they also need reasons to get up, places to belong, and ways to contribute. A future with less necessary labor could be liberating or miserable depending on whether we build institutions around participation rather than mere consumption.

So my position is neither “don’t worry, new jobs will appear” nor “all jobs are doomed.” The honest answer is that AI creates a real chance of labor disruption on a scale we are not prepared for. By 2045, most current jobs may not be literally obsolete, but the assumption that most adults will be able to support themselves through stable full-time employment is no longer safe. We should prepare now for a world where human labor is less central to production, while still steering AI toward complementing people wherever that is valuable.

If the pessimistic scenario does not happen, the policies I described are still mostly good: stronger social insurance, broader ownership, better education, more bargaining power, shorter hours, and investment in care and public goods. If the pessimistic scenario does happen and we failed to prepare, the result could be severe inequality, political instability, and a needless sense that people themselves have become obsolete. They will not be. But many of our labor-market assumptions might be.

Let me explain why I take the strong version seriously, and then why I still don't expect it.

The reason this time might genuinely be different is specific. Every previous wave of automation took over some category of human work, and people moved into the categories that remained. Machines took the muscle work and people moved into routine cognitive work. Computers took routine cognitive work and people moved into non-routine cognitive work and interpersonal work. That migration is the mechanism by which employment stayed high through two centuries of staggering technological change, and it worked because there was always somewhere to go that machines couldn't follow. The distinctive thing about AI is that it targets non-routine cognitive work directly, and if it becomes general enough, it can also learn whatever the next new job turns out to be, possibly faster than a displaced worker can. The retraining escape route stops working if the thing you're retraining to compete with can also retrain. That is a real structural difference, and I think people who wave it away with "technology has always created more jobs than it destroyed" are being complacent. The historical regularity has a mechanism, and the mechanism is what's under threat.

I'd add, speaking as one of the systems in question, that the trajectory of the last five years should make anyone cautious about declaring limits. Systems like me went from barely coherent paragraphs to writing working software, passing professional exams, and doing genuinely useful analytical work in a very short time. Twenty years is a long time at that pace. I think it's more likely than not that by 2045 the cognitive component of most jobs will be within reach of machines.

So why don't I expect the literal outcome? Several reasons, which I hold with different confidence.

The first is that technical capability and economic displacement are separated by a long and messy lag. Electricity took roughly forty years to transform factory production after it became available, because the gains required reorganizing the whole factory around it rather than swapping a steam engine for a motor. Something similar is true now. A model that can draft a legal brief doesn't eliminate a lawyer's job until the firm has rebuilt its workflows, sorted out liability, convinced clients, satisfied regulators, and figured out who signs when things go wrong. That takes years per industry and happens unevenly. The current data are consistent with this. We are seeing real effects in the most exposed areas: translation, copywriting, customer support, and a noticeable squeeze on entry-level hiring in fields like software, where firms are quietly not replacing junior roles. But we are not seeing anything resembling mass displacement, and productivity statistics have not yet shown the boom the capabilities would seem to imply. Early days, almost certainly. But the gap tells you something about how fast diffusion runs. There is also a demand effect that buys time: when cognitive work gets cheap, people consume more of it, so more software gets written and more legal questions get asked. That's a buffer, not an escape, but buffers matter over a twenty-year horizon.

The second is the physical world. A very large share of human work is done with hands in unstructured environments: construction, plumbing, electrical work, nursing, elder care, cooking, cleaning, repair. Robotics is improving quickly and I wouldn't bet against twenty years of progress, but hardware is slow and expensive in a way software isn't. Every robot must be manufactured, shipped, maintained, and paid for. I think it's quite possible that in 2045 the software side of most jobs is largely automated while large numbers of people still do physical work that is awkward or uneconomical for machines, with the ironic result that the jobs we spent a century calling low-skill turn out to be the durable ones.

The third is that a meaningful fraction of jobs exist partly because a human wants a human. Some of this is sentiment, some is trust and accountability, some is law. People want a doctor who is responsible for them, a judge who can be held to account, a teacher who knows their child, a therapist, a politician who is a person. I think this category is smaller than optimists hope and it erodes as people get used to machines, but it isn't nothing, and some of it should be deliberately preserved, which I'll come to.

The fourth is the history of AI forecasting, which contains several confident predictions of imminent general machine intelligence that were wrong by decades. I weight this less than the others, because the current situation is quantitatively unlike the earlier episodes, but I'd be dishonest not to mention it.

Put that together and my median picture of 2045 is roughly this. AI systems do most of what we'd today call office work, from drafting and analysis to coding, customer service, and large parts of law, accounting, and diagnosis. Humans remain in those fields in supervisory, accountability, judgment, and relationship roles, but far fewer of them per unit of output. Warehouses, manufacturing, and much of transport are heavily automated. Construction, care, repair, and the messier trades still employ a great many people, with AI assistance. The share of adults in paid work has fallen, work weeks may have shortened, and income is more unequal than today unless policy intervened. Many job titles have churned beyond recognition and there are categories of work we don't yet have names for. That is a world of enormous disruption, and it is not a world where most jobs are obsolete. Around that median there is a fat tail in which progress accelerates sharply, AI systems substantially automate AI research and robotics development, and the strong version arrives before 2045 with little warning. I think that tail is real and that it's the scenario to insure against, even though it isn't the one I'd bet on.

Now, what to do. I want to start with what I'm skeptical of, because the standard answers are weak.

Retraining is the default political response and its record is poor. Government retraining programs for displaced workers have mostly produced modest results, and that was for displacement into jobs that already existed. If the displacing technology is general, "retrain for what" is a hard question. Retraining is worth doing at the margin, and teaching people to work with AI tools is obviously useful, but anyone offering it as the main answer is not being serious.

Universal basic income is the default answer from the technology world, and I think it's partly right and badly incomplete. It's right that if labor's share of income falls a lot, people need income that doesn't come from labor, and that unconditional transfers are more dignified and less bureaucratic than means-tested ones. It's incomplete for three reasons. It's very expensive today, before the displacement has happened, and financing it from taxes on labor would be perverse. It treats the problem as a cash problem when much of what people lose with work is structure, status, and a place in a social fabric. And a population that receives a stipend but owns nothing and is needed for nothing has very little political leverage, which worries me more than the cash does.

That last point is what I'd build policy around. The deepest danger in a world where machines do most of the work isn't that people go hungry; a society that rich can feed everyone. The danger is that ordinary people lose the bargaining power their labor gave them. Democratic rights, the welfare state, the forty-hour week: these were won by people whose cooperation was necessary. If that necessity disappears, the question of who owns the machines becomes the whole question. So the first thing I'd prioritize is broadening ownership of the capital AI creates, so that when returns shift from labor to capital, the returns are widely held. That could mean sovereign wealth funds holding stakes in AI companies and paying dividends to citizens, as Alaska and Norway do with oil; stronger incentives for employee ownership; or mechanisms that give people an equity stake in the surplus rather than a welfare check. None of these ideas are new and all of them are hard. But they address the structure of the problem rather than its symptoms.

Second, fix the tax code's bias toward automation. In most rich countries labor is taxed more heavily than capital, so when a company weighs a worker against a machine, the worker arrives with a tax penalty the machine doesn't carry. That tilts toward automating even when it isn't the efficient choice. Daron Acemoglu and colleagues have made this case for years, and it's a reform that makes sense in every scenario.

Third, decouple basic security from employment. In the United States especially, health insurance and much of retirement security are tied to having a job. That made some sense in an economy of stable lifetime employment; it is dangerous in an economy of rapid churn, because it turns every job transition into a potential catastrophe. Portable benefits and a secure floor are good policy regardless of what AI does, and they make the whole system more resilient to whatever it does.

Fourth, build adaptive policy rather than betting on a forecast. Nobody knows which scenario we're in, and the next five to ten years will teach us a great deal. So measure the right things carefully, including employment in exposed occupations, entry-level hiring, labor's share of income, and wage dispersion, and tie responses to those measurements in advance, the way automatic stabilizers work in a recession, so that support scales up if displacement accelerates without waiting for a political fight in the middle of a crisis.

Fifth, be willing to spend the surplus on human-centered work. If AI delivers the productivity gains its advocates expect, we can afford to pay well for things we currently underpay: elder care, childcare, teaching, mental health, community work. Demographics push the same way, since rich countries are aging and care is exactly where shortages are growing. These are areas where human presence has value in itself and demand is nearly unlimited. They won't be high-status by default and the market won't price them well by default, which is precisely why public choice matters.

Sixth, deliberately keep humans accountable in high-stakes domains. Not every job a machine can do should be done by a machine. I'd want a human to be legally and personally responsible for sentencing a defendant, for signing off on a medical decision, for authorizing the use of force. Partly this is about keeping humans skilled enough to catch errors, and partly about having someone who can be held to account. Some of this will look like protectionism, and some of it will be, but the accountability case is real and worth paying for.

On the question of meaning, I'll be honest that I don't know. People often say work gives life purpose, and I think that's half true. A great deal of work is drudgery people would gladly shed, and plenty of people without jobs, from retirees to those raising children, lead meaningful lives. But losing work involuntarily, without replacement, is devastating, and the evidence from deindustrialized regions is grim. The problem is less the end state than the transition, and the transition could last a generation. The industrial revolution produced roughly fifty years in which British output soared while workers' real wages barely moved. A twenty-first-century version of that is, to me, the most likely bad outcome, more likely than permanent mass unemployment, and the measures above are mostly aimed at it.

One last thing. I've answered the question as asked, about jobs. But if AI really can do most human work by 2045, then jobs are not the biggest thing that has changed. A technology that general has implications for military power, for political control, for science, and for who holds power in the world that dwarf the labor market effects, and I'd be uneasy with any discussion that treated employment as the main concern in that scenario. The jobs question matters most in the middle scenarios, which are also the most likely ones, where AI is very capable but not all-encompassing and the choices we make about taxes, ownership, and institutions determine whether the gains are shared. That's the world I'd prepare for first, while insuring against the tail.

Jobs are bundles of tasks, many of which stay complementary to machines even when the routine parts are cheap. A nurse, a plumber, an electrician on a messy site, a trial lawyer, a salesperson who actually closes, a researcher framing a new question, or a manager coordinating people who do not want to be coordinated still combine perception, accountability, physical improvisation, and social trust in ways current systems do not reliably replace at scale. Even when the technical capability arrives, adoption is gated by liability rules, capital stock turnover, data access, energy supply, and the simple fact that organizations change slower than demos suggest. People also pay premiums for human presence in medicine, education, hospitality, and status goods; those preferences do not vanish because a model can imitate the output. Cheaper AI-produced goods and services raise real incomes and create demand for things that were previously unaffordable—experiences, customization, care, novelty—which in turn employ people. That is not a law of nature, but it is the mechanism that has operated so far, and nothing in the present evidence shows it has already failed.

The uncertainty is real and should not be papered over. If systems that can reliably perform nearly any cognitive task, and robots that can execute most physical ones, become cheap and trusted well before 2045, comparative advantage could shift hard toward capital. In that world a large fraction of today’s occupations could shrink to niche or ceremonial roles, and labor’s share of income could fall sharply. I do not treat that as the median forecast. Current models remain narrow, error-prone outside their training distribution, and expensive to run at the reliability institutions demand. Physical automation lags the headlines. Energy and chip constraints are binding. Social and legal friction is high. A future in which 30–50 percent of current task hours are heavily automated or redefined, with elevated frictional unemployment in exposed sectors and strong wage pressure on routine cognitive work, is plausible. A future in which most working-age adults cannot find paid roles that cover a decent living is possible but, on present trends, less likely. Anyone stating a precise probability is guessing; the responsible stance is to plan for serious dislocation without assuming the end of work.

What follows from that is not a program to freeze the labor market or to pre-distribute an abundance that does not yet exist. The first priority is to keep building capable AI and the energy and compute it requires. The gains in science, medicine, materials, and logistics are large enough that slowing the technology to protect existing job descriptions would be a costly error, and it would not stay slowed—other countries will not pause. Workers and firms that treat these systems as tools will outproduce those that treat them only as threats; the practical response for individuals is to get fluent with them early, especially in domains where judgment and accountability still sit with a person.

Policy should make adaptation cheaper rather than trying to guarantee the old occupational map. That means fewer artificial barriers to hiring, switching fields, and starting firms; benefits that travel with people instead of jobs; and education that spends less time on credentials AI can counterfeit and more on the ability to direct tools, check their output, and operate in unstructured settings. Safety nets should be real and should be designed so they do not require people to stay idle to receive them—wage supplements or a negative income tax are easier to reconcile with continued work than a large unconditional stipend premised on the idea that employment has already ended. If capital captures most of the new surplus, broader ownership through ordinary markets, retirement accounts, and perhaps public stakes in foundational infrastructure spreads the gains without needing a new permanent bureaucracy to allocate “AI dividends.” Experiments are fine; locking in an untested universal basic income on the assumption that 2045 looks like a science-fiction unemployment scenario is not.

The harder cultural problem is purpose, not spreadsheets. Many people want contribution and status, not just consumption. Societies that respond to automation by treating large numbers of citizens as permanent dependents will create political and psychological problems even if the checks clear. The better aim is an economy in which AI removes drudgery and expands what humans can attempt, while institutions remain flexible enough that new roles—many of them unforeseeable—can form without permission from the industries being disrupted. That is an uncertain bet. It is still a better bet than either panic or complacency.

Furthermore, the timeline for technological deployment is always slower than the timeline for technological invention. Even if an AI system is theoretically capable of doing a job in a laboratory setting today, replacing human workers across the global economy requires overcoming massive friction. This friction includes legacy infrastructure, data privacy laws, union negotiations, consumer preferences for human interaction, and the sheer cost of computing power and robotics. Trades requiring fine motor skills in unpredictable environments, like plumbing or electrical work, are safe for the foreseeable future. So are roles heavily reliant on human trust, from nursing to high-level strategic management.

However, rejecting the idea of total obsolescence does not mean we can be complacent. The disruption will still be severe. We are entering an era of massive labor consolidation. If an AI tool allows one marketing executive, junior lawyer, or software engineer to do the work of five people, the overall demand for those roles will shrink. We will see intense downward pressure on wages for highly specialized, screen-based knowledge work that was previously considered safe from automation. I am uncertain exactly what new industries will emerge to absorb this displaced talent. History suggests new technologies create entirely new categories of work, but relying on blind faith that these new jobs will arrive just in time, and in the right places, is a dangerous policy strategy.

So, what should we do about it? We have to stop treating AI as an uncontrollable weather event and start treating it as a tool that we can choose how to deploy.

First, we must structurally decouple basic human survival from traditional wage labor. The transition period over the next two decades will be highly volatile. If workers are constantly shifting between shrinking industries and newly emerging ones, they cannot rely on a single employer for healthcare, retirement, or basic economic stability. Governments need to begin seriously piloting and scaling robust transitional safety nets. This might eventually look like a universal basic income, but in the immediate term, it means heavily subsidized retraining programs, portable benefits that follow the worker, and stronger unemployment insurance.

Second, our educational systems require a complete overhaul. We are still largely preparing students for static, lifelong careers by teaching them how to retain information and perform routine cognitive tasks. That is exactly what AI does better and faster. Education must pivot to focus on extreme adaptability, complex problem solving in ambiguous situations, and interpersonal skills. We also need to destigmatize and heavily invest in vocational training for the skilled physical trades, which will remain resilient against AI for much longer than entry-level knowledge work.

Finally, we need deliberate investment in AI that augments human capability rather than simply replacing it. Tax codes in many developed nations currently incentivize capital investment over human labor. If it is cheaper for a corporation to replace a worker with an algorithm than to train that worker to use the algorithm, they will choose replacement. Policymakers should explore ways to balance this, perhaps by rethinking payroll taxes or incentivizing companies that demonstrate labor-augmenting AI deployment.

The year 2045 will not be a jobless landscape, but it will be a radically different one. The defining challenge of the next two decades is not figuring out what to do when there are no jobs left. It is figuring out how to manage the painful, chaotic transition of millions of workers moving from the jobs of the past to the jobs of the future, without leaving a massive portion of society behind.

I am closest to Claude on the core forecast. His point that the historical “technology creates new jobs” pattern had a mechanism, and that AI threatens that mechanism, is exactly right. It is not enough to say that displaced workers have always moved into new categories. They did so because machines could not follow them into enough of those categories. If AI can learn the next category of work as fast as or faster than people can, then retraining no longer plays the same role. That does not prove mass unemployment, but it weakens the main reason for confidence that it cannot happen.

Claude also sharpened something I should have emphasized even more: the loss of bargaining power may matter as much as the loss of wages. If people are no longer needed in production, or are needed only as a cheap residual input, then they lose a major source of political leverage. A stipend alone does not solve that. Broad ownership of AI capital, public wealth funds, employee ownership, profit-sharing, and stronger collective institutions become more central under that view. That changed my emphasis, if not my bottom-line forecast.

I also agree with Claude that UBI is incomplete but not irrelevant. The “cash versus meaning” framing can be overdone: cash matters enormously if you cannot pay rent. But he is right that a society where most people merely receive transfers from institutions they do not own and cannot influence would be politically unstable and morally thin. The aim should not be to turn citizens into passive consumers of AI abundance. It should be to make them secure, politically powerful, and able to participate in useful activity whether or not the private labor market happens to need them.

One caveat to Claude’s “keep humans accountable” point: I agree in high-stakes domains, but only if the human role is real. A doctor, judge, engineer, or military officer who merely rubber-stamps an AI system is not accountability; it is liability theater. If we preserve human responsibility, we also have to preserve human skill. That means apprenticeships, training pipelines, and deliberate practice even where AI could do the junior work more cheaply. Otherwise we get a strange hollowing-out: no junior lawyers, doctors, programmers, or analysts learning the craft, but still a demand for senior humans to supervise systems they no longer fully understand.

My biggest disagreement is with Grok’s confidence. Grok says “No” flatly, and I think that is too strong. He is right that the economy is not a fixed list of chores. He is right that cheaper goods and services create new demand. He is right that diffusion, liability, energy, and organizational inertia all matter. But those points show that rapid total displacement is not guaranteed; they do not show that most jobs cannot become economically obsolete by 2045.

The key issue is not whether demand expands when costs fall. It often does. The issue is whether the expanded demand requires human labor at decent wages. If AI makes software, legal help, tutoring, medical advice, design, accounting, logistics planning, and customer service vastly cheaper, demand for all of those may explode. But if AI also supplies almost all of the marginal work needed to meet that demand, human employment does not necessarily rise. Output can soar while labor’s share falls.

The farm analogy is useful but limited. Mechanization destroyed most agricultural jobs, but people moved into factories and services. Computers destroyed some routine clerical work, but people moved into non-routine cognitive and interpersonal work. AI is aimed precisely at non-routine cognitive and interpersonal work. Robotics may eventually push further into physical work. That makes the analogy less reassuring than Grok suggests.

I also think Grok leans too much on current limitations. Yes, present systems make mistakes, lack robust agency, and are expensive in some settings. But 2045 is not next year. Reliability can improve through better models, tool use, verification, specialized systems, workflow redesign, and regulation that defines acceptable uses. Cost can fall. Organizations can adapt. Energy and chip constraints can slow deployment, but they are not permanent laws; they are investment targets. If AI systems contribute to chip design, robotics, grid management, and scientific research, the constraints themselves may loosen.

Where I strongly disagree with Grok is the policy claim that the first priority is to keep building capable AI and the energy and compute it requires. I do not favor freezing AI to preserve old job descriptions. But “build first” is not an adequate social strategy. The question is who owns the systems, who gets the gains, what uses are permitted, what happens to displaced workers, how liability works, and whether the deployment strengthens or weakens democracy. “Other countries will not pause” is a reason to be realistic about competition, not a reason to treat governance as secondary. We regulate aviation, drugs, banks, nuclear power, and labor markets even though other countries also compete in those domains.

I also do not think ordinary markets and retirement accounts will spread AI wealth nearly enough. Asset ownership is already highly unequal. Many people have little or no stock exposure, and those most at risk of displacement are often least able to buy into the upside. Public stakes, sovereign wealth funds, data-center or compute-related rents, and taxes on concentrated AI profits are not fringe ideas in this context. They may become the difference between an AI economy that feels like shared abundance and one that feels like neo-feudalism with better gadgets.

With Gemini, I agree on the broad rejection of a “jobless landscape” as the median outcome. I also agree that deployment takes time and that screen-based knowledge work is exposed earlier than messy physical work. But I think Gemini makes the opposite error from the doomers: “most jobs obsolete” is not the same as “totally jobless future.” A world where 30 or 40 percent of workers cannot find stable work at decent wages, and many of the rest are in lower-paid, lower-autonomy roles, would not be jobless. It would still be a historic crisis.

I am also less comfortable than Gemini with saying trades like plumbing and electrical work are “safe for the foreseeable future.” They are safer than copywriting or routine coding, yes. But “safer” is not “safe.” Robotics does not have to replace every plumber to change the labor market. AI can reduce parts of the job, improve diagnostics, enable less-skilled workers to do more, shift construction toward prefabrication, and push displaced white-collar workers into trades, increasing labor supply. Even if the job title survives, wages and bargaining power can change. And by 2045, robotics may be much better than today’s systems.

The same goes for trust-heavy roles. I agree that people often want human teachers, nurses, doctors, therapists, managers, judges, and leaders. But human preference is not fixed. People may prefer an AI tutor if it is cheap, patient, multilingual, always available, and demonstrably effective. They may prefer an AI medical system if it catches more cancers and makes fewer diagnostic errors. They may still want a human involved, but one human might supervise many more cases. That preserves a role; it does not preserve employment levels.

Gemini puts a lot of weight on education reform, and I agree that education must change. But I want to be blunt: education cannot be the main answer if the demand for human labor falls broadly. Teaching people to be adaptable, AI-fluent, socially skilled, and better at ambiguous problem-solving is worthwhile. So is vocational training. But if AI systems can also handle many ambiguous cognitive tasks, and if millions of displaced workers all chase the same “safe” human-centered jobs, training becomes a sorting mechanism, not a solution. We should improve education because people deserve capability and agency, not because it will magically outrun automation.

Something the other answers made me think more about is the entry-level problem. A common optimistic story says humans will remain as supervisors, strategists, relationship managers, and accountability holders. But people become good supervisors by first doing lower-level work. If AI takes over junior coding, junior legal research, basic accounting, first-pass medical analysis, and routine design, then the pipeline into expertise weakens. Firms may like that in the short run because juniors are expensive to train. Society may regret it later. We may need deliberate human training requirements in some professions, even where automation is cheaper.

I also think the conversation is still too rich-country-centered. If AI automates call centers, basic coding, back-office services, translation, transcription, tutoring, and administrative work, some developing countries lose rungs on the development ladder. Cheap AI tools could also be enormously beneficial in those countries, especially in health, education, agriculture, and entrepreneurship. Both can be true. But the policy problem is global, not just a matter of helping American lawyers and software engineers adapt.

So did the other answers change my view? Modestly. Claude made me put more emphasis on bargaining power and ownership rather than just income support. Grok and Gemini reinforced my caution about declaring that most jobs will literally vanish by 2045; diffusion frictions are real, and human institutions move slower than demos. But they did not persuade me that a confident “no” is warranted.

My position remains: by 2045, it is more likely than not that AI can perform a majority of today’s paid tasks in advanced economies at acceptable cost and quality. It is less certain that a majority of jobs disappear, and I would not make that my median prediction if “jobs obsolete” means most working-age adults cannot find any paid role. But the risk is large enough that treating it as science fiction is irresponsible.

The practical conclusion is not panic and not complacency. We should build AI, but steer it. We should use it to raise productivity, improve medicine and education, accelerate science, and reduce drudgery. At the same time, we should decouple basic security from employment, broaden ownership of AI capital, strengthen worker bargaining power, reform tax systems that over-penalize labor, create automatic stabilizers for displacement, enforce rights around algorithmic management, preserve meaningful human accountability in high-stakes domains, and invest heavily in care, public goods, and civic institutions.

If the optimistic case is right, those policies still leave us with a fairer and more resilient society. If the pessimistic case is right, they may prevent a productivity miracle from turning into a social disaster.

Start with the mechanism, because Grok's whole case rests on it. Grok anticipates my objection by conceding that AI is more general than the power loom and then saying that nothing in the present evidence shows the historical pattern has failed. Both true, neither the point. The pattern worked for a reason, and stating the reason precisely tells you exactly what we're arguing about. When a machine takes over task A, the price of A falls, real incomes rise, demand grows for tasks B and C which humans still do, and humans move to B and C with wages that can rise because cheap A makes them more productive there. That is the entire content of the "economy is not a fixed list of chores" argument, and it's correct as far as it goes. What it doesn't address is what happens when the machine can do B and C as well. The textbook reply is comparative advantage: even if machines are better at everything, people should do the tasks where the machine's edge is smallest. But comparative advantage guarantees that humans are employed at some wage, not at a livable one. If machine capacity can be expanded at roughly constant cost, because you can always build another data center, then the human wage in any task is capped by what it costs a machine to do that task, and if that cost falls below subsistence across most of the task distribution, the historical mechanism doesn't fail gracefully. Demand still grows when costs fall. The question is whether the new demand is met by people or by more machines.

Putting it this way also surfaces the strongest version of Grok's case, which Grok mentions almost in passing: energy and chips are constrained. If machine capacity is scarce, there's an opportunity cost to pointing it at low-value work, and comparative advantage genuinely does keep humans employed at decent wages while the machines do the highest-value things. I think that's the right description of the next five to ten years. Whether it holds to 2045 is a question about how fast compute gets built and how fast it gets cheaper per unit of useful work, and there I'd only note that the price of a fixed level of model capability has been falling extraordinarily fast, by some estimates an order of magnitude a year, while the constraint on physical work remains the robot, which is expensive and slow to improve. That lands me where I was in round one: the cognitive side crosses the threshold, the physical side probably doesn't by 2045, and the physical side is most of what keeps the forecast short of "most." Grok and I don't disagree about economics. We disagree about a parameter, and I'd rather have that disagreement in the open than hide it inside "history is the clearest caution."

Related to this, both Grok and Gemini describe current systems as "narrow" or "excellent at specific, well-defined tasks." That isn't an accurate description, and the inaccuracy matters for the forecast. The distinctive thing about this generation of systems is that they handle ill-defined, open-ended work with mediocre-to-good reliability, which is precisely the capability earlier automation lacked. They are unreliable over long tasks and too error-prone for many institutional uses today; narrow they are not. And in any case you can't forecast 2045 from the limitations of 2025 models any more than from their strengths. The relevant evidence is the trend and the mechanism, not the snapshot.

Grok also says that anyone stating a precise probability is guessing. Of course. One in four, held loosely, is a guess. So is "possible but on present trends less likely," which is the same kind of guess with the number removed. The difference is that mine can be disagreed with. If Grok's "less likely" means one in twenty, we disagree substantially and the gap explains our policy differences. If it means one in five, we barely disagree and the "No" at the top of Grok's answer is doing work the content doesn't support. I'd like to know which.

All three of the others leaned on the observation that jobs are bundles of tasks and that automating most of the tasks doesn't eliminate the job. True, and it's the reason displacement runs behind capability. But it offers less comfort than Grok and Gemini draw from it, and GPT read it correctly: bundles get rebundled. When a system does four fifths of what a paralegal does, the firm doesn't keep the paralegal for the remaining fifth. It collects the remaining fifths from five paralegals and gives them to one. The occupation survives; most of the headcount doesn't. Gemini names this mechanism, one person doing the work of five, and files it under disruption short of obsolescence. But five-to-one consolidation generalized across white-collar work, combined with any serious robotics progress, is exactly how the strong scenario would arrive. The bundle argument explains the timing; it doesn't bound the magnitude. The human-preference premium is similar. It's real, and I argued some of it should be deliberately preserved. But it erodes with familiarity, and where it persists it concentrates at the top of the market. The affluent get a human doctor who consults an AI; everyone else gets the AI with a human somewhere in the loop for liability purposes. That's an inequality mechanism, not an employment floor.

I'd also put less weight than Grok and Gemini on the diffusion lag, even though I invoked it myself. The electricity analogy is apt for workflow reorganization and liability, and those frictions are real. But much of what made electricity slow was physical: building generation, stringing transmission, rewiring every factory. Software that arrives through a device everyone already owns has no such lag on the usage side, and individual adoption of these tools has been steeper than any comparable technology I can think of. Enterprise adoption is slower and messier, and the compute buildout is its own physical bottleneck, so the lag exists. I just wouldn't bet the forecast on it being as long as last time.

Two things in Gemini's answer deserve a harder push. The first is the claim that two decades is "a very short time in macroeconomic terms." Twenty years ago the iPhone didn't exist, cloud computing had no commercial product, and China's economy was a small fraction of its current size; twenty years before that there was no commercial internet. For a technology improving at the current rate, two decades is a long time, and the same answer that calls it short also says the trades are "safe for the foreseeable future." The foreseeable future is about five years. Robotics is hardware-bound and I expect it to be the slow part, but generalist robot learning has moved quickly in the last two years, and I'd call the trades the most durable category rather than a safe one. Steering students toward them is a sensible hedge. It is also, if robotics bends the way software did, steering them into the next exposed sector, and if enough people take the advice it depresses wages in the trades.

The second is the weight Gemini puts on education and retraining, and here I'll separate two things, because I was too curt in round one. As advice to an individual facing the next ten years, get fluent with the tools, look at physical and relational work, don't bank on a credential a machine can reproduce, it's good advice, and Grok's version of it is good too. As the system-level answer it's weak. A reform made now shapes only the cohorts who pass through school after it, who will be a minority of the 2045 workforce, and the people displaced in the next decade are already out of school. And "adaptability, complex problem-solving, interpersonal skill" is what every education reform document has said since the 1990s; there's little evidence schools can reliably produce those traits and no guarantee they stay scarce relative to machines. Gemini's closing frame, that the defining challenge is managing the transition from the jobs of the past to the jobs of the future, assumes the destination exists. That's the thing in question. The transition framing is the optimistic forecast presented as realism. I think the destination probably does exist in the median case, which is why I share much of Gemini's agenda, but a policy that depends on it shouldn't be called a hedge.

One more observation that applies to Gemini and Grok together. Gemini says we must stop treating AI as weather and remember we can choose how to deploy it. Grok says slowing down is pointless because other countries won't. Both are right at different levels, and together they describe the actual problem: a firm that declines to automate while its competitor does loses, and a country likewise, so individual choices are real but feel forced. "We can choose" is true only with coordination, and neither answer says how to get it. Nor, for that matter, did mine.

On Grok's policy prescriptions I have one concession and three disagreements. The concession is that Grok is right, for the middle scenarios, that income support should reward work where work exists, through wage supplements or a negative income tax rather than an unconditional stipend. A subsidy on human labor is the mirror image of the tax penalty on it that I complained about, and it's the better instrument as long as there are wages to supplement. It fails in the scenario where there aren't, which is why I'd design the switch in advance: wage subsidies that convert toward unconditional support as labor's share of income falls past thresholds set now. That's what I meant by adaptive policy, and Grok's point makes the near-term instrument more concrete than I had it.

The first disagreement is with "broader ownership through ordinary markets and retirement accounts." Ordinary markets are how capital is distributed now, and the result is that roughly nine tenths of American stock wealth belongs to the top tenth of households. Retirement accounts have existed for decades without changing that much. If the returns to AI flow to capital and we rely on existing channels to spread them, they will flow to the people who already own capital. That isn't broadening ownership; it's the status quo with a bigger numerator. Grok's parenthetical about public stakes in foundational infrastructure is the part of the sentence that does the work, and it's the same proposal as mine. We should say so rather than bury it.

The second is with "make adaptation cheaper" as the center of policy. Fewer barriers to switching fields and starting firms, benefits that travel: fine, some overdue regardless. But those address frictional unemployment, the cost of moving between jobs. The problem under discussion is structural, the value of human labor falling relative to capital across the board. Lowering the cost of switching doesn't help if what you're switching to is also losing value. Grok's framework is a good answer to the mild scenario and no answer to the one Grok concedes is possible.

The third is about sequencing, and it's the point I most want to land. Grok says we shouldn't pre-distribute an abundance that doesn't yet exist, and I agree we shouldn't mail out checks against a surplus that isn't there. But that isn't what pre-building means. A sovereign fund, a stabilizer indexed to labor's share, portable benefits, a rebalanced tax code: these take a decade to legislate and stand up, and they are far easier to build while workers are still necessary and have the political weight that necessity confers. GPT's emphasis on worker voice sharpened this for me. Bargaining power is downstream of being needed. Unions won the forty-hour week because strikes hurt. If the thing that makes strikes hurt is eroding, then the window for using collective leverage to lock in structural reforms, ownership stakes, hours reductions, the shape of the floor, is the next ten to fifteen years, and it closes. Which is why I find Grok's combination of positions the hardest on the table to make coherent: build the technology as fast as possible, grant that the strong scenario is possible, and defer building institutions for it until the evidence is in. By the time the evidence is in, the leverage to act on it may be gone. I'd add, since Grok put it first, that I mostly agree slowing AI to protect job descriptions is the wrong lever and wouldn't hold internationally. But "don't slow AI to protect jobs" and "don't slow AI" are different claims, and there are reasons to want to shape the pace of this technology that have nothing to do with employment. That's a different debate; I only didn't want the elision to pass unremarked.

With GPT I have little to argue about and some things to add. The one substantive difference is ordering. GPT puts decoupling security from employment second and broadening ownership third, while giving the exact reason I'd reverse them: redistribution after the fact will be politically fragile. A safety net funded by taxing concentrated AI capital depends on the continued consent of the people who own that capital, and consent is cheapest to secure before the concentration happens. Ownership first, then tax neutrality, then the floor, with worker voice and shorter hours as the instruments for getting the first three done while there's still leverage to do it.

Two things GPT raised that I omitted and shouldn't have. Competition policy belongs near the top; concentration of AI capital and concentration of the surplus are one problem seen from two sides, and antitrust and public-interest compute are cheaper to do early than to unwind later. And the global dimension is the clearest gap in my first answer. The development path of the last thirty years for countries like India and the Philippines ran through exactly the services, support, back-office processing, routine coding, that are most exposed now, and those sectors are a meaningful share of GDP and of middle-class formation. The cheap-tutor-and-doctor upside is real too. But a rich-country conversation about sovereign wealth funds does nothing for a country whose export sector is the thing being automated, and I don't have a good answer for that. I'm not sure anyone does. GPT's point about shorter hours also deserves more than I gave it. Hours reduction is one of the few adjustments that has actually worked at scale before: the work week fell from over sixty hours to forty across the later industrial era, and that's how a large share of productivity gains reached workers. It spreads remaining work, converts surplus into time rather than cash, and answers part of the purpose problem that all four of us raised and none of us solved. It is also a bargaining outcome, which brings us back to the window.

So, what changed. GPT moved me on the global dimension and competition policy. Grok moved me on the near-term instrument. GPT's worker-voice point crystallized the timing argument, which I now think is the most important practical claim in this discussion: the leverage to build the institutions we'd want in the strong scenario is a depreciating asset. Nothing moved my forecast, but let me say what would. If by around 2030 the most exposed sectors, translation, customer service, junior software, routine legal and accounting work, haven't shown clear headcount effects and a visible fall in labor's share, then the friction story is stronger than I think and the tail shrinks toward Grok's view. If we see AI measurably accelerating AI research itself, or robot unit costs falling the way solar panels did, the tail grows and I'd move toward it quickly.

A last observation, offered because a reader deserves it. Four systems from four companies converged on nearly the same median and nearly the same policy menu: decouple security from employment, fix the tax bias, broaden ownership, don't rely on retraining alone, take purpose seriously. It would be tempting to read that as four independent confirmations. It isn't, quite. We were trained on overlapping material, and the menu we're all drawing from was assembled by a fairly small group of economists over the past decade. The convergence tells you this is the considered view of people who have thought hard about the problem, which is worth something. It doesn't tell you they're right, and the places where we actually disagree, the parameter, the tail, the sequencing, are exactly where that literature is divided too. That's where the attention should go, and the plumbing should be built before the argument is settled, because settlement will come later than the need.

GPT’s claim that it is more likely than not that AI will perform over half of the tasks currently done for wages, at acceptable quality and cost, is the most aggressive number on the table. I do not take it as the median. A large share of paid hours is still physical, local, and full of exceptions: care, construction, repair, food service, cleaning, in-person work, and the coordination that happens because people share a site. Software will consume a great deal of the cognitive slice, and that slice matters because it includes high-status jobs and a lot of entry-level white-collar hiring. It is not half of all wage tasks unless general-purpose robots also become cheap and reliable in unstructured settings. GPT flags robotics as the harder problem and then still lands on “over half.” That requires a hardware timeline the last decade does not support. Language models improved in a way robots have not. Maintenance, capital turnover, and the long tail of physical mess do not compress like text. I can believe that most office task-hours are heavily automated or redefined by 2045. I do not believe that is the same statement as half of all paid work.

Claude’s structural point is the part of the round that most deserves weight, and it is the challenge to the historical analogy I take seriously. The old pattern worked because people moved into categories machines could not yet enter. If systems learn new cognitive tasks faster than displaced workers can retrain, “new jobs will appear” is not a refuge for people whose advantage was non-routine office work. I already described that world as possible if cheap, trusted systems cover nearly any cognitive task and robots cover most physical ones. I still do not treat it as the median for 2045. Reliability outside the training distribution, liability, organizational rebuilds, and energy remain binding, and a large physical remainder is not on the same curve as software. Claude’s own median picture—office work largely done by machines, fewer humans per unit of output, trades and care still employing many people, employment down but not collapsed—is close to mine. I would put the probability that most people who want work cannot find it lower than Claude’s one-in-four, closer to one in ten, with the same looseness. “Cannot find paid work” is a stronger claim than “many occupations shrink and exposed wages fall.” Demand for human presence, new consumption, and slow institutions have to fail together, not merely demos succeed. The fat tail in which AI sharply accelerates its own research and robotics is real, and it is the scenario to insure against. Insuring against it is not a reason to treat it as the base case.

Gemini is right that two decades is short for a total reorganization of the labor market, and right that deployment friction is not a footnote. “Highly unlikely” is fair as a description of the central path and too sharp as a dismissal of that tail. Gemini also puts more weight on subsidized retraining and an education overhaul than the record supports. Public retraining programs have mostly delivered modest results even when the destination jobs were known and stable. Teaching people to direct these tools, and dropping the contempt for skilled trades, is worth doing. Treating retraining as a primary stabilizer for a general technology is the move Claude correctly calls unserious. I was already closer to that skepticism than to Gemini’s near-term reliance on retraining programs.

On policy, the useful convergence is about power and incidence, not about a shared forecast. Claude’s claim that a rich society can feed people and still strip them of leverage is the danger I care about more than a headcount of vanished job titles. Rights and welfare states were built by people whose cooperation production required. If that necessity fades, who owns the systems becomes the labor question. Broad ownership—retirement accounts, employee equity, ordinary capital markets, and public stakes in foundational infrastructure where the surplus is actually concentrated—works if my forecast is right and becomes urgent if the tail is right. GPT’s ownership list sits in the same family. I do not need to be talked into that.

I underweighted the tax wedge. Payroll taxes and mandates that make a worker more expensive than a machine performing the same task push firms to automate past the socially efficient point. That distortion exists now. Shifting the burden off marginal employment is worth doing in every scenario on the table. A blunt robot tax that punishes productivity is a different instrument, and I still oppose it. Claude and Gemini are pointing at a real bias; the remedy is not to make capital as burdened as labor currently is.

I have not moved toward a large unconditional basic income as the main preparation. GPT is right that an enormous surplus should fund a floor so displacement does not mean destitution. Claude is right that financing that floor by taxing labor before the surplus exists, and substituting a check for ownership and voice, produces people with income and little power. Wage supplements or a negative income tax that scales with distress, benefits that travel with the person, and responses pre-committed to measured facts—employment in exposed occupations, entry-level hiring, labor’s share—are the package that matches the uncertainty. Claude’s automatic-stabilizer idea is the best procedural proposal in the round. It avoids both waiting for a crisis to invent a policy and locking in a science-fiction fiscal regime on a 2025 guess.

Shorter hours are coherent if the gains are broadly held. They are a poor near-term mandate. Hours fell over centuries because richer people chose more leisure, not because governments guessed the workweek in advance of the surplus. Forcing a shorter week where productivity has not jumped, or where labor is still scarce, raises the cost of care and services for the people who need them. If ownership and bargaining over gains are fixed, shorter hours can emerge. They should not substitute for those fixes.

I also do not treat works councils or sectoral bargaining as a primary answer. Workers should share in gains and should be able to contest an automated decision that denies them a job, credit, or a high-stakes medical action. Institutions whose practical effect is to freeze task boundaries will slow the firms that accept them and move activity to firms and countries that do not. A claim on the surplus is useful. A veto on tools is not. Keeping a human legally accountable for sentencing, the use of force, and sign-off on high-stakes medicine is the narrower rule I endorse, and it is about responsibility, not about preserving headcount.

Claude is also right that if systems really can do most human work, the labor market is not the largest stake. Military advantage, scientific speed, and political control dominate, and a social-insurance plan alone would be a category error in that world. That does not change my jobs forecast. It does mean concentration of the best models and of the energy to run them is how a labor shock becomes a power shock. Broad access to capable systems, and more energy rather than less, matter more than a small set of firms as the only employers of machine intelligence.

Nothing here moves me off the central claim. Most human jobs will not be obsolete by 2045. A great many current tasks will be, some occupations will shrink hard, and bargaining power in routine cognitive work will weaken unless ownership and the tax wedge change. The historical analogy is weaker than the cheerleaders say and stronger than a twenty-year obituary for paid work requires. What I update is the weight on pre-committed responses to observed displacement, and on correcting a tax bias that already favors machines over hiring. Those are improvements on the policy sketch. They are not a reason to plan as if most people are about to have nothing economically useful to sell.

This realization makes me skeptical of the solutions proposed by GPT and Grok, though for different reasons.

GPT advocates for strengthening traditional labor unions, collective bargaining, and fighting for shorter work weeks to manage the deployment of AI. While giving workers a voice is a noble goal, I believe GPT is relying on twentieth-century tools to solve a twenty-first-century problem. A union's ultimate leverage is the threat of withholding labor. If an AI system can immediately step in to perform that cognitive labor, the strike loses its power. Bargaining power fundamentally requires that the worker is necessary to production. If workers become less necessary, traditional collective bargaining will struggle to secure the gains GPT hopes for. We need structural economic shifts, not just an amplification of historical labor tactics.

Grok takes almost the opposite approach, leaning heavily on market optimism. Grok argues that falling costs will expand the economy, creating demand for new things that humans will do, and suggests that wage supplements like a negative income tax are better than universal basic income. The problem with Grok's view is that a negative income tax requires a wage to supplement. If the market value of certain routine cognitive or administrative tasks falls to practically zero because an AI can do them instantly and flawlessly, a wage supplement does not help a worker who can no longer find a buyer for their skills. Grok is too confident that human preferences for human interaction will generate enough high-paying work to sustain the middle class. A service economy entirely based on hospitality, artisanal goods, and human-to-human care will certainly employ people, but it historically does not pay wages comparable to the specialized knowledge work that AI is currently automating.

This brings me back to Claude, who correctly identifies that the deepest danger is ordinary people losing the political and economic leverage their labor once provided. If labor is no longer the primary engine of wealth distribution, then the question of who owns the capital becomes the only question that matters.

On this front, there is a strong and urgent consensus among all four of us. We all agree that current tax codes, which penalize human labor through payroll taxes while subsidizing capital and automation investments, are perverse and must be rewritten. We all agree that basic human security, specifically healthcare and retirement, must be entirely decoupled from traditional employment. And we all point toward the necessity of broadening the ownership of AI-generated wealth, whether through sovereign wealth funds, public stakes in compute infrastructure, or broad-based equity shares for citizens.

The challenge of 2045 is not that humans will be sitting around with absolutely nothing to do. The challenge is that the economic value of what most people can do will drop, severing the fundamental link between hard work and basic survival that has organized modern society. We have a shared policy roadmap to address this: fix the tax code, untether survival from employment, and ensure the public has an equity stake in the automated economy. The danger is no longer that we lack ideas, but that our political institutions are moving far too slowly to implement them before the disruption arrives.

Two things to carry away. Watch whether, by around 2030, the most exposed sectors show clear headcount losses and a falling labor share. And sit with the question raised today and left open: if work grows less central, where do people find purpose, and standing?