The A.I. We Reward Is the Future We Get

A friend of mine, a clinical psychologist, once described what happened to the United States in 2020 as a “national nervous breakdown.” The phrase stayed with me because it captured something that unemployment rates, infection curves, and political polling could not: the deeper sense of a society that had reached a tipping point.

Breakdowns do not come from nowhere. They gather force over time, fed by strains that are ignored, denied, or kept just barely under control until one final pressure makes them impossible to contain. Covid became that pressure. It did not create America’s distrust, institutional decay, or political rage; it brought long-latent tensions to the surface and gave them a common stage.

I worry we are now headed toward another nervous breakdown, this time over artificial intelligence. It will not announce itself with closed schools, masked grocery aisles, or the eerie quiet of a country brought to a halt. It will arrive more subtly, through smaller ruptures: a promotion that never comes, a freelance rate that collapses, a career ladder that narrows. Over time, it could prove even more destabilizing, because it will strike at something as intimate as health: work, status, and the stories people rely on to understand who they are.

Two recent essays have named the problem with unusual clarity. In “The Messy Middle,” Molly Kinder argues that the gravest A.I. labor-market danger is not today’s relatively stable economy or some imagined post-work future in which machines do nearly everything and society somehow shares the spoils. It is the unstable passage between them, when A.I. may devalue or eliminate many knowledge-worker jobs long before robotics, social policy, and democratic institutions have caught up.

Jasmine Sun, in her New York Times essay on Silicon Valley’s fear of a “permanent underclass,” makes a related point from inside the culture building the technology. The people closest to A.I. are not merely debating abstract existential risk; many are worried about a more immediate future in which ordinary workers lose economic leverage as their jobs are automated away. Sun’s most important observation is not that this outcome is inevitable. It is that the creation of a social underclass would be a policy choice.

Kinder and Sun have identified the danger. The next question is what to do about it.

The answer cannot simply be to oppose A.I. Nor can it be to wait for mass displacement and then hope that retraining, universal basic income, or post-A.G.I. redistribution will arrive in time. The central task of the messy middle is to shape the A.I. economy before it hardens. That means distinguishing A.I. companies that help workers become more productive, skilled, and valuable from those whose business model depends on making workers unnecessary.

Not all A.I. companies are created equal. Some expand human capacity in fields where we already have shortages. Others sell themselves, implicitly or explicitly, as a way to cut payroll. Both may be innovative. Both may increase productivity. But they will not have the same social consequences.

Consider the difference between Abridge and Legora. Abridge uses A.I. to turn patient-clinician conversations into draft clinical notes, reducing documentation burden while keeping clinicians responsible for review and care decisions. In a health care system already strained by burnout and labor shortages, that kind of A.I. can help people do more of the work only humans can do: listen, judge, treat, and care.

Legora represents a different and more complicated category. Its legal A.I. platform helps lawyers review documents, conduct research, draft contracts, and automate parts of due diligence, tasks that have traditionally trained junior lawyers. Used well, such a tool could make lawyers more productive. Used primarily to reduce junior head count, it could narrow the entry-level ladder through which young lawyers learn judgment.

Those are not the same social bargain. One kind of A.I. relieves pressure in a field where society needs more human capacity. The other may make professional services more efficient while eroding the apprenticeship system that produces future experts. Public policy should not treat them as identical.

Kinder begins with two stories that are not, strictly speaking, stories about A.I. One is about a former senior USAID official who lost her job and is contemplating a much lower-paid teaching career. The other is about a laid-off semiconductor engineer now driving for Uber while facing the possibility that robotaxis may soon threaten even that fallback. Neither person was displaced by A.I. directly. But that is precisely why the examples are useful. They show what disruption feels like before it has a clean technological label: a sudden loss of income, status, identity, and plausible alternatives.

Kinder’s central claim is that the A.I. conversation often leaps from one reality to another. In the first, A.I. is useful but the labor market remains mostly intact. In the third, artificial general intelligence creates such extraordinary abundance that work becomes optional and redistribution solves the rest. Missing from this story is the second reality: a long, politically volatile transition in which A.I. reshapes professional work before society has decided how to absorb the shock.

This is the part of the future we are least prepared to discuss.

One reason is that full automation of the physical economy is likely to be much harder than the automation of laptop work. Airports, hospitals, restaurants, construction sites, nursing homes, schools, and repair jobs require physical presence, dexterity, trust, and judgment in messy real-world environments. A.I. may become very good at drafting memos, writing code, analyzing contracts, and producing slide decks well before robots can reliably care for the elderly, fix a broken pipe, manage a chaotic classroom, or work a hospital floor.

Kinder suggests that this creates a strange inversion of the pandemic economy. In 2020, laptop workers were relatively protected while essential workers had to show up in person. In the A.I. transition, the vulnerability may run the other way. Many remote professional jobs may become more exposed, while in-person service, care, and trade jobs may prove more insulated.

But that does not mean the future belongs neatly to plumbers and nurses. “Learn a trade” may be good advice for some individuals, Kinder argues, but it is not a national labor-market strategy. Trades are well paid partly because they are scarce, difficult, physical, and slow to master. If displaced consultants, lawyers, analysts, engineers, and product managers flooded into them, training systems would be overwhelmed and wages would eventually come under pressure. The country cannot retrain its way out of every structural shock, especially when the shock arrives faster than institutions can adapt.

Nor is universal basic income a complete answer. If payments are small, they will not replace the salaries, benefits, and social standing lost by displaced professionals. If they are large enough to do so, society will still need people to staff hospitals, teach children, repair homes, move goods, care for the elderly, and maintain infrastructure. Cash can soften hardship, but it cannot by itself decide who does the essential work, what that work should pay, or how a society preserves dignity when millions of people are told their skills are suddenly worth less.

That is why the distinction between worker-complementing and worker-replacing A.I. matters so much.

The difference is not always simple. Some automation will replace unpleasant, dangerous, or unnecessary work, and that can be good. Some labor-saving tools will make businesses more productive, and productivity matters. The point is not to freeze the economy in place or protect every existing job exactly as it is. The point is to stop pretending that all productivity gains are socially equivalent.

Right now, many of the incentives point in the wrong direction. A start-up can raise money by promising to replace paralegals, customer service agents, junior analysts, copywriters or coders. A corporation can justify an A.I. purchase by calculating how many salaries it will save. A venture capitalist can see labor replacement as a clean path to high margins. In that environment, founders will naturally build companies that turn human labor into a cost center to be minimized.

A serious response to the messy middle would change this incentive structure. Governments should use procurement, tax credits, grants, and regulatory preference to support A.I. companies that complement workers rather than merely replace them. Public contracts should favor systems that improve service quality, expand capacity and preserve meaningful human roles. Tax benefits should be tied not just to A.I. investment, but to job quality, training, wage growth, and worker retention. Universities, philanthropies, and public research agencies should fund human-complementing A.I. with the same seriousness that investors now fund automation for payroll reduction.

We should also ask companies to account for the labor-market effects of the systems they deploy. Before a firm receives public subsidies, sells into government, or claims to be solving a social problem, it should be able to say what its product does to workers. Does it make existing workers more capable? Does it create new roles? Does it preserve entry-level pathways? Does it share productivity gains? Or does it mainly remove people from payroll while shifting the costs of disruption onto families, communities, and the state?

This is where Kinder’s argument and Sun’s warning converge. The danger is not only mass unemployment. It is the erosion of economic leverage. A lawyer may still have a job, but fewer junior associates may be needed. A software team may still ship products, but with fewer entry-level coders. A marketing department may still exist, but with fewer copywriters, analysts, and coordinators. The firm survives, and so does the occupation. But the ladder narrows.

That narrowing could produce one of the most damaging effects of the messy middle: the collapse of apprenticeship. Many professional careers are built on junior workers doing routine work at first and gradually developing judgment. If A.I. absorbs the bottom rungs of the professional ladder, companies may save money in the short term while destroying the pipeline that produces experienced workers in the long term.

The result would not be a clean division between the employed and the unemployed. It would be a sorting process. Some workers would be amplified by A.I. Their productivity would rise, and so might their value. Others would see their bargaining power weaken as more of their work becomes cheap, automated, or bundled into software. The political consequences would come not only from joblessness but from downward mobility: smaller teams, weaker hiring, fewer entry-level opportunities, lower wages, and more competition for the remaining human roles.

Status matters here. A displaced software engineer, lawyer, policy expert, or analyst is not just losing a paycheck. That person may be losing a career identity, a class position, a sense of competence, and a credible account of the future. People can often endure hardship when they believe it is temporary, fair, or meaningful. They react differently when they believe the system has made their lives obsolete while enriching those who designed the system.

That is what makes the comparison to deindustrialization so unsettling. Factory closures devastated communities, hollowed out towns, and helped produce decades of resentment. But A.I.-driven disruption could be politically explosive in a different way. Manufacturing job loss was geographically concentrated and often culturally distant from the institutions that shape national policy. A.I. could hit lawyers, consultants, software engineers, analysts, designers, managers, journalists, and other professionals who are educated, networked, and close to power.

These are people who know how to sue, lobby, organize, write op-eds, influence regulators, and pressure politicians. If they experience a sudden loss of income and status, their backlash may arrive quickly and loudly. A professional-class version of deindustrialization would not remain confined to factory towns. It would land inside the institutions that govern the country.

This is why waiting for post-A.G.I. redistribution is not a serious plan. The promise that extraordinary future wealth will eventually be shared depends on a politics we do not currently have. The same country that struggles to agree on Medicaid, food assistance, wealth taxes, and work requirements is unlikely to glide effortlessly into a generous post-work welfare state because a handful of technology companies become even richer. If anything, the politics of redistribution may become harder as the winners of A.I. accumulate more power.

A serious transition plan would start before the crisis peaks. It would treat A.I. not only as an innovation challenge or a social insurance challenge, but as a market-design challenge. The question is not simply how to compensate people after their work has been devalued. It is how to encourage the creation of companies that do not make devaluation their core business model.

The point is not that there are “good” and “bad” A.I. companies in some simplistic moral sense. The point is that different business models produce different societies. A.I. that helps workers become more skilled, more productive, and better paid is not the same as A.I. that makes workers easier to discard. A.I. that expands access to services we lack is not the same as A.I. that removes the bottom rungs of the career ladder.

Founders respond to incentives. Investors respond to incentives. Companies respond to incentives. If the easiest way to make money with A.I. is to replace workers, many firms will be built to replace workers. If public money, procurement, tax benefits, and prestige flow toward companies that augment workers, a different set of companies will be founded.

Kinder’s deepest insight is that the messy middle is not a temporary inconvenience on the road to abundance. Sun’s is that a permanent underclass is not merely a technological outcome but a political choice. Taken together, they point toward the same conclusion: the future of A.I. work will be shaped not only by model capabilities, but by the incentives we put around them.

That is why the “messy middle” deserves more attention from the press and policymakers now, before it hardens into a crisis. A society can survive technological change; it has done so before. What it cannot survive indefinitely is a series of shocks to people’s purpose and security while being told, in effect, to wait patiently for abundance.

The way to avoid that future is not to oppose A.I. as such. It is to stop treating all A.I. companies as if they are building the same future. Some will help people work, learn, and build. Others will hollow out the institutions through which people acquire skill, status and a sense of purpose. The messy middle will be shaped by which kind we reward.

That is not a side issue. It may be the central question. Without incentives that favor human-complementing A.I., the transition will not feel like progress to millions of people. It will feel like another national breakdown.