Do markets price things sensibly? Not always.

The occupations that artificial intelligence finds hardest to touch - caring for the elderly, teaching children, supporting people with disabilities - are also the occupations we have chosen to pay the least: A home health aide in the United States earns a median of $34,900 a year. A teacher's assistant earns $35,240. Both sit nearly $25,000 below the national median wage, for work that no algorithm can replicate and no society can do without. Meanwhile, the jobs AI is hungriest for, software development, financial analysis, legal research, are the jobs we pay handsomely. Developers earn a median of $130,160; financial analysts $99,890.

The inverse correlation between automation vulnerability and compensation is not a coincidence. It is a market failure introduced by bias, accumulated over decades, that the AI transition is now making impossible to ignore.

This is the automation paradox. And unlike most paradoxes, this one has consequences that compound.

What makes care and education genuinely hard to automate

The canonical framework for understanding this comes from Frey and Osborne's 2017 study, which identified three "engineering bottlenecks" to computerisation: perception and manipulation, creative intelligence, and social intelligence.

Care work sits behind all three. A home health aide must read a patient's non-verbal distress, adapt a care plan in real time, and maintain the kind of relational trust that is built across months of daily presence.

Same for the education jobs: a teacher must hold a room of thirty children simultaneously, diagnose each student's confusion differently, and supply the motivational scaffolding that separates learning from mere information transfer.

Empirical evidence supports. The more recent GPT-exposure literature, led by Eloundou, Manning, Mishkin, and Rock at OpenAI, confirmed the pattern using large language models as the evaluator. Their 2023 analysis found that the jobs exposed to LLMs across more than half their tasks are heavily in white-collar, high-wage professional work. Personal care, education support, and direct care barely register.

Three reasons the market got the price wrong

The undervaluation of care and education work is not new. Nor is it accidental. At least three structural mechanisms explain why the price signal failed.

First, positive externalities go unpriced. A teacher who produces a curious, socially capable adult generates value that accrues to the student's future employer, to the tax base, to the civic fabric. None of that value flows back to the teacher's wage. Care work is similar: the home health aide who keeps an elderly person out of a hospital bed saves the healthcare system tens of thousands of dollars in avoided acute care. That saving is not captured by the aide's employer, who is often a Medicaid contractor operating on thin margins.

Second, these are structurally monopsonistic labour markets. School districts, hospital systems, and Medicaid-funded care agencies are often the dominant - sometimes sole - buyers of care and educational labour in a given geography. Monopsony depresses wages below the competitive level even when workers are productive and in demand. The result is persistent undercompensation that persists when the buyer is powerful.

Third, the historical feminisation of these occupations anchored wages at a level that compounded across generations. PHI's research on the direct care workforce found that 85% of home care workers and 91% of nursing home assistants are women, and 67% are people of colour. Research on wage anchoring consistently shows that wages set when a workforce was predominantly female, and minority tend to persist even as the demographic rationale erodes, a form of institutional inertia baked into collective agreements, job classifications, and public reimbursement rates.

Occupations are constructs, not facts

David Autor's research offers a clarifying lens. Most jobs Americans hold today did not exist in 1940. Among health services workers alone, 85% of current roles have emerged since then.

Occupations are not discovered - they are built, and the terms on which they are built determine who gets paid what.

The mechanism that turns a cluster of tasks into a well-compensated profession is institutional: licensing, credentialing, professional associations, and political organisation. The clearest case is medicine. Before the 1910 Flexner Report commissioned by the Carnegie Foundation and effectively written in advance by senior AMA officials, mechanics earned more than doctors. The AMA used the report to push state governments into closing more than half of American medical schools, restricting supply, and handing licensing authority to AMA-controlled boards. As a result, physician incomes were permanently elevated by institutional construction.

Care work and teaching never got that moment. When men abandoned teaching after 1840 for better industrial wages, school boards didn't raise pay to compete, they recruited women precisely because they could be paid less. The same logic shaped home care and personal services: work coded as an extension of domestic virtue, performed disproportionately by women and workers of colour, was never granted the associations strong enough to lobby or the credentials rigorous enough to restrict supply. After 1980, new work bifurcated into high-paid professional and low-paid service roles. Care and teaching ended up in the second category, not because of their skill content, but because of who had built them, and who had not.

This maps directly onto the three mechanisms above. The measurability problem left no credentialing substrate on which to build a profession. The substitution illusion removed the pressure to professionalise. And the monopsony structure of public funding eliminated the price signal that scarcity might otherwise have triggered.

Autor's 2024 research closes the loop. Historically, even poorly constructed occupations could eventually be lifted by the broader tide of new, well-paying job creation. That tide is weakening: the demand-eroding effects of automation have intensified over four decades, while the augmentation side has not kept pace.

The escalator that once lifted new occupations into the middle is breaking down, and with it, the passive correction mechanism that might have eventually repriced care and teaching work upward.

Without deliberate intervention, there is no path. No professional association powerful enough to lobby, no credentialing system robust enough to restrict supply, no wage floor backed by sustained public investment. Kraft and Lyon's 2024 data captures the result: teacher prestige, interest, and preparation are at fifty-year lows; real wages sit at their lowest since 1987; the cost of a teaching degree now consumes 27% of a starting salary, up from 10% in the early 1980s. The pipeline is draining precisely as demographic pressure builds.

The market did not create this problem. It will not fix it.

The contradiction is becoming a crisis, NOW

Previous automation transitions unfolded over decades, giving labour markets and policy frameworks time, however imperfect, to adjust.

This moment is different in kind.

Displacement is already measurable. Erik Brynjolfsson's Canaries in the Coal Mine study, using ADP payroll data, found a 13% relative employment decline for early-career workers in the AI-exposed occupations. The roles being hollowed out are software development, customer support, and paralegal work, precisely those that middle-class households relied on as stable entry points. Home health aides were barely touched.

The demographic pressure compounds this. The U.S. population of adults aged 65 and older is projected to grow from 57.8 million today to 88.8 million by 2060; the 85-and-older cohort will nearly triple. The direct care workforce is already projected to add more new jobs than any other single sector between 2024 and 2034, over 772,000 positions. AI is compressing the timeline on one side while ageing population stretches the demand horizon on the other. The workforce development system sitting between these two pressures was not built for this pace. Yet, nearly half of existing direct care workers live in or near poverty, and 49% rely on public assistance.

Demand is surging. The work is irreplaceable. The wages are not moving. This is an absence of signal, and the window to address it is narrowing.

What a new social contract would actually require

The phrase “social contract” has been used so often that it has started to hollow out.

What would concretely mean repricing care and education work?

It means reforming how national accounts measure care work. GDP currently treats most care work as a cost to be minimised rather than a productive activity to be invested in. The unpaid care work performed predominantly by women, estimated at anywhere from 10% to 39% of GDP depending on methodology, does not appear in the accounts at all. When the state funds home care, it shows up as public expenditure rather than as investment in human capital. This accounting choice shapes every downstream policy conversation.

It means changing wage standards in publicly funded care and education sectors. The public sector is not merely a passive payer in these markets; it is often the dominant buyer. Minimum wage floors for Medicaid-funded care work, meaningful reclassification of teaching assistant roles, procurement rules that condition public contracts on living wages, these are levers that governments control directly.

It means shifting the cultural narrative from vocation to scarce skill. The language of “calling” and “passion” applied to teaching and care work has long functioned as a wage-suppression mechanism, implying that intrinsic rewards substitute for financial compensation in a way no one would accept applied to software engineering. The AI transition provides a new, economically rigorous argument for reframing this: care work is not just valuable, it is genuinely scarce in a way that white-collar cognitive work is rapidly ceasing to be.

We are at an unusual moment. For the first time in the history of automation, the labour most vulnerable to displacement is not the work of the lowest-paid workers. It is the work of the moderately and highly paid professional class. The occupations left standing, the work of care, of teaching, of attending to another person's body and development and emotional state, are the occupations we have systematically undervalued, but serves as the foundation for technical advancement.

AI has handed us an unusually clear argument for making a different choice. The question is whether we will use it.