Loss of Meaning and Purpose

The claim is not about income. It rests on a body of research going back to Marie Jahoda's studies of unemployment, which held that a job delivers latent benefits beyond wages: a shape to the day, contact outside the household, shared goals, status and enforced activity. Karsten Paul and Klaus Moser tested the resulting prediction across 237 cross-sectional and 87 longitudinal studies in the Journal of Vocational Behavior in 2009, finding an average effect size of d equals 0.51 for the difference in distress between unemployed and employed people, with 34% of unemployed people showing psychological problems against 16% of employed people, and evidence from longitudinal studies and natural experiments that the relationship runs from unemployment to distress rather than only the reverse.^[1]

The AI-specific version adds a second step. A person can remain employed and still lose the part of the job that supplied the meaning. If a model drafts the argument, writes the code, reads the scan and produces the design, what remains for the human is checking, approving and taking responsibility. Researchers at Microsoft Research and Carnegie Mellon University described exactly that shift in a study of 319 knowledge workers presented at the 2025 CHI conference: generative AI moves effort from gathering information to verifying it, from solving problems to integrating a model's response, and from doing tasks to supervising them.^[2]

A third and more distant version concerns a world where human effort is not needed for anything practical. Nick Bostrom's Deep Utopia, published in 2024, calls this the post-instrumental condition and treats the question of whether meaning survives it as the problem that remains after all other problems are solved.^[3] That version is a philosophical argument, not a measurement, and it should be kept separate from the workplace claims, which are empirical and testable now.

Public expectation is clear and negative, but expectation is not outcome. In a Pew Research Center survey of 5,023 US adults conducted 9-15 June 2025, 53% said AI would worsen people's ability to think creatively against 16% who said it would improve it, and 50% said it would worsen the ability to form meaningful relationships against 5% who said it would improve it. On problem solving the split was narrower, 38% worse against 29% better.^[4] Pew reported that the most common concern volunteered by respondents was the weakening of human skills and connections.^[4]

One study went looking for the mechanism directly. Mateja Vodiskar and Caroline Ruiner randomly assigned 677 participants to complete a writing task with or without an AI text generator, published in AI and Society in 2025. Using the generator had no direct effect on how meaningful people found the task. Mental effort was the mediator: AI changed perceived meaningfulness only through its effect on cognitive engagement.^[10] That result cuts both ways. It suggests that using a model does not by itself drain meaning from work, and it suggests that the way to drain it is to remove the part of the task that requires thinking.

The first route runs through the bottom rung. Expertise is usually acquired by doing tasks that experts find tedious: reviewing documents, writing first drafts, running the routine analysis. Those are the tasks generative systems perform most cheaply. If the apprenticeship tasks disappear, so does the path to the mastery that makes work absorbing later on. The payroll evidence from Brynjolfsson, Chandar and Chen is consistent with that pattern, showing relative employment declines concentrated in workers aged 22-25 in the most exposed occupations while experienced workers were unaffected.^[7]

The second route is supervision without authorship. A person who checks a model's output all day carries responsibility for work they did not make. The Microsoft and Carnegie Mellon findings describe that redistribution of effort, and the experimental work in Scientific Reports found reduced intrinsic motivation after periods of AI collaboration.^[2][5] Add the finding by researchers at BetterUp Labs and the Stanford Social Media Lab that 41% of workers have been sent AI-generated content that looked polished but lacked substance, costing nearly two hours of rework each time, and the social layer of work degrades alongside the individual one.^[6]

The third route is the one Jahoda's model predicts most directly: income arrives without structure. Paul and Moser found that the mental health penalty of unemployment was larger in countries with weaker unemployment protection, which suggests that money mitigates the harm without eliminating it.^[1] A society that solves the income problem through transfers and leaves the structure problem untouched would still expect higher rates of distress, particularly among men and workers in manual occupations, the groups where the meta-analysis found the largest effects.^[1]

The fourth route is the slowest and least testable. If systems eventually perform every task better than any person, then achievement, contribution and expertise stop being available as sources of self-respect. Bostrom treats this post-instrumental condition as a genuine difficulty rather than an illusory one, and locates the challenge in philosophy rather than in engineering.^[3] No experiment currently planned can settle it, which is why it belongs in a different category from the workplace findings above.

The strongest objection is that the fear romanticises existing work. John Danaher argued in Automation and Utopia in 2019 that work is a source of misery and oppression for most people, that hastening its demise would therefore be an improvement, and that the freed time could go to games, play and virtual worlds.^[11] The empirical support is partial but real. Robert Dur and Max van Lent, using representative data on 100,000 workers in 47 countries, found that about 8% consider their job socially useless and another 17% are doubtful about it, and that a large majority care about holding a socially useful job and suffer when they judge their own to be useless.^[12] Meaning is not evenly distributed across the jobs AI is displacing.

Removing the compulsion to work does not automatically produce distress. Finland's basic income experiment gave 2,000 randomly selected unemployed people a tax-exempt 560 euro a month unconditionally across 2017 and 2018, with the control group drawn from everyone else receiving the same Kela unemployment benefit. The evaluation published in May 2020 found small employment effects, and recipients reported greater life satisfaction, better perceived economic security and less mental strain, depression, sadness and loneliness than the control group.^[13] That is a test of unconditional income for people already out of work, not of a workless society, but it points against the assumption that only employment can supply psychological security.

The timeline may also be wrong. John Maynard Keynes predicted in 1930 that his grandchildren would work about fifteen hours a week. Nicholas Crafts revisited the prediction in Economica in 2022 and concluded it will not be realised, but not because leisure failed to grow: expected lifetime hours of leisure and home production in the United Kingdom rose 58% between 1931 and 2011, rather more than Keynes would have expected, taken as longer retirement at the end of life rather than as a shorter working week throughout it.^[14] Forecasts about how a society will spend its productivity gains have a poor record on timing and on form, and the version of this fear that assumes mass idleness by a particular date inherits both problems.

Being outperformed by a machine has not reliably destroyed the meaning of an activity. Minkyu Shin, Jin Kim, Bas van Opheusden and Thomas Griffiths analysed more than 5.8 million move decisions by professional Go players between 1950 and 2021, published in the Proceedings of the National Academy of Sciences in 2023. Players began making significantly better decisions after superhuman programs appeared, and novel moves became more frequent. The authors put the causal reading carefully: superhuman AI may have prompted players to break from traditional strategies, which in turn may have improved their decision making.^[15] Human Go did not stop mattering to the people who play it when a program could beat them all.

Finally, the direction of the workplace evidence is not settled. Anthropic's Economic Index report of 26 June 2026, based on its own users, reported that people who delegated a larger share of sessions to Claude were more optimistic about pay, job security, meaning and autonomy, that 57% said AI had made their skills more valuable, and that 68% said they were learning more, with learning flat across levels of delegation.^[16] Those are self-reports from roughly 9,700 self-selected users of one company's product, which is a serious limitation, but they are evidence against the strong version of the deskilling story. David Autor's argument in a 2024 National Bureau of Economic Research paper points the same way: AI could extend the reach of human expertise, letting more workers do the judgment-heavy tasks now reserved for elite professionals, which would increase rather than reduce the number of jobs containing discretion.^[17]

No jurisdiction regulates for meaning, and it is not obvious what such a rule would say. The closest existing instruments give workers notice and a say. Article 26(7) of the EU AI Act requires deployers who are employers to inform workers' representatives and the affected workers before putting a high-risk AI system into service at the workplace, and Article 26(11) requires that people subject to Annex III systems used in decisions about them be told.^[18] Notice is thin, but it converts an invisible change in the content of a job into something that can be bargained over.

The second live channel is income policy tested empirically rather than asserted. Finland's experiment is the best documented example of a government evaluating unconditional income against a control group and publishing wellbeing outcomes alongside employment outcomes.^[13] Its lesson for this fear is methodological as much as substantive: wellbeing was measured, not assumed, and the measured effects on employment and on mental health pointed in different directions.

Measurement of how AI is actually used at work is improving. The Anthropic Economic Index publishes recurring reports classifying interactions into automation and augmentation patterns and, since 2026, pairs them with user surveys on pay, security, meaning and autonomy.^[16] Gallup continues to track engagement and wellbeing on a consistent global panel running back to 2009, which gives a baseline that predates widespread generative AI use.^[9] The SHRM survey of 5,875 US workers in March and April 2026 provides comparable data on how workers themselves rate the threat to their field.^[8] None of these were designed to measure meaning, and building instruments that do is the obvious gap.

At the level of practice, the most concrete responses have come through collective bargaining and professional bodies rather than legislation, where the question is who gets to remain the author of the work. The evidence on which parts of a job supply meaning is directly relevant here: Vodiskar and Ruiner found mental effort to be the channel through which AI use reaches meaningfulness at all, which implies that where a system is inserted into a task matters more than whether it is used.^[10]

Treat the entry-level rung as the leading indicator rather than the aggregate unemployment rate. Total employment can be stable while the tasks that build expertise disappear, and the psychological claim in this fear concerns the content of work rather than its quantity. The payroll evidence on 22-25 year olds is the closest thing to an early warning that exists, and it should be checked against later data rather than treated as settled.^[7]

Be sceptical in both directions about survey evidence. Public expectation that AI will damage creativity and relationships is strong and consistent, but people are poor forecasters of their own adaptation.^[4] Equally, favourable self-reports from heavy users of a product are affected by who chooses to use it.^[16] The findings worth weighting most are the ones with a control condition and a behavioural measure, which at present are few.

The philosophical version and the workplace version should not be scored together. Whether meaning is possible in a world without necessary work is an open question that Bostrom, Danaher and others treat seriously and answer differently.^[3][11] Whether checking model output all day is less satisfying than writing the report yourself is a question that can be settled with a field experiment, and has not been.