Stanford research finds AI is reshaping, not erasing, the bottom of the labour market
A new working paper from Stanford economist Lukas Althoff suggests AI is reshaping the work of lower-skilled employees rather than replacing them outright, complicating the dominant narrative of mass technological unemployment.
The working paper lands at a moment when the political class in Washington and Brussels has spent eighteen months rehearsing one version of the future: that generative AI will hollow out the labour market from the middle, sweeping through call centres, paralegal pools and back-office clerical work, and leaving behind a smaller, more polarised workforce of highly paid specialists and everyone else.
Lukas Althoff, a Stanford economist, has now put a counter-weight on the scale. In research circulated on 14 July 2026 and reported by the Phys.org network, Althoff concludes that artificial intelligence is more likely to reshape jobs than to eliminate them outright, with the most counter-intuitive effects showing up at the lower end of the skill distribution, not the middle.
The bottom of the ladder gets a raise
The intuition is unfashionable, which is part of why it is worth taking seriously. If machines can read a contract or draft an email, the obvious worry is that the clerical worker who used to do that loses their job. Althoff's finding pushes the other way: where AI lifts the productivity of a less-skilled worker, the worker captures more of the surplus than a highly paid specialist would, because there is more slack to take up in the first place.
In plain terms, the carpenter, the junior nurse's aide, the line cook and the apprentice electrician were already paid less than the value their work generated. AI that helps them work faster or with fewer errors narrows that gap. The senior architect or partner-track lawyer, by contrast, was already operating close to the ceiling of what their employer could pay them; an AI co-pilot trims their workload without lifting their pay cheque in the same proportion.
That is not a guarantee. It is a mechanism. Whether the mechanism survives contact with the real economy depends on bargaining power, on whether productivity gains are shared with workers or skimmed by employers, and on the regulatory choices still to be made in Sacramento, Brussels and Beijing.
Why this cuts against the dominant framing
The default story on AI and jobs has been a story about middle-class erosion. It has shaped White House talking points, European Parliament hearings, and the script of every management consultancy deck since ChatGPT's public release. The argument runs that routine cognitive work disappears first, then the displaced workers compete downward into service jobs that AI can also touch.
Althoff's data challenges that ordering. The early productivity gains, in his reading of the evidence, are concentrated in tasks done by the people who were already the cheapest to hire. For a stretched employer in home care, construction or hospitality, even a modest AI lift is transformative, because the labour they are augmenting is genuinely scarce.
The counter-narrative, advanced by some labour economists and trade-union analysts, is more pessimistic. They argue that AI's first victims will be the lower-skilled workers in the most exposed sub-sectors: translators, junior coders, copywriters, telemarketers, customer-service agents. In that reading, Althoff is right about the long-run reshaping but wrong about the near-run displacement, and the two effects will be felt simultaneously rather than sequentially. The paper does not yet adjudicate between these views. It is a working paper, not a verdict.
The structural frame, in plain language
There is a pattern here that goes beyond AI specifically. Every general-purpose technology of the last century, from the electric motor to the spreadsheet, ended up redistributing income inside firms rather than simply shrinking the headcount. The puzzle each time is whether the workers or the owners capture the surplus.
That is a political question as much as a technical one. It turns on minimum wage floors, on the density of collective bargaining, on antitrust enforcement, on whether a platform economy routes gains to workers or to the firms that intermediate them, and on the tax-and-transfer regime that catches whatever the labour market does not. None of that is written into the model. Althoff describes a mechanism; the political system decides who benefits from it.
What to watch next
Three dates matter. The first is the peer-review cycle of the Stanford paper itself; the working-paper stage is where ambitious claims live and die, and several of the economists cited in the surrounding debate will look for replication on independent firm-level data. The second is the autumn 2026 release of US Bureau of Labor Statistics occupational projections, which will give the first official read on whether the early AI rollouts have touched headcounts in the exposed clerical sub-sectors. The third is the European Commission's consultation on AI-in-the-workplace rules, expected to close in early 2027, which will set the disclosure regime for firms using AI to set wages, schedules or performance targets.
What remains genuinely uncertain is the speed of substitution. AI tools that assist a worker are easy to deploy; AI tools that replace one are harder, because most occupations are bundles of tasks rather than single functions. The honest reading of the evidence right now is that the technology augments before it displaces, and that augmentation, for once, looks most generous to the people at the bottom of the wage ladder. How long that window stays open is a question for regulators and union negotiators as much as for engineers.