AI’s Impact on the Labour Market: Too Many Forecasts, Too Little Clarity

Three recent accounts of AI and work point in three different directions: a research paper finds almost no displacement, a consultancy finds a labour market splitting in two, and a tech founder predicts a labour shortage while his own company sheds tens of thousands of roles. Each rests on a different measure, and each measure was built to capture something narrow. What follows reads them against one another, and asks why the public conversation they feed is so much more confident than any of them.

In March 2026, Anthropic published a paper on the influence of AI on the labour market, proposing a conceptual framework for understanding how — and to what extent — artificial intelligence might reshape it. Authored by Maxim Massenkoff and Peter McCrory, the study introduces a new metric called observed exposure, which asks a narrower question than earlier measures: of the tasks that LLMs could in principle speed up, which ones are actually seeing use in professional settings? The framework rests on three data sources: Anthropic’s internal usage statistics, the O*NET database (which catalogues the tasks associated with around 800 occupations in the United States), and a 2023 paper by Eloundou et al. that scores each task according to whether an LLM could plausibly perform it autonomously, semi-autonomously (with the aid of additional tools or software), or only with substantial human assistance.

The paper’s central finding is cautiously framed: there has been no systematic increase in unemployment among highly exposed workers since late 2022, although there is suggestive evidence that hiring of younger workers, those aged 22 to 25, has slowed in the most exposed occupations. The authors are careful about what this does and does not establish. They calibrate the method and report that it would detect a doubling of unemployment in the high-exposure group, from roughly 3% to 6%: it is built to catch a large shock, and a null result therefore means that no large shock has been detected, not that nothing is happening. Anthropic also acknowledges that its framework cannot capture every channel through which AI might reshape the labour market — a humility largely absent from how the study is being repackaged in popular commentary.

Two distinct things are worth separating here, because they are routinely run together.

1. The first concerns the outcome being measured rather than the method: unemployment.
   A worker who becomes a supervisor or overseer of AI output has not become unemployed, yet has plainly undergone a transformation — one that may involve wage reduction, reallocation across functions, or the erosion of professional identity. A framework keyed to unemployment will not register this. Nor is it self-evident that those who once performed a task will be retrained to monitor an AI doing it instead, or that the resulting role carries equivalent standing: the worker who oversees the tool depends on it, and on the firm that owns it, in a way the worker who performed the task did not.
2. The second is a question about the method itself, and it is the more contestable of the two: whether work decomposes into discrete tasks at all.

Practical expertise is a form of transmissible knowledge that, while expressed through the execution of tasks, does not reduce to their mere aggregation. Otherwise, expertise would collapse into the metric of execution speed. This task-additive view lends labour a markedly mechanical, almost Fordist quality: the worker becomes the figure who assembles components within a fixed time, pulls a lever, or shunts items along a conveyor belt — an apt description of Amazon warehouse staff, whom Chris Smalls, founder of the Amazon Labour Union and author of When the Revolution Comes: A Fight for the Future of the Working Class (Pantheon), has described as reduced to conditions approaching servitude. The reference is not decorative. If work is modelled as a sum of discrete operations, the model will describe some workplaces very well — warehouse floors organised on precisely that principle — and will quietly assume that all work is like that. What it cannot register is what is lost when work that was not like that is reorganised until it is.

Neither observation is a flaw in the paper. Observed exposure is an indicator of one economic phenomenon, deliberately scoped, and its authors are explicit about what it does not cover. The difficulty begins when a narrow measure travels: what leaves the paper as no detectable rise in unemployment among exposed workers arrives in public commentary as AI is not taking jobs.

Splitting the job market in two

If the trouble is that one measure cannot see everything, the obvious response is to look at another. PwC’s 2026 Global AI Jobs Barometer, released in mid-June and based on the analysis of around a billion job advertisements across 27 countries, approaches the question from the opposite end: not from what AI can do, but from what employers are advertising for. It divides the labour market into two macro-categories: “professionalised” roles, in which AI acts as a force multiplier for experts and intensifies demand for human-intensive skills, and “democratised” roles, which become easier for non-experts to perform. The former are growing twice as fast in headcount and roughly 42% faster in wages than the latter. The report also documents a striking divergence at the entry level: in its analysis of US data, AI-exposed junior roles are seven times more likely to require traditionally senior-level skills such as judgement, leadership and creativity, and have expanded by 35% since 2019, while other entry-level roles have contracted by 10%.

As Pete Brown, PwC’s Global Workforce Leader, observes: “The traditional relationship between experience and expertise is changing. AI is removing some of the routine work that once acted as an apprenticeship, while increasing demand for judgement, leadership and adaptability much earlier in careers. Organisations need to rethink how they develop talent if they want people to thrive in this new environment.”

The diagnosis is plausible, yet it leaves the underlying problem largely intact. If less-qualified work is restructured from its apprenticeship base, the distinction between an assembler of engine parts and the worker who supervises an AI performing the same job is not so easily drawn — both occupy a similar position of dependency on the tool and on the firm that owns it. That dependency is the thread both measures keep missing: it does not register as unemployment, and it does not register as a change in the skills an advertisement asks for.

While Anthropic concludes that AI is not, on present evidence, displacing workers, Jeff Bezos’s recent remarks at the VivaTech conference in Paris are more arresting. The Amazon founder argued that AI will not cut jobs but will instead create a labour shortage, by removing the technological constraints that have prevented entirely new categories of work from emerging. This was said even as Amazon’s CEO Andy Jassy described an ongoing shrinking of the company’s corporate workforce, with more than 30,000 roles cut in 2026 alone; data from Challenger, Gray & Christmas indicate that around 40% of the 97,000 layoffs announced by US employers in May 2026 were linked to AI-related restructuring.

The juxtaposition is not incidental, and it is worth saying what it shows.
Bezos predicts a labour shortage; his own company is shrinking; and the figure most often cited against him records what employers said about their own decisions, not any independently verified cause — which is why the concern about “AI washing”, firms attributing to automation what weak demand or over-hiring would explain equally well, is hard to dismiss. Every number in this debate is produced by a party with an interest in what it shows, and no two of them measure the same thing. That is the real difficulty: not that the evidence points one way rather than another, but that the most quoted figures are the least verified.

AI, job market, and the accuracy of information

Taken together, these signals suggest a more complicated picture: AI appears to augment specialist work while transforming, or thinning out, the labour of skilled trades, potentially reducing operatives to a small cohort of supervisors who are themselves easily replaced. That is a reading of present evidence rather than a forecast, and it should be held as loosely as the forecasts it is offered against. Yet the dichotomy between specialist technicians and manual labourers is reductive. White-collar employees may end up no better positioned than industrial workers, since AI is already strong in management and decision-making. The services sector — sales agents, financial intermediaries, estate agents — is a particularly under-examined blind spot: a category that the professionalised/democratised divide fails to capture.

Settling for the political slogan that AI will not cut jobs but will create new and better ones is the icing on a cake whose recipe even the leaders of Big Tech do not appear to know. The result is a public that remains in the dark, and a public anxiety about AI that grows precisely because the loudest voices are reductive. Better information — both in education and in mainstream commentary, and from analysts with genuine command of the subject — would allow people to form a sharper picture of what is changing, and to make different choices about training and life accordingly.