The question of what artificial intelligence is actually doing to employment has generated more heat than light over the past few years. Sweeping predictions of mass displacement sit alongside equally confident claims that AI will create more jobs than it destroys, and most of these claims share a common flaw: they are not grounded in current labour market data. A recent policy brief from Stanford's SIEPR, authored by economists Neale Mahoney, Erika McEntarfer, and Karsen Wahal, attempts to cut through this noise with an empirically grounded assessment. The result is a document worth reading carefully, both for what it finds and for the methodological questions it raises.
The Empirical Baseline: What Labour Markets Are Actually Doing
The brief's core contribution is its insistence on anchoring the AI-and-jobs debate in observable data rather than theoretical projections. McEntarfer's background as the outgoing Commissioner of the Bureau of Labor Statistics gives the team unusually direct access to the machinery of labour market measurement, and that institutional knowledge shows. Rather than extrapolating from task-based exposure models, the authors examine what aggregate employment figures, job posting data, and sectoral trends are actually recording right now.
The picture they describe is, at this point in time, one of relative stability at the macro level. Headline unemployment figures have not shown the kind of structural rupture that more alarmist forecasts would predict. This is consistent with what several other labour economists have observed: the diffusion of general-purpose technologies into production processes tends to be slow, uneven, and mediated by institutional factors that pure capability assessments ignore. The history of electrification and computerisation both support this reading. The fact that a technology can perform a task does not mean firms will immediately reorganise around that capability, particularly when complementary investments in training, workflow redesign, and organisational change are required.
That said, aggregate stability can mask significant distributional turbulence beneath the surface. The brief is careful to flag this. Even if total employment remains broadly stable, the composition of work, the wage distribution across occupations, and the geographic concentration of displacement effects can shift substantially without moving headline numbers in ways that trigger alarm.
Separating Capability Claims from Deployment Reality
One of the more analytically useful moves in the brief is its explicit distinction between what AI systems are capable of doing in controlled benchmark conditions and what is actually being deployed at scale in workplaces. This gap is frequently elided in popular discourse, and it matters enormously for any serious labour market analysis.
Benchmark performance on coding tasks, legal document review, or medical image analysis tells you something about the frontier of AI capability. It tells you very little about the pace at which those capabilities are being integrated into production environments, the degree to which they are substituting for rather than augmenting human labour, or the extent to which firms are capturing productivity gains by reducing headcount versus expanding output. The authors are right to treat these as empirically separate questions.
This connects to a broader methodological problem in the AI-and-labour literature. Studies that estimate automation exposure by mapping occupational task descriptions onto model capabilities, such as the Acemoglu and Restrepo task framework or the more recent GPT-exposure indices developed by Eloundou and colleagues, are measuring potential displacement, not actual displacement. They are useful for identifying which workers face the greatest structural risk, but they should not be read as forecasts of near-term employment outcomes. The SIEPR brief implicitly endorses this distinction, though it could be made more explicit.
Where the Data Does Show Stress
The brief does not present an entirely sanguine picture. Several sectors and occupational categories are showing signs of genuine disruption, and the authors identify these with appropriate care. A few patterns stand out:
- Knowledge work at the routine end of the spectrum is seeing measurable pressure. Tasks that involve information retrieval, summarisation, first-draft generation, and structured data processing are precisely the tasks where current large language models perform most reliably, and job postings in roles heavily weighted toward these tasks have shown softening in some analyses.
- Entry-level positions in white-collar sectors appear more exposed than mid-career roles, which has significant implications for the pipeline of skill development. If junior roles in law, finance, and software development shrink because AI can handle the work those roles previously trained people to do, the long-run supply of experienced practitioners in those fields may be affected in ways that are not yet visible in current employment data.
- Freelance and gig-adjacent knowledge work, including categories like copywriting, basic graphic design, and certain categories of software contracting, has shown more visible contraction. These workers often lack the institutional buffers that salaried employees have, making them earlier and more legible indicators of AI-driven substitution.
These are not trivial effects, even if they do not yet show up as macroeconomic shocks. The distributional and lifecycle implications deserve serious policy attention.
Methodological Limitations and What the Brief Cannot Tell Us
A critical reading of the brief requires acknowledging what it cannot do. Labour market statistics, even excellent ones, are inherently backward-looking. The BLS data that underpins much of this analysis captures what has already happened, not what is in the process of happening. Given that the most capable AI systems have been widely deployed for only two to three years, and that organisational adaptation to new technologies typically operates on multi-year timescales, the current data may simply be too early to show effects that are already in motion.
There is also a measurement problem specific to AI-driven productivity changes. If AI allows a firm to produce the same output with fewer workers, that shows up clearly in employment figures. But if AI allows the same number of workers to produce significantly more output, the employment effect is neutral or positive while the distributional effect, specifically who captures the productivity surplus, depends entirely on labour market bargaining dynamics. Standard employment statistics are largely silent on this second scenario, which many economists consider the more likely near-term pathway.
The brief would benefit from a more explicit engagement with the endogeneity of firm-level adoption decisions. Firms that adopt AI aggressively are not a random sample. They tend to be larger, more capital-intensive, and operating in sectors with higher margins. The employment effects observed in early-adopting firms may not generalise to the broader economy, and selection effects can make aggregate data misleading in either direction.
Policy Implications and the Road Ahead
The brief's policy conclusions are measured and, on balance, sensible. The authors resist both the fatalism of inevitable mass unemployment and the complacency of assuming that markets will automatically route displaced workers into new productive roles. The historical record on technology transitions suggests that adjustment costs are real, unequally distributed, and often persistent at the community level even when aggregate outcomes look acceptable.
From a policy design perspective, the most actionable implication is the need for much better real-time labour market data infrastructure. The current statistical apparatus was built for an economy where structural change happened over decades. AI diffusion, if it accelerates as many expect, may compress those timescales significantly. Investing in higher-frequency, more granular occupational and sectoral data collection is not a glamorous policy intervention, but it is a prerequisite for any evidence-based response to whatever disruption does materialise.
The brief also points, implicitly, toward the importance of distinguishing between cyclical and structural unemployment in AI-affected sectors. Standard unemployment insurance and active labour market programmes are designed around the assumption that displaced workers can retrain for similar roles in the same or adjacent sectors. If AI systematically degrades the value of certain cognitive skill sets, that assumption breaks down and the policy toolkit needs to be redesigned accordingly.
What the SIEPR team has produced is a valuable corrective to the extremes of the current discourse. The data, read carefully, does not support either the apocalyptic or the dismissive reading of AI's labour market effects. What it does support is a conclusion that should motivate serious researchers and policymakers alike: we are in an early and genuinely uncertain period, the effects are already differentiated by occupation and sector, and the statistical infrastructure we rely on to track what is happening needs urgent improvement. The most intellectually honest position right now is not confidence in any particular outcome. It is a commitment to watching the data closely and building the capacity to respond quickly when the picture becomes clearer.