The Gist: The way we measure AI’s impact on jobs shapes what we see, and what we see shapes what we do. A new Anthropic report reveals that theoretical capability and actual usage tell very different stories. Getting the measurement right matters for policy, for education, and for avoiding the kind of premature conclusions that have undermined previous rounds of automation research.
Numbers have authority. Once a finding makes it into a widely cited report or onto a conference slide, it tends to travel independently of its caveats. This is particularly true in a field like AI and employment, where everyone from governments to universities to individual workers is trying to figure out whether they should be worried, and how much.
Getting the measurement right,
then, is not merely a methodological concern.
It has real stakes.
Two Different Questions
A task-based approach to measuring AI exposure asks: could an AI perform this task? Could it perform this task faster, or to the same quality? Eloundou et al.’s (2023) influential work operationalised this as a binary: can an LLM alone double the speed of this task? If yes, the task is exposed.
This is a sensible way to map the frontier of AI capability. But it is asking a different question from: is AI actually being used for this task, in a professional context, in a way that could displace a human worker?
Massenkoff and McCrory (2026) formalise this distinction. Their observed exposure measure requires not just theoretical feasibility but evidence of sustained work-related use in automated (rather than merely augmentative) contexts. Tasks that people might theoretically use AI for, but haven’t yet, do not count. Tasks that are used in augmentative ways receive half the weight of fully automated ones. The overall measure reflects the fraction of a job’s tasks that meet this higher bar.
The Result Is a Very Different Map
The difference between the two maps is striking. Eloundou et al.’s measure suggests that 94% of tasks in Computer and Mathematics occupations are theoretically within LLM reach. Massenkoff and McCrory’s observed measure puts actual coverage at 33%. In Office and Administrative Support, the gap is similarly large: theoretical coverage near 90%, observed coverage well below that.
For researchers and policymakers reading this, the key implication is that previous estimates of AI exposure may have substantially overstated what is actually happening now. That does not mean the theoretical ceiling is wrong. It means the gap between potential and actuality is large, and the gap is informative.
Why does actuality fall short of potential? The paper points to several mechanisms: model limitations, legal constraints, software integration requirements, human verification steps, and simple institutional inertia. Theoretical capability assumes all these barriers are zero. Observed usage measures them directly.
There is also a profound empirical problem in this field. Major economic disruptions are hard to study even in hindsight. Research on industrial robots produced opposing conclusions about job displacement despite using largely the same data. The debate about the China trade shock has continued for well over a decade (Autor et al., 2013; Acemoglu & Restrepo, 2019). If retrospective analysis of completed technological transitions remains contested, prospective analysis of one currently under way requires considerable humility.
Why ‘Augmentation’ and ‘Automation’ Are Not the Same
The distinction between augmentative and automative AI use runs through the recent literature as one of its most important analytical cleavages. Brynjolfsson, Chandar and Chen (2025) find that employment declines for early-career workers are concentrated in occupations where AI automates tasks rather than augmenting human capabilities. Where AI primarily augments, the employment picture looks different.
This is not merely an economic observation. It is a design observation. Whether a system is built to automate a task entirely or to assist a human in completing it is, to a significant extent, a choice. It reflects assumptions made by developers, procurement decisions made by employers, and norms established by regulators and professional bodies.
From an educational perspective, this framing suggests that the relevant question is not simply which jobs will survive AI but which forms of human-AI collaboration are worth building towards. The OECD (2023) has emphasised the importance of preparing workers not just for a labour market shaped by AI but for a role as active participants in how AI-augmented work is designed. That requires a different kind of education than the one most institutions currently provide.
The Data Availability Problem
One of the things I appreciate about the Massenkoff and McCrory paper is its honesty about what it cannot yet see. The Current Population Survey is well suited to tracking unemployment across occupational categories. It is less well suited to detecting subtle changes in hiring patterns or to distinguishing between workers who are unemployed and workers who have simply stopped looking.
The suggestive evidence on young workers comes partly from this data and partly from a growing literature using more granular sources. The Brynjolfsson et al. (2025) paper, for instance, uses payroll data from ADP covering millions of workers at thousands of firms. Hampole et al. (2025) use job posting data. Johnston and Makridis (2025) use US administrative data. These are different instruments measuring overlapping but distinct aspects of the same phenomenon.
The field does not yet have a single definitive dataset. What it is developing is a family of approaches that, together, triangulate on a phenomenon that no single instrument could capture alone.
What This Means for the UK Context
The report draws on US data, as most of this literature does. But the task-based approach is not inherently country-specific. Massenkoff and McCrory note that their occupation-level exposure measures can be extended to different countries. The O*NET taxonomy they use has analogues in the UK’s Standard Occupational Classification.
UK higher education is in the midst of its own reckoning with AI and graduate employment. The Russell Group’s AI principles, published in 2023, acknowledged the need for graduates to engage critically with AI tools. The UK government’s AI Opportunities Action Plan (2025) emphasised AI skills as a national priority. What neither document has done in detail is connect the theoretical ambitions for AI in education to the growing empirical evidence about what AI is actually doing to the graduate labour market.
That connection needs to be made. And it needs to be made with the kind of careful measurement that the Anthropic paper is beginning to build.
The tools for understanding AI’s economic impact are improving. They are not there yet. But the combination of usage data, occupational databases, and employment surveys is starting to produce a picture that is genuinely more honest than the forecasts we have been working with.
That is worth paying attention to, even if the picture remains incomplete.
Thanks for reading Data in Motion.
Subscribe for free to receive new posts and support my work.