MIT researchers published a study finding that AI can already technically replace 11.7% of the US workforce — roughly 19 million jobs — across finance, healthcare, and professional services. The Washington Post built an interactive tool mapping which occupations face the highest automation exposure. CNBC ran a segment on the labour market impact being "like a tsunami." And then Yale University's Budget Lab looked at the actual employment data from 2022 to 2025 and found that the occupational distribution of US workers had not shifted significantly since ChatGPT's debut. Anthropic's own labour market research reached a similar conclusion: limited empirical evidence of AI-driven employment disruption so far. The gap between theoretical exposure and measured impact is the most important data point in the AI and jobs debate, and it is not getting enough attention.
The explanation for the gap involves several distinct mechanisms. First, technical replaceability is not the same as economic substitution. Even if AI can perform a task, deploying it at scale requires workflow redesign, regulatory approval in some cases, capital investment, and change management — all of which take time and create friction that slows adoption well below what capability assessments suggest is technically possible. Second, most jobs are bundles of tasks, not single tasks. AI may automate some components of a role while leaving other components unchanged, resulting in job transformation rather than job elimination — the person does different work, not no work. Third, the productivity gains from AI in many white-collar roles are currently being captured by employers as efficiency rather than headcount reduction: the same number of workers get more done, and the surplus goes to output or profit rather than layoffs.
There is a genuine exception to the "limited displacement so far" story, and it sits in the information services sector specifically. That sector has lost an average of 5,000 jobs per month over the past twelve months, with 11,000 cut in a recent single month. This is the sector that includes software development support, content production, data entry, and related knowledge work that AI tools have most directly affected. The overall labour market is large enough that these losses are statistically small, but they are real and concentrated in exactly the occupations that economists identified as high AI exposure — which suggests the broader displacement story is beginning rather than not happening.
Anthropic's labour market research introduces a measurement framework that is more useful than simple occupational exposure scores. Rather than asking which jobs AI can theoretically do, the research asks which specific tasks within jobs AI is actually being used to do, at what frequency, and with what effect on output and employment. This task-level analysis is harder to compute but more predictive of actual economic impact, because it captures the bundle-of-tasks dynamic and the complementarity effects that occupation-level analysis misses. The early findings suggest that AI use in workplaces is predominantly complementary — augmenting workers rather than replacing them — in most sectors, with substitution effects concentrated in specific task types within specific occupational clusters.
Bloomberg's guide to getting a job in 2026 as AI transforms hiring captures the pragmatic reality well: the advice is not to avoid AI-exposed roles, but to develop skills in directing, evaluating, and collaborating with AI systems rather than competing with them. This is the "Excel analyst" analogy in practice — the spreadsheet automated some analytical tasks while creating demand for people who could use spreadsheets to do analysis that was previously impossible. AI is doing the same at a larger scale and faster pace, but the direction of travel is similar. The people who thrived in the spreadsheet transition were not the ones who were unaffected; they were the ones who adapted fastest.
The skilled trades argument that has gained traction in recent commentary is worth taking seriously on its own terms rather than as mere reassurance. Plumbing, electrical work, HVAC, welding, and construction require physical dexterity, spatial reasoning in variable environments, and real-time problem-solving that current AI and robotics cannot replicate cost-effectively. These occupations were already facing labour shortages before the AI boom; if AI accelerates white-collar displacement while leaving trades unaffected, the relative labour market value of skilled trades will increase. That is not a comfortable message for people in AI-exposed professional roles, but it is probably accurate — and it suggests the job market rebalancing, when it comes, will be stranger and more sector-specific than the broad headlines about AI and unemployment currently imply.