The Future of Work

Two things are true about the evidence on AI and employment. There is a real, measured signal concentrated in early-career workers in exposed occupations. And the popular version of that finding overstates it in three specific ways that the authors themselves flag.

The strongest evidence that something is happening

Stanford's Digital Economy Lab, using ADP payroll records covering 3.5 to 5 million employees monthly, found that employment of 22 to 25-year-olds in the most AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers. Between November 2022 and June 2026, the most-exposed quintile fell around 11% for that age group while the least-exposed grew around 10%.

The gap has widened, from 15% at the July 2025 data vintage to 19% by June 2026. It operates through reduced hiring rather than increased separations, and it shows up in employment rather than base pay.

The mechanism they propose is elegant: AI substitutes for codified knowledge, the kind you can learn from a textbook, and complements tacit knowledge, the kind you acquire by doing. Entry-level roles are disproportionately made of codified knowledge.

Three ways the headline overstates it

The authors are careful and most coverage is not.

• 19% is relative, not absolute. It is a divergence from a counterfactual trend, not a count of destroyed jobs. The absolute decline was around 11% over three and a half years.
• They explicitly decline to claim causation. They call these "descriptive facts" and say ongoing work is needed to determine how much is caused by generative AI rather than merely correlated with it. Their own first finding is that there is no economy-wide displacement.
• Superseded numbers are circulating. The IMF, the Bank of England and the Dallas Fed have all cited the earlier 13% figure. If you see 13%, it is an old vintage.

The evidence pointing the other way

This is where it gets genuinely difficult, because the contrary evidence is equally serious.

The Economic Policy Institute, using CPS microdata, found that 98% of the increase in young college-graduate unemployment between 2024 and 2026 was driven by growth in the labour force, not by job losses. Employment levels held steady; more graduates entered the market looking for work. And young non-college workers saw similar unemployment increases despite far lower AI exposure scores. If AI were driving it, the exposed group should have fared worse.

The New York Fed analysed job postings and found the relative decline for high-exposure occupations began before ChatGPT's release and stabilised after 2023, with junior and senior roles declining in parallel. That is the opposite of a gradual-displacement signature.

Norway's full private-sector employment register, roughly 3.1 million monthly records, found no detectable employment decline for workers aged 21 to 30 in the most-exposed occupations. The one exception was software development, at around 18% below.

The honest summary: there is a real signal in early-career exposed roles, it is contested, and nobody has cleanly separated it from post-pandemic correction and interest rates.

Why jobs recompose instead of disappearing

There is a structural reason, and it is the single most clarifying fact in this literature: only 2% of occupations have a single task accounting for more than half of work time. Jobs are bundles. Automating one task removes a slice, not a role.

Denmark provides the cleanest evidence of what actually happens. Linked survey and administrative data covering 25,000 workers across 7,000 workplaces found precise null effects on earnings and hours, ruling out anything above 2% two years after ChatGPT's launch. And 85% of chatbot users reallocated their time savings to other job tasks. The saved time became different work.

Across roughly 6,000 executives in four countries, over 90% report AI has had no effect on employment over the past three years. Looking forward three years they expect a fall of around 0.7%, with two-thirds of that coming through reduced hiring rather than layoffs.

Adoption is widespread and shallow

Generative AI now saves 2.2% of total US work hours, up from 1.6% in late 2024. Within most occupations and tasks, fewer than half of workers have adopted it at all. Fewer than 3% of tasks reach 50% adoption.

The reason is often not capability. The St. Louis Fed found medical secretaries adopting at 16.8% against 61% predicted by task exposure, because privacy and the cost of an error dominate. Where mistakes are expensive, exposure does not become adoption.

What the evidence says actually helps

Most advice in this area is generic. These four things have evidence behind them.

• Structure beats enthusiasm. In the Danish data, without employer initiatives 41% of workers used chatbots and 7% used them daily. With encouragement, enterprise tools and training, 93% used them and 28% daily. The difference is organisational, not motivational. For comparison, only 11% of UK businesses have trained more than half their workforce.
• Build tacit knowledge deliberately. If AI substitutes for codified knowledge and complements tacit knowledge, the durable skills are the ones you cannot get from a manual: judgement, client relationships, knowing why a decision was made.
• Being a builder pays more than being a user. IMF analysis found AI-developer skills carry a wage premium above 8%, while AI-user skills carry about 2%.
• Expect the gains to be largest if you are early in your career. Across support work, software development and general professional tasks, the consistent finding is that AI helps novices far more than experts. One randomised experiment found it compressed the education-based performance gap by about 75%.

The uncomfortable implication of that last point is that the same technology narrowing the gap between junior and senior output is the one making it harder to get hired as a junior. Both findings are in the data, and anyone claiming a tidy story is not reading it carefully.