AI and the Labor Market

Start with what is not in the data: an unemployment wave. Overall employment in the US and European economies kept growing through 2024–2026, and employers reshaped teams more often than they cut them: in ZipRecruiter's summer 2026 survey, 92% of companies reported adopting AI in some form — and most hired more, not fewer.

Now what is in it. Stanford's study of payroll data across millions of workers found that since late 2022, employment of 22–25-year-olds in the most AI-exposed occupations fell about 6%, while older colleagues in the same roles gained 6–9%; in relative terms, junior hiring dropped 13–16%. In the UK, job postings in exposed roles fell 38%. Q1 2026 brought 78.5 thousand US tech layoffs, nearly half explicitly citing AI. And employers plan to hire the Class of 2026 just 1.6% above last year — with graduate numbers rising, that flat line is a functional contraction.

The sum: AI is not firing people en masse — it is quietly removing the bottom rung of the ladder. The rung where routine work used to pay for mentorship.

The same Stanford dataset produced the year's most practical finding. Where AI automates a task end-to-end — writing routine code, running customer chats — hiring of beginners falls. Where AI augments a human — checking, suggesting, accelerating — employment holds or grows. A role's fate is decided not by industry or job title but by the share of its tasks that can be handed over with no human in the loop.

PwC's barometer (one billion job ads, 27 countries) shows the same fork from the other side: "professionalised" roles, where AI raises the expert's value, are growing twice as fast in postings and 42% faster in wages than "democratised" ones, where AI lowers the entry bar. In practical terms: if a model lets anyone do your job, that is bad news; if it lets you do the work of a team, that is good news.

The baseline is the World Economic Forum's projection: by 2030 technology displaces about 92 million jobs and creates about 170 million, net plus 78 million. The trap in that arithmetic is that the minus and the plus land on different people: a laid-off back-office clerk does not automatically become an ML-infrastructure engineer. The gap between displacement and creation is a reskilling gap — and it, not the total headcount, will be the decade's main social problem. The pessimistic pole was set by Amodei: up to half of entry-level office positions at risk within one to five years, unemployment of 10–20%. A year on, the aggregate data does not confirm that scenario — but the Stanford numbers on the young run exactly in its direction. An MIT estimate adds scale: today's models are already technically capable of work equivalent to 11.7% of employment.

And the mandatory humility shot — the radiology lesson. In 2016 Geoffrey Hinton called it "completely obvious" that neural networks would outperform radiologists within five years and advised against training them. Nine years later radiology is short-staffed and growing: AI took part of the image reading, but not the accountability, the patient, or the adjacent tasks. The moral is not that automation is a myth — it is that both alarmists and skeptics systematically miss on timing and completeness. Professions rarely disappear whole; they get reassembled.

The replace-or-reinforce logic gets clearer on concrete roles. Lawyers: contract review and case search went to the models; accountability and negotiating position did not — an experienced lawyer with AI runs several times more deals, but the routine review that used to train associates is gone, and firms are reinventing apprenticeship. Accountants: data entry is pure automation and clerical demand is shrinking, while tax planning and gray-zone judgment are reinforcement. Developers: routine code automates — hence the Stanford junior squeeze — while architecture, decomposition and reviewing machine-written code carry a rising premium. Designers: drafts and production routine are commoditized; concept, systems and taste are not. Support: agents close the routine (up to 78% in our case files), leaving fewer but more expensive humans — plus a role that did not exist three years ago, the bot trainer who labels errors and maintains the knowledge base.

The common denominator: every profession now contains an internal fork — down toward execution, which is getting cheaper, or up toward task-setting and accountability, which are getting dearer. Choosing the rung now matters more than choosing the profession.

Sources: Stanford Digital Economy Lab (Brynjolfsson et al., ADP payroll data), PwC Global AI Jobs Barometer 2026, WEF Future of Jobs Report, ZipRecruiter AI Employer Report 2026, Stanford AI Index / Lightcast, MIT estimate (Nov 2025), public statements by D. Amodei. Interpretations and recommendations are the editors' own.