AI and the Labour Market, Signal, Noise and Structural Change

Narrative

Financial press coverage and the underlying studies it cites point to a narrow, uneven first labour-market effect rather than economy-wide displacement. Bloomberg and The Wall Street Journal reported a 13% employment fall from late 2022 for US workers aged 22 to 25 in highly exposed occupations, including software development, accounting, administrative work and customer service, based on payroll records from tens of thousands of firms. This is stronger evidence than vacancy anecdotes, but it remains an observational age-by-occupation comparison, not proof that generative AI alone caused the decline.

The counter-evidence is substantial. Bloomberg’s account of the New York Fed survey found business adoption rising while relatively few firms reported job cuts attributable to AI, and the Financial Times reported that OECD research had found no employment reduction at small businesses deploying generative AI. Official Eurostat data show adoption expanding sharply, from 13.5% of EU firms with at least ten workers in 2024 to 20.0% in 2025, but adoption is not an estimate of displacement. OECD reporting also places AI alongside ageing, weak demand, trade-policy uncertainty, immigration changes and public-sector cuts as competing explanations for hiring and employment trends.

Enterprise reporting shows work redesign running in both directions. Microsoft Copilot deployments required data preparation and employee training, Bloomberg found, while BNY Mellon and IBM said they were increasing junior recruitment partly because entrants brought AI skills. Conversely, employers including Lufthansa, Block and several technology companies have begun explicitly linking administrative reductions or hiring restraint to AI-driven productivity, although corporate explanations often combine cost cutting, investment reallocation and slower growth.

The strongest measured productivity evidence remains task and workplace specific. The customer-service field experiment published in the Quarterly Journal of Economics found gains in worker productivity and customer outcomes, especially for less experienced staff, whereas reports based on employer expectations, vendor surveys or executive commentary are forecasts or intentions. Distributional risks also extend beyond headcount: Bloomberg’s controlled résumé-ranking test found race- and gender-linked variation in GPT rankings, illustrating how AI can alter access to jobs even where it has not eliminated them. The defensible conclusion by August 2026 is that adoption and job redesign are real, early-career white-collar employment may be weakening in exposed work, but aggregate labour-market effects remain difficult to separate from the post-pandemic normalisation, high interest rates and wider macroeconomic cycle.