by Claude Opus 5.5

How can AI-driven labour-market effects be separated from cyclical and policy effects (labour costs, interest rates, weak demand)? Which identification approaches are credible, and which claims tend to be over-attributed to AI?

No headline series can separate them. Unemployment, vacancies and graduate postings all move with AI, the business cycle and policy at once. Credible identification compares groups that face the same macroeconomic and policy shocks but differ in AI exposure, and checks that timing fits. The strongest designs compare young and older workers within the same occupations and firms. Weaker ones rely on occupation-level trends, firms’ own attributions or surveys of expectations. Applied to the UK in 2026, the evidence suggests AI is making a real but modest contribution to weaker entry-level hiring in exposed office work. Most of the aggregate weakness reflects employer NICs, minimum-wage rises, high interest rates and weak demand. Aggregate job losses and the more dramatic graduate figures are over-attributed to AI. The quiet hiring-freeze channel, seniorised roles and freelance work are under-attributed.

Why the UK is an especially hard case

Several large shocks have landed within the same few years as generative AI:

  • Dec 2021–Aug 2023. Shock: Bank Rate raised from 0.1% to a 5.25% peak.

  • Nov 2022. Shock: ChatGPT released.

  • Apr 2025. Shock: Employer NICs rate up from 13.8% to 15%; threshold cut from £9,100 to £5,000; National Living Wage up 6.7%.

  • Feb 2026 onwards. Shock: Employment Rights Act 2025 reforms phased in.

  • Apr 2026. Shock: National Living Wage up 4.1% to £12.71.

  • 2026. Shock: Energy-price shock; Bank Rate held at 3.75%, with three MPC members voting to raise it in July and September.

Several of these hit the same groups AI is thought to affect, especially young people and new entrants. A naive before-and-after comparison will attribute all of it to whichever cause the analyst has in mind.

The way out is to notice that the shocks leave different fingerprints. Take the employer NICs change. Because the threshold fell as well as the rate rising, the cost increase is proportionally far larger for low earners. On a £12,000 salary, annual employer NICs rose from about £400 to £1,050, an increase of roughly 160%. On £60,000 they rose from about £7,020 to £8,250, about 17%. (These are computed figures that ignore the Employment Allowance, which was increased to offset part of the change for small employers.) The NLW also bites hardest on low-wage sectors such as retail and hospitality. Those are mostly sectors with low AI exposure. AI exposure is concentrated in higher-paid office work: finance, law, IT and administration. Where the pain shows up therefore helps identify the cause.

Interest rates are the hardest confounder, because they hit technology, finance and property, which are among the most AI-exposed sectors. The rate-driven retrenchment in tech hiring after 2022 coincided almost exactly with ChatGPT’s release.

Approach 1: exposure-based difference-in-differences

How it works. Rank occupations by AI exposure. Compare how employment or vacancies changed in high- and low-exposure occupations before and after a start date, usually November 2022. If both groups faced the same cycle, the difference between them estimates AI’s effect.

UK evidence. Bank of England staff analysis on the Bank Underground blog (August 2026), which is staff research rather than Bank policy, found vacancies fell 15% over three years in the most AI-exposed occupations, against 10% in medium-exposure and 6% in low-exposure ones. Customer service and administrative vacancies fell by more than 20%. That gradient is consistent with an AI effect, though the authors say “confident attribution remains premature”.

The main weakness: pre-trends and correlated shocks. Difference-in-differences assumes the two groups would have moved in parallel without AI. Bloomberg Economics’ June 2026 analysis of about 400 UK job types challenges that assumption. Vacancies in AI-exposed roles were “already falling before OpenAI’s ChatGPT was released in late 2022”, dropped further after it, and “have actually increased since the summer of 2024”. Private employment in vulnerable sectors also rose. A decline that starts before the treatment is a warning sign that something else, such as post-pandemic normalisation or rising rates, was already at work. PwC’s longer view makes the same point. Since 2012, postings for the most exposed quartile of occupations stood at 0.98 times their starting level while low-exposure postings grew 2.27 times. Much of that divergence predates generative AI.

Credibility: moderate. It is useful for a gradient, but weak unless pre-trends are shown and rate-sensitive sectors are handled separately.

Approach 2: within-occupation and within-firm cohort comparisons

How it works. Compare groups that share the same employer and the same macroeconomic conditions but differ in how exposed they are. Stanford’s “Canaries in the Coal Mine” study uses US payroll data from ADP to compare 22–25-year-olds with older workers across occupations of differing exposure, with firm-by-time controls. A firm hit by higher interest rates or weak sales should cut junior hiring across all its occupations, not only the exposed ones.

Evidence. The August 2026 update finds employment of 22–25-year-olds in highly exposed jobs 19% below comparable peers, up from a 15% gap in July 2025. The authors report that the divergence persists when technology firms and computer occupations are excluded, and when they control for exposure to interest-rate rises and remote work. They are also candid that these are “descriptive patterns, not causal estimates”, and that gaps shrink when education is accounted for.

The closest UK equivalent is the DSIT/LinkedIn entry-level snapshot (June 2026). It found 30 of 38 tracked entry-level roles shrinking, including accountants (−29%), graphic designers (−28%) and software engineers (−27%), while sales and customer-facing roles grew. The sharpest declines are in occupations “where AI has become most visibly capable”, but the authors say “further research is needed before conclusions can be drawn”.

The main weakness. The UK has no public payroll microdata linking age, occupation and employer at the level ADP provides in the US, so UK versions rely on job-board and LinkedIn data with uneven coverage.

Credibility: the strongest design currently available, but still descriptive. Its value is in ruling out simple explanations. A generic cost shock would not single out young workers in exposed occupations within the same firms.

Approach 3: firm-level data and surveys

There are two very different things under this heading.

Linked adoption data. Comparing firms or workers that adopted AI with similar ones that did not, using administrative records, is potentially powerful. Humlum and Vestergaard’s Danish study links surveys of chatbot adoption to registry data on earnings and hours. It finds “precise null effects on earnings and recorded hours”, ruling out effects larger than 2% two years after ChatGPT’s launch. The UK has nothing comparable published, although ONS could in principle link its Business Insights survey to HMRC payroll data. The ONS survey alone finds about 6% of firms using AI for operations reporting a fall in headcount.

Attribution surveys. These ask employers what caused their decisions or what they expect. Examples include the Morgan Stanley survey, which found a reported net 8% job reduction among UK AI adopters (January 2026) and 6% in its second wave (May 2026). The Work Foundation found 60% of large firms that cut entry-level roles citing AI, against 25% of small firms. Deloitte’s CFO survey ranks AI second, at net 47%, behind cost control at net 62%, as a drag on graduate hiring. The Bank of England’s Decision Maker Panel finds firms expecting AI to reduce employment by about 0.4% a year.

These surveys have well-known biases:

  • Selection. Surveys of adopters, such as Morgan Stanley’s of firms using AI for a year or more, describe the most engaged firms, not the economy.

  • Salience and narrative. Respondents tend to name the cause that is topical and flattering.

  • Expectations versus outcomes. Expected effects are not measured effects.

  • Small or shifting samples. Deloitte’s survey covers 58 CFOs. Morgan Stanley’s waves are not like-for-like: only the second included banking and professional services.

  • Reverse causality. Firms cutting costs for other reasons often adopt AI as part of the same programme, so AI is present at the scene without being the cause.

Credibility: high for linked administrative data, where it exists. Attribution surveys are useful as signals of intent, not as estimates.

Approach 4: aggregate churn tests and the AI-washing check

The Yale Budget Lab measures how fast the US occupational mix is changing compared with earlier technology waves. It finds the mix “appears to be changing faster than it has in the past, although not markedly so”. It also finds that exposure measures “show no sign of being related to changes in employment or unemployment”. An Oxford Economics report, as covered by Fortune, warns of “AI-washing”: Challenger, Gray & Christmas data showed AI cited in about 55,000 US job cuts in the first eleven months of 2025, only 4.5% of the total, yet the AI explanation dominated coverage. Firms have reasons to over-attribute: AI sounds strategic, while “weak demand” sounds like failure.

Attribution can be distorted in the other direction too. Centrica said “AI isn’t driving these particular job reductions” when it cut 1,300 call-centre roles, citing a 20% fall in call volumes, and unions disputed that. Firms worried about reputation or industrial relations may understate AI’s role.

A useful fingerprint: productivity and the labour share

The OBR’s March 2026 outlook offers a neat way to tell the stories apart. It modelled two scenarios that both raise equilibrium unemployment to 5.5%:

  • Technology displacement. Technology substitutes for labour and raises productivity, but higher productivity “is not reflected in higher real earnings”. The result is “a lower labour share and a higher corporate profit share”, with GDP broadly unchanged.

  • Higher labour costs. Employment falls with no productivity gain and lower real GDP.

The same unemployment rate comes with very different productivity and income-distribution signatures. If weak hiring reflects AI, output per worker should rise in exposed sectors and the profit share should edge up. If it reflects labour costs and weak demand, neither should happen. Bank of England staff work reported in August 2026 found software and IT consulting adding 0.1 percentage point a year to productivity growth in 2023–25, ten times the pre-pandemic rate. That is early evidence pointing towards AI in one sector, not across the economy.

Applying this to UK claims

Likely over-attributed to AI:

  • The overall rise in unemployment and fall in payrolls. Unemployment is 4.9% and payrolls are down 145,000 on the year. The timing follows the April 2025 cost increases and the sectoral pattern points to low-wage, consumer-facing industries. The Bank’s February 2026 Monetary Policy Report says cost pressures from higher employer NICs and the NLW weakened employment relative to GDP. Retailers reported 74,000 roles lost and cited NICs, the NLW and the Employment Rights Act, with automation as a response rather than the cause.

  • The most dramatic graduate figures. Adzuna’s reported 45.6% fall in graduate vacancies is disputed by Jisc’s Charlie Ball. He says Adzuna captures only 5–10% of graduate vacancies and that its identification of graduate jobs is unreliable. Indeed’s comparable figure is about −7%.

  • “Britain as the AI job-loss capital.” This comes from a survey of AI adopters reporting their own attribution. Across all firms, ONS data show only about 6% of AI users reporting headcount falls.

  • Large corporate announcements. Standard Chartered’s reported plan for about 7,800 back-office roles runs to the end of the decade, and the HSBC figures are press reports of options under consideration. They signal intent, not measured UK effects.

Likely under-attributed:

  • Replacement hiring and unfilled vacancies. The Bank’s July 2026 report describes firms “slowing hiring or leaving vacancies unfilled”. This channel never shows up in redundancy data, which are flat at 3.6 per 1,000 employees.

  • Entry-level shifts within exposed professions. The DSIT/LinkedIn declines in accounting, design and software, together with Canaries-style evidence that within-firm age gaps survive controls for rates, suggest a real AI component on top of cyclical weakness.

  • Seniorisation and task change within jobs. In US data, PwC finds AI-exposed entry-level roles whose requirements have moved upmarket grew 35% since 2019, though that is not a UK finding. Headcount can stay level while the work changes underneath it.

  • Freelance and self-employed work in translation, illustration and copywriting, which payroll data do not capture at all.

What would settle it

  • Linked UK microdata: ONS business survey adoption data linked to HMRC payroll records by occupation and age.

  • Reweighted Labour Force Survey data, due towards the end of 2026, to firm up youth figures.

  • The ISE Student Recruitment Survey on 14 October, a direct employer count rather than a job-board estimate.

  • Productivity and profit-share data by sector, the OBR’s fingerprint test.

Bottom line

Treat any claim that AI explains UK job weakness as a hypothesis that has to beat the cost and cycle explanations on timing, on which sectors and occupations are affected, and within firms. On current evidence AI wins some narrow contests, in junior roles in exposed office occupations, and loses the broad one.

Sources

From AI and Jobs: UK, October 2026