AI and inequality — what the evidence says about jobs, wages and who captures the gains, 2020–2026

Abstract

1. Introduction

Every wave of automation has produced two competing stories about inequality. In the first, machines replace the workers whose tasks they can perform, wages for those workers fall, and the owners of the machines collect the difference. In the second, the machine makes ordinary workers more productive, demand for what they produce grows, and the gains spread. The industrial robot, the subject of a companion paper on this site (Will robots take manufacturing jobs?), produced a version of the first story concentrated on routine manual work in manufacturing. Generative AI, the large language models (LLMs) that write, summarise, code and advise, raises the same question about a different group of people: the graduates, clerks, analysts and professionals whose work is words and numbers rather than metal.

The question matters for three reasons. The exposed group is large: about 60 per cent of jobs in advanced economies contain tasks that AI can perform (IMF, 2024), a share that the Department for Science, Innovation and Technology (DSIT) puts at around 70 per cent for the UK (DSIT, 2026b). It is well paid: exposure rises with earnings, education and city location, which is the opposite of the 1980s and 1990s pattern in which technology hollowed out the middle (Goos, Manning and Salomons, 2014). And the money involved is unprecedented: private investment in information-processing equipment contributed around 57 per cent of US real gross domestic product (GDP) growth after the fourth quarter of 2024 (IMF, 2025), and hyperscaler capital expenditure could exceed $3.4 trillion by 2030 (IMF, 2026a). Whether that investment ends up in wages, profits or prices is the inequality question in its plainest form.

Public discussion has run ahead of the evidence. Exposure estimates, which measure what AI could do, are routinely reported as job losses, which the ILO and Microsoft's own researchers have both warned against (ILO, 2025; Microsoft Research, 2025). Field experiments showing large gains for individual workers are extrapolated to the economy, although the first population-level studies find almost nothing in pay data. And the one clear signal, a fall in entry-level hiring in exposed occupations, coincides with a post-pandemic hiring correction and higher interest rates.

This paper addresses four questions. Which occupations and income groups are exposed to generative AI, and how does that exposure differ from earlier automation? What does the empirical evidence from 2023 to 2026 show about employment, wages, task composition and the dispersion of productivity, including the finding that AI narrows performance gaps within occupations? How are the gains being distributed between labour, capital and firms? And what does the evidence imply for UK income distribution, regions and policy, and for UK SMEs in particular? Its contribution is a reconciliation of sources that measure different things, a decomposition that shows why within-occupation compression and between-group inequality can rise together, and a UK reading of the international literature.

Section 2 reviews the literature by theme. Section 3 sets out the method and three equations used to organise the evidence. Section 4 presents findings with four figures and three tables. Section 5 discusses the implications, presents the strongest counter-argument, and includes a sub-section for SMEs. Section 6 states the limitations and Section 7 concludes.

2. Literature review

2.1 Theory: tasks, displacement and reinstatement

The modern framework treats production as a set of tasks that can be performed by labour or by capital. Acemoglu and Restrepo (2019) distinguish a displacement effect, in which automation lets capital take over tasks previously done by labour, which "always reduces the labor share in value added", from a reinstatement effect, in which new tasks are created for which labour has a comparative advantage, which "always raises the labor share and labor demand". Their reading of the previous three decades in the US is that employment growth slowed because of "an acceleration in the displacement effect, especially in manufacturing, a weaker reinstatement effect, and slower growth of productivity". Karabarbounis and Neiman (2013) documented the global side: the labour share has "significantly declined since the early 1980s" in most countries and industries, and the falling price of investment goods, attributed to information technology, explains roughly half of it.

The earlier wave of computerisation was routine-biased. Goos, Manning and Salomons (2014) documented "the pervasiveness of job polarization in 16 Western European countries over the period 1993-2010", with employment growing at the top and bottom of the wage distribution and shrinking in the routine middle. The UK's record fits: Giupponi and Machin (2022) find that "earnings inequality is considerably higher in the UK than it was 40 years ago", with sharp rises in the 1980s, upper-tail growth in the 1990s and 2000s, and a modest narrowing after 2010 helped by the 1999 minimum wage.

Two recent theoretical contributions frame what generative AI might do differently. Acemoglu (2024) applies the task model to current exposure estimates and finds gains that are "nontrivial but modest": no more than a 0.66 per cent increase in total factor productivity (TFP) over ten years, later revised to less than 0.53 per cent, with AI "predicted to widen the gap between capital and labor income". Autor (2024) argues the opposite possibility: that AI's distinctive capability "to weave information and rules with acquired experience to support decision-making" could let workers without elite credentials perform higher-stakes tasks, and that "AI, if used well, can assist with restoring the middle-skill, middle-class heart of the US labor market". He is explicit that this is "an argument about what is possible", not a forecast. The Governor of the Bank of England set out the same channels in May 2026, with the warning that in the First Industrial Revolution "real wages stagnated until the 1840s" despite productivity gains (Bailey, 2026).

2.2 Measuring exposure

Exposure indices map what AI systems can do onto the tasks and abilities that occupations require. Felten, Raj and Seamans (2021) built the AI Occupational Exposure (AIOE) index from ten AI applications and the abilities occupations use, which the Department for Education (DfE) later applied to the UK (DfE, 2023). Eloundou et al. (2024) asked whether an LLM could materially cut the time a task takes and found that "around 80% of the U.S. workforce could have at least 10% of their work tasks affected" and "approximately 19% of workers may see at least 50% of their tasks impacted", with "higher-income jobs potentially facing greater exposure". With software built on top of LLMs, the share of tasks that could be done faster rises from about 15 per cent to "between 47 and 56% of all tasks".

The IMF's Staff Discussion Note added a complementarity dimension: whether exposed tasks are ones where AI is likely to assist a human whose judgement is still needed, or ones where it can substitute. Its headline is that "almost 40 percent of global employment is exposed to AI", about 60 per cent in advanced economies, 40 per cent in emerging markets and 26 per cent in low-income countries. In advanced economies "27 percent of employment is in high-exposure, high-complementarity occupations, 33 percent in high-exposure, low-complementarity jobs" (IMF, 2024). The ILO's refined index takes a narrower view: "one in four workers across the world are in an occupation with some degree of GenAI exposure", 34 per cent in high-income countries and 11 per cent in low-income ones, with only 3.3 per cent of global employment in the highest exposure gradient (Gmyrek et al., 2025). Its 2023 paper had already concluded that "the overwhelming effect of the technology will be to augment occupations, rather than to automate them" (Gmyrek, Berg and Bescond, 2023), and the ILO stresses that "the figures reflect potential exposure, not actual job losses" (ILO, 2025). The Bank for International Settlements (BIS) modelled the same question by wage quartile and found that at high capability 36 per cent of skills are exposed on average "and up to 45% in the highest wage quartile", but that lower-wage workers face substitution rather than complementarity because "AI can typically perform their core skills" (Auer, Köpfer and Švéda, 2024).

For the UK, Henseke et al. (2025) built the Generative AI Susceptibility Index (GAISI), defined as "the share of job activities where an LLM or LLM-powered system can reduce task completion time by at least 25%". They find that 94 per cent of UK jobs had some exposure in 2023/24 but only 13 per cent were highly exposed, and that high-skill occupations scored 0.474 on average against 0.236 for low-skill ones. The Greater London Authority (GLA) applied the ILO framework to London and found 46 per cent of the capital's workforce exposed against 38 per cent nationally, with 77 per cent of jobs in information and communication and in finance and insurance exposed, and women holding 60 per cent of the highest-exposure roles (Dwan-O'Reilly, 2026).

2.3 Field experiments and the compression finding

Three studies published in 2023 established the pattern that runs through this paper. Brynjolfsson, Li and Raymond (2025), studying 5,179 customer support agents given a conversational assistant, found productivity up "14% on average, including a 34% improvement for novice and low-skilled workers but with minimal impact on experienced and highly skilled workers", and interpret the tool as one that "disseminates the best practices of more able workers". Noy and Zhang (2023) assigned writing tasks to 453 college-educated professionals: "the average time taken decreased by 40% and output quality rose by 18%. Inequality between workers decreased". Dell'Acqua et al. (2023), with 758 consultants, found tasks completed 25.1 per cent more quickly and at "more than 40% higher quality", with below-average performers improving by 43 per cent and above-average ones by 17 per cent; but on a task chosen to lie outside the model's competence, consultants using AI were "19 percentage points less likely to produce correct solutions". Peng et al. (2023) found developers with an AI pair programmer "completed the task 55.8% faster than the control group".

Observational data agree. Georgieff (2024), for the Organisation for Economic Co-operation and Development (OECD), examined 19 countries from 2014 to 2018 and found "no indication that AI has affected wage inequality between occupations so far", but some evidence of reduced inequality within occupations, because "low performers have more to gain from using AI because AI systems are trained to embody the more accurate practices", or because low performers left roles they could not adapt to.

The compression result is not universal. Otis et al. (2023) gave Kenyan entrepreneurs a GPT-4 business adviser and found no average effect, but "the treatment effect for entrepreneurs who were high performing at baseline to be 0.27 standard deviations greater than for low performers": high performers gained about 15 per cent and low performers lost about 8 per cent. Where the task is open-ended and the user must judge which advice to act on, AI can widen gaps. Hui, Reshef and Zhou (2024) found that freelancers in highly affected occupations suffered "reductions in both employment and earnings", with "suggestive evidence that top freelancers are disproportionately affected by AI".

2.4 Population-level evidence, 2023 to 2026

The first study to link adoption surveys to administrative earnings is Humlum and Vestergaard (2025), covering 25,000 Danish workers in 7,000 workplaces across 11 exposed occupations. By 2024, 49.1 per cent had used chatbots; where employers encouraged use, adoption rose "from 47% to 83%". Users reported time savings of "2.8% of work hours", only "3–7% of workers' productivity gains" reached earnings, and the confidence intervals ruled out effects on earnings and hours larger than 1 per cent. The March 2026 revision, two years after ChatGPT, reports "precise null effects on earnings and recorded hours at both the worker and workplace levels", with employers absorbing AI through "new tasks in content generation, AI oversight, and AI integration" (Humlum and Vestergaard, 2026).

US payroll data tell a different story about the young. Brynjolfsson, Chandar and Chen (2025), using records from the payroll processor ADP for millions of workers, documented "substantial declines in employment for early-career workers (ages 22-25) in occupations most exposed to AI" while employment for experienced workers in the same occupations rose, with adjustment "primarily via employment rather than compensation". Between late 2022 and September 2025, 22 to 25-year-olds in exposed jobs saw a 6 per cent fall against rises of 6 to 9 per cent for older workers; controlling for firm-level shocks the relative decline was 16 per cent. The August 2026 update, with data to June 2026, finds "no evidence of widespread, economy-wide job displacement" but puts young workers in exposed roles "19% below where it would be had it kept pace" with less exposed peers, a gap working "primarily through reduced hiring of young workers rather than increased separations" (Brynjolfsson, Chandar and Chen, 2026). Davis (2026) finds the wage counterpart: employment in AI-exposed sectors fell 1 per cent since late 2022 while the economy added 2.5 per cent, but weekly wages in computer systems design rose 16.7 per cent against 7.5 per cent nationally, and exposure lowered wage growth where there was no experience premium and raised it where the premium was large.

Against this, the Yale Budget Lab finds the occupational mix has changed only about 1 percentage point more since November 2022 than during the internet's adoption and concludes there has been "no discernible disruption" 33 months after ChatGPT (Gimbel et al., 2025). Earlier AI pointed the same way: Acemoglu et al. (2022) found establishments adopting AI from 2010 to 2018 "reduce hiring in non-AI positions", but aggregate effects "currently too small to be detectable".

2.5 Adoption and usage

Adoption data condition everything above. Bick, Blandin and Deming (2024) found "nearly 40 percent of the U.S. population age 18-64 uses generative AI", 23 per cent of workers had used it at work in the previous week, and "between 1 and 5 percent of all work hours are currently assisted by generative AI", with reported time savings of 1.4 per cent of hours. Chatterji et al. (2025) report ChatGPT "adopted by around 10% of the world's adult population" by July 2025, with non-work use rising to more than 70 per cent of messages and work use "more common for educated users in highly-paid professional occupations". The Anthropic Economic Index shows the same skew: computer and mathematical occupations account for about 35 per cent of conversations, per-capita usage rises with income at an elasticity of 0.7, and the UK's usage index of 2.67 sits below the US (3.62) and well below Israel (7.0) (Anthropic, 2025; 2026). In the UK, 35 per cent of businesses with ten or more employees used AI by June 2026, up from about 12 per cent in September 2023, but only 28 per cent of micro-businesses against 49 per cent of large ones (ONS, 2026b).

2.6 Gaps

Three gaps remain. Almost all experimental evidence comes from single tasks over weeks, not careers over years. The population studies cover Denmark and the US; the UK has exposure indices and vacancy data but no linked earnings study. And no source reconciles the compression inside occupations with the factor-share effects in the aggregate. The method below is designed around that third gap.

3. Method

3.1 Design and data

This is desk research: a secondary analysis of published sources retrieved on 25 August 2026 and listed in the references. Numerical claims are taken from primary or official sources (IMF, ILO, BIS, ONS, gov.uk, Bank of England, Federal Reserve, university working papers and peer-reviewed journals), with corporate research (Anthropic, Microsoft, PwC) used only for usage patterns and labelled as such. Where a working paper has been revised, both versions are cited. The Organisation for Economic Co-operation and Development's Employment Outlook 2023 could not be retrieved in full and is represented by the OECD working paper of Georgieff (2024).

3.2 Exposure as a task-weighted index

Every exposure index in Section 2.2 reduces to the same form:

E with subscript o is the exposure score of occupation o; the sum runs over the K tasks (subscript o) that make up the occupation; w with subscripts o,k is the share of working time spent on task k; and a with subscript k is an indicator that equals one if an AI system can perform the task or cut its time by a stated threshold (25 per cent in Henseke et al., 2025) and zero otherwise. Two properties follow. Exposure is a statement about tasks, not jobs, so it can be high in an occupation that AI cannot perform end to end. And the threshold assumed for a drives the level: the 80 per cent of US workers with at least a tenth of tasks affected (Eloundou et al., 2024) and the 25 per cent of global employment with some exposure (Gmyrek et al., 2025) are not contradictory, they use different thresholds.

3.3 Within- and between-worker components of productivity

To reconcile the compression found in experiments with the inequality found in aggregates, the paper uses a standard shift-share decomposition of average productivity across a group of workers:

P is average output per worker in an occupation or firm; the sum runs over n worker types i (for example novice and experienced); s with subscript i is the employment share of type i; p with subscript i is the productivity of that type; and Δ denotes the change over the period. The first term is the within-worker effect, the productivity gain each type receives holding the workforce mix fixed; the second is the between-worker or composition effect, the change that comes from reallocating employment between types. The experiments measure the first term. The payroll studies measure the second. The novelty of generative AI is that the first term is largest for the lowest-productivity workers (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2023), which compresses within-occupation dispersion, while employers appear to respond by reducing the share of exactly those workers (Brynjolfsson, Chandar and Chen, 2026), which raises average productivity through the second term without raising anyone's wage.

3.4 The labour-share identity

Who captures the gains is a question about the labour share, which can be written as a ratio of the real wage to labour productivity:

LS is the labour share of value added; w is the nominal wage; L is employment; p is the price of output; Y is real output; w/p is the real product wage; Y/L is labour productivity; and Δ ln denotes a proportional change. The labour share rises only if real wages grow faster than productivity. The Danish pass-through estimate gives a direct reading: if AI raises a user's productivity by 2.8 per cent and 3 to 7 per cent of that reaches pay (Humlum and Vestergaard, 2025), the wage term rises by 0.08 to 0.2 per cent while the productivity term rises by 2.8 per cent, and the labour share of the affected activity falls by roughly 2.6 to 2.7 per cent unless output prices fall by the same proportion. This is the arithmetic behind the IMF's finding that "capital income and wealth inequality always increase with AI adoption" (IMF, 2024) and it is the paper's central analytical claim. The assumption is that the productivity gain is retained by the firm rather than competed away in prices; Section 5 returns to that assumption.

4. Findings

4.1 Exposure: who, where and how it differs from earlier automation

Figure 1 sets the UK and IMF exposure estimates side by side using the common complementarity split. In the UK, 35 per cent of workers are in high-exposure, high-complementarity occupations, 32 per cent in high-exposure, low-complementarity ones and 33 per cent in low-exposure jobs, so "around 70% of UK workers are in occupations containing tasks that AI could potentially perform or enhance", above the US figure of about 60 per cent (DSIT, 2026b). The difference reflects the UK's large professional and administrative services sector and small manufacturing one.

Table 1 summarises how exposure is distributed within the UK. Three features distinguish it from the automation of the 1980s to 2000s. First, exposure rises with qualification and pay: "employees with more advanced qualifications are typically in jobs more exposed to AI" (DfE, 2023), high-skill occupations score twice as high on GAISI as low-skill ones (Henseke et al., 2025), and the BIS finds exposure of 45 per cent in the top wage quartile against 26 per cent in the bottom (Auer, Köpfer and Švéda, 2024). Second, it is urban and southern: "workers in London and the South East have the highest exposure to AI" and those in the North East the least, although the regional variation is "much smaller than the variation observed across occupations or industries" (DfE, 2023); in London, 46 per cent of the workforce is exposed against 38 per cent nationally (Dwan-O'Reilly, 2026). Third, it is gendered through clerical work: in high-income countries "jobs at the highest risk of automation make up 9.6 per cent of female employment" against 3.5 per cent for men (ILO, 2025), and in London women hold 60 per cent of the highest-exposure roles because they are concentrated in administrative and customer-service occupations, of which "61% in Level 4 and a further 27% in Level 3" are exposed (Dwan-O'Reilly, 2026).

The BIS row is the caveat to the "AI hits the well paid" story. High earners have more exposed tasks, but their core tasks, the judgement, negotiation and accountability that define the role, remain outside the frontier, so exposure operates as assistance. Lower-paid clerical workers have fewer exposed tasks, but those tasks are the job. Exposure and vulnerability are not the same ranking, which is why the complementarity split matters more than the raw exposure share.

4.2 Productivity: large in the task, small in the payslip

Figure 2 plots the two field studies that report gains by prior performance, alongside the Kenyan counter-example. The pattern is a large gain for the less able or less experienced and a small one for the most able. Table 2 places these next to the population studies.

Two things about Table 2 deserve emphasis. The productivity column and the wage column do not connect: users save time, the firm captures the output, and pay is unchanged; the Danish estimate that 3 to 7 per cent of gains reach earnings is the only direct measurement of pass-through in the literature (Humlum and Vestergaard, 2025). And the employment column is bimodal. For workers already in post nothing has happened; for people trying to enter exposed occupations, a great deal has.

4.3 Employment: the ladder is being pulled up, not the floor

Figure 3 collects the hiring indicators. In the UK, online job adverts have fallen 15 per cent in high-exposure occupations against 10 and 6 per cent in mid- and low-exposure ones, with customer service (23 per cent) and administrative occupations (22 per cent) the worst affected, and the rank correlation between exposure and the size of the fall is −0.70 across 26 occupational groups (Schlicht, 2026). DSIT's measure, using a different window and index, gives a 38 per cent fall in high-exposure postings against 21 per cent in low-exposure ones from 2022 to 2025, concentrated "in high-salary occupations, with no significant change observed in low-salary roles", and a 44 per cent annual fall in 16 to 24-year-olds employed in programming (DSIT, 2026b). Egan (2026), using a third method, finds postings in the most exposed roles down about 40 per cent since mid-2022 against about 20 per cent for the least exposed. The US early-career gap of 19 per cent completes the picture (Brynjolfsson, Chandar and Chen, 2026).

Three qualifications apply. UK vacancies "nearly halved since their 2022 peak" for reasons that include a post-pandemic correction and higher interest rates, and Schlicht (2026) calls confident attribution to AI "premature". Unemployment rose only from 4.1 per cent in mid-2024 to 5.0 per cent in late 2025 and stood at 4.9 per cent in the second quarter of 2026, with 707,000 vacancies and payrolled employment down 78,000 in a year (Cominetti and Slaughter, 2025; ONS, 2026a): a soft labour market, not a rupture. And the firm surveys are calm: the Bank of England's Decision Maker Panel finds "AI has had little effect on employment so far, with only a minor reduction expected in coming years", only 4 per cent of AI-using firms report having cut staff because of it (Egan, 2026), and the ONS finds about half of adopters reporting no headcount change and about 7 per cent a decrease (ONS, 2026b). The consistent mechanism is fewer entry-level hires, not redundancies: the between-worker term of Equation 2 operating on the youngest cohort.

4.4 Wages and the distribution of earnings in the UK

UK pay data through April 2025 show no AI signature. Median full-time weekly earnings were £766.60, up 5.3 per cent in cash terms and 1.1 per cent in real terms (ONS, 2025a). Hourly pay for full-time jobs grew 6.9 per cent at the 90th percentile and 6.1 per cent at the 10th, the low-paid share fell to 2.5 per cent, "the lowest since the series began in 1997", and the high-paid share rose to 23.2 per cent, with 40.8 per cent of London jobs high-paid (ONS, 2025b). The compression at the bottom is the minimum wage, not AI; the growth at the top predates ChatGPT. By mid-2026 regular pay growth had slowed to 3.5 per cent, or 0.5 per cent in real terms (ONS, 2026a), and the Office for Budget Responsibility expected real wages to grow "2.0 per cent in total between now and 2031" (Cominetti and Slaughter, 2025).

Two UK findings hint at what may be coming. Henseke et al. (2025) find that the pay premium attached to AI-exposed tasks fell by about 11.6 per cent between 2017 and 2023/24, and that vacancies were 5.5 per cent lower in the second quarter of 2025 than hiring patterns before ChatGPT implied. And the IMF finds that postings requiring new skills carry wage offers 3 to 3.4 per cent higher in the US and the UK, that 85 per cent of workers listing new skills hold a degree against 60 per cent of others, and that the benefits go to "both high- and—through higher consumption of services—low-skilled workers", with "no significant benefits for middle-skilled workers" (Jaumotte et al., 2026): the polarisation of Goos, Manning and Salomons (2014) returning by a different route.

4.5 Who captures the gains: labour, capital and firms

Figure 4 shows the three channels and where the evidence puts the money.

Labour. The UK is an outlier in labour-share history: its share rose from about 53.8 per cent in 1997 to about 59.5 per cent in 2023, "around 8% higher in 2023 than in 1995", while the US share fell 10 per cent and Japan's 5 per cent. Over half of the rise came from a shift towards services "where the labour share tends to be higher, particularly the professional and administrative support, and the government, health and education industries" (ONS, 2024). Those are the industries most exposed to generative AI, so the composition effect that has protected the UK labour share is the one now at risk. By Equation 3, productivity gains of a few per cent in exposed activities with negligible wage pass-through lower the labour share of those activities by almost the full amount.

Capital. The IMF's model finds that "capital income and wealth inequality always increase with AI adoption", whatever happens to labour income; in its most favourable scenario, high complementarity with strong productivity gains, total income rises for everyone but "ranging from 2 percent for low-income workers to almost 14 percent for high-income workers" (IMF, 2024). The financial data show the concentration: information technology carries 35 per cent of the S&P 500 and seven companies alone 33 per cent (IMF, 2025); AI-related technology investment "added an estimated 0.5 percentage point to US GDP growth in 2025" (IMF, 2026b); and firms along the AI value chain "have increasingly relied on circular financing arrangements" (IMF, 2026a). Market power concentrates among "a small number of dominant firms and 'hyper-scalers'" (Tamirisa et al., 2026). A companion paper examines the financial-stability side (AI, debt and financial markets).

Firms and consumers. Between labour and the hyperscalers sit the adopting firms, where the evidence is most favourable to a broad distribution of gains. Inference prices for some frontier models have fallen "by over 99 percent" (Tamirisa et al., 2026), a transfer from model owners to users. PwC reports productivity growth "40% higher at companies most exposed to AI versus least" (PwC, 2026), and DSIT finds 56 per cent of UK adopters reporting gains, mostly up to 20 per cent (DSIT, 2026b). Whether those gains reach consumers as prices, workers as pay or owners as margin depends on competition in each product market, which no source yet measures.

5. Discussion

5.1 Answering the four questions

Exposure. Generative AI exposes the top of the distribution more than the bottom, the South East more than the North, women more than men through clerical work, and around 70 per cent of UK workers in some degree (DSIT, 2026b; DfE, 2023; Dwan-O'Reilly, 2026). The difference from earlier automation is not only which occupations are exposed but how: for high earners the exposed tasks are peripheral and the effect is assistance; for clerical workers they are central and the effect is substitution (Auer, Köpfer and Švéda, 2024). Exposure rankings overstate the risk to professionals and understate it to administrators.

Evidence. Inside the task, AI's gains are real, large and concentrated on the least experienced (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2023; Noy and Zhang, 2023). Inside the payslip, nothing has changed: the Danish nulls are precise (Humlum and Vestergaard, 2026) and the UK earnings distribution through 2025 is shaped by the minimum wage and the pre-existing growth of high pay, not by AI (ONS, 2025b). Inside the hiring pipeline, the youngest workers in exposed occupations are being hired less in the US and advertised for less in the UK (Brynjolfsson, Chandar and Chen, 2026; Schlicht, 2026; DSIT, 2026b). Equation 2 reconciles these: the within-worker term compresses productivity gaps, the between-worker term changes who is employed, and neither has yet moved the wage.

Distribution. The gains have gone first to the owners of the models and the infrastructure, then to adopting firms as retained productivity, and last to workers, where the only measured pass-through is 3 to 7 per cent (Humlum and Vestergaard, 2025). Capital and wealth inequality rise in every IMF scenario (IMF, 2024). The UK's unusual rising labour share rests on the services composition that AI now targets (ONS, 2024).

The UK. Regional inequality is unlikely to widen much through exposure alone, because geographic variation is small relative to occupational variation (DfE, 2023), but London, where 40.8 per cent of jobs are high-paid (ONS, 2025b), is where both the losses from displacement and the gains from augmentation will concentrate. The clearest distributional risk is generational: a cohort entering exposed occupations after 2022 faces fewer openings and, on the US evidence, no compensating wage premium.

5.2 Comparison with the literature

The paper's reading sits between Acemoglu (2024) and Autor (2024). Acemoglu's "nontrivial but modest" aggregate is consistent with the Danish nulls and the Yale finding of no broad disruption, and his prediction that AI widens the capital-labour gap is what the pass-through arithmetic implies. Autor's possibility, that AI extends expertise to workers without elite credentials, is what the field experiments show inside the task; the difficulty is that employers appear to be responding by hiring fewer such workers rather than paying them more. The IMF's new-tasks evidence (Jaumotte et al., 2026) and the Danish finding that adopters moved into higher-paying occupations (Humlum and Vestergaard, 2026) are the reinstatement channel of Acemoglu and Restrepo (2019) beginning to operate, but at 8.4 per cent of workers receiving new tasks it does not yet offset the displacement seen in hiring.

5.3 The strongest counter-argument

The strongest case against this paper's conclusion is that it mistakes a cyclical hiring correction for a structural change and a transitional pattern for a permanent one. On the first point, Schlicht (2026) is candid that remote working, post-pandemic over-hiring and higher rates overlap with AI, and the UK vacancy fall of 13 per cent in the year to late 2025 is modest against the 37 per cent fall in the first year of the financial crisis (Cominetti and Slaughter, 2025); the Yale Budget Lab finds the occupational mix moving no faster than in the internet era and notes that shifts "were well on their way during 2021, before the release" of ChatGPT (Gimbel et al., 2025). On the second, every previous general-purpose technology produced a transition in which the reinstatement effect lagged the displacement effect: 60 per cent of 2018 employment is in occupations that did not exist in 1940 (Jaumotte et al., 2026), and the skills demanded in exposed roles are already "changing more than twice as fast" as elsewhere (PwC, 2026). On this view the compression inside occupations is the durable finding, the entry-level squeeze is temporary, the pass-through will rise as labour markets tighten and workers with AI skills become scarce, and inequality could fall as Autor (2024) hopes. If that view is right, the policy priority is to shorten the transition through skills and hiring incentives rather than to redistribute.

The paper's response is threefold. The entry-level decline is specific to exposed occupations and concentrated where AI substitutes rather than assists, which a general cycle would not produce (Brynjolfsson, Chandar and Chen, 2025). The capital-share effect does not depend on the transition: it follows from Equation 3 whenever productivity outruns pay, and the IMF finds it in every scenario (IMF, 2024). And the optimistic case rests on competition to pass gains through, in a market where seven firms hold a third of the US index (IMF, 2025). The counter-argument is strongest on timing and weakest on distribution.

5.4 Implications for UK policy

Five implications follow. First, measure pass-through. The UK has exposure indices and vacancy data but no study linking AI adoption to administrative earnings as Denmark has; the ONS and HM Revenue and Customs hold the data to do it. Second, treat the entry-level squeeze as the live problem. Apprenticeships rose from 3 per cent of AI hires in 2020 to 19 per cent in 2025 (DSIT, 2026a), the right direction; the risk is that the "seniorised" junior roles PwC (2026) describes lock out those without experience. Third, direct skills policy at the 32 per cent in low-complementarity exposed jobs (DSIT, 2026b), most of them clerical and disproportionately women (Dwan-O'Reilly, 2026), rather than at the professionals who will be assisted regardless; the Governor's view that skills investment "unambiguously will help on the employment question" (Bailey, 2026) applies most to this group. Fourth, protect the labour share through competition policy rather than wage policy: the pass-through problem is a market-power problem at the model layer (Tamirisa et al., 2026) and a margin problem at the adopter layer. Fifth, be honest about the £400 billion by 2030 figure cited in the AI Opportunities Action Plan (DSIT, 2025): the IMF's medium-term range for productivity is 0.1 to 0.8 percentage points a year (IMF, 2026b), DSIT's own range is 0.4 to 1.2 (DSIT, 2026b), and growth in potential output has been 1.3 per cent a year since the financial crisis (Bailey, 2026). Those figures are compatible with a substantial prize and with a very unequal one.

5.5 Implications for UK small and medium-sized businesses

The UK adoption gap is a size gap: 28 per cent of micro-businesses used AI in June 2026 against 49 per cent of large ones, and DSIT puts firms using or planning AI at 14 per cent of micro-businesses, 23 per cent of mid-sized and 36 per cent of large ones, with fewer than a third of staff using it even in adopting firms (ONS, 2026b; DSIT, 2026b). Five practical points follow.

The productivity gain is real and is largest for the people you find hardest to hire. The experiments show the biggest gains for novices and below-average performers (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2023). For a 30-person firm that cannot match London salaries for experienced staff, that is the most useful finding in this paper: AI assistance narrows the gap between the people you can afford and the people you cannot. The condition is supervision, because the same consultants were 19 percentage points more likely to be wrong on tasks the model could not do (Dell'Acqua et al., 2023) and Kenyan entrepreneurs who acted on poor advice lost ground (Otis et al., 2023). The practical safeguards — what staff may paste into a model, who checks the output, what is logged — are set out in a separate piece on AI use policies for small firms.

Expertise is the constraint, not cost. Lack of expertise is the most cited barrier and cost is cited by only 7 to 14 per cent of firms (ONS, 2026b); 97 per cent of organisations employing AI staff report a skills gap and 35 per cent struggle to fill roles (DSIT, 2026a). Encouragement and training matter more than tools: Danish employers that encouraged use raised adoption from 47 to 83 per cent, and training cut the gender gap in use from 11.9 to 5 percentage points (Humlum and Vestergaard, 2025). The three-fifths of UK firms citing expertise gaps that retrained staff (ONS, 2026b) chose the right response.

Do not let the entry-level squeeze become your own skills shortage. The temptation is to stop hiring juniors because AI can do their first-year tasks, and the hiring data show many firms doing exactly that (Brynjolfsson, Chandar and Chen, 2026; Schlicht, 2026). Firms that keep training juniors with AI assistance will hold the experienced staff of 2030 that others will bid for; the premium for experience is already rising in exposed sectors (Davis, 2026).

Expect the gains to stay with you only if your market lets them. Equation 3 cuts both ways. If a firm's productivity rises and its wages do not, its margin rises; if its competitors do the same, prices fall and the customer captures the gain. The 99 per cent fall in model inference prices (Tamirisa et al., 2026) is what that looks like at the supplier layer. SMEs in competitive services should plan for the second outcome and invest the transition in what AI does not supply: relationships, accountability and the physical work the exposure indices score near zero.

Get the IT foundations right first. Every adoption figure above assumes accessible data, secure systems and equipped staff. Our business IT support and cloud services pages describe that groundwork; the companion paper on data centre electricity demand explains why model costs may not stay as low as they are.

6. Limitations

The paper is a secondary analysis and inherits the limits of its sources. Exposure indices are forecasts of capability, not measurements of use; the ILO's mean automation score fell from 0.30 in 2023 to 0.29 in 2025 as its method was refined (Gmyrek et al., 2025), and Figure 1 combines an IMF split from 2024 with a DSIT split from 2026 that used related but not identical methods. The field experiments cover single tasks over weeks in a handful of occupations and cannot speak to learning, deskilling or career progression; the compression result in Figure 2 rests on two studies with performance splits and one counter-example. The Danish study is the only linked-earnings evidence and Denmark's institutions differ from the UK's. The Canaries findings moved from a 13 per cent relative decline in the original working paper to 16 per cent controlling for firm shocks and 19 per cent in the August 2026 update, which shows both that the effect is growing and that the estimate is not settled (Brynjolfsson, Chandar and Chen, 2025; 2026). The UK vacancy measures in Figure 3 use three different exposure indices and windows and are not comparable with each other or with the US employment gap. The labour-share arithmetic in Section 3.4 applies the Danish pass-through to a single activity and ignores price effects, reallocation and second-round demand. The OECD Employment Outlook 2023 could not be retrieved and its survey findings are not used. Corporate sources (Anthropic, Microsoft, PwC) describe their own users and customers and are not population samples. Finally, the paper does not model taxation and transfers, which determine whether market-income inequality becomes disposable-income inequality.

7. Conclusion

Generative AI exposes a different part of the labour market from earlier automation: the well paid, the well qualified, the urban and, through clerical work, women. Inside the task the evidence is unusually consistent: AI raises the output of the least experienced most and narrows the gap between workers doing the same job. Inside the economy the evidence is equally consistent in the other direction: earnings and hours have not moved, only a small fraction of productivity gains reaches pay, hiring of young workers in exposed occupations has fallen sharply in the US and job adverts for exposed occupations have fallen roughly twice as far as others in the UK, and capital and wealth inequality rise in every scenario modelled. The reconciliation is that compression within occupations and concentration between factors are not in conflict; they are the within and between terms of the same decomposition. For the UK, whose labour share has risen for a generation on the strength of the service industries that AI now targets, the distributional risk lies less in regional divergence than in the gap between those who own or retain the gains and those whose entry to the exposed occupations is being deferred. For small firms the findings are practical: the technology helps the staff you can afford most, expertise rather than cost is the barrier, and the firms that keep training juniors through the transition will own the scarce experience at the end of it.

References

• Acemoglu, D. (2024) The Simple Macroeconomics of AI. NBER Working Paper 32487. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w32487 (accessed 25 August 2026).
• Acemoglu, D. and Restrepo, P. (2019) 'Automation and New Tasks: How Technology Displaces and Reinstates Labor', Journal of Economic Perspectives, 33(2), pp. 3–30. Available at: https://www.aeaweb.org/articles?id=10.1257/jep.33.2.3 (accessed 25 August 2026).
• Acemoglu, D., Autor, D., Hazell, J. and Restrepo, P. (2022) 'Artificial Intelligence and Jobs: Evidence from Online Vacancies', Journal of Labor Economics, 40(S1), pp. S293–S340. Record available at: https://researchonline.lse.ac.uk/id/eprint/113325/ (accessed 25 August 2026).
• Anthropic (2025) Anthropic Economic Index report: Uneven geographic and enterprise AI adoption. 15 September 2025. Available at: https://www.anthropic.com/research/anthropic-economic-index-september-2025-report (accessed 25 August 2026).
• Anthropic (2026) Anthropic Economic Index report: Learning curves. 24 March 2026. Available at: https://www.anthropic.com/research/economic-index-march-2026-report (accessed 25 August 2026).
• Auer, R., Köpfer, D. and Švéda, J. (2024) The rise of generative AI: modelling exposure, substitution and inequality effects on the US labour market. BIS Working Paper 1207. Basel: Bank for International Settlements. Available at: https://www.bis.org/publ/work1207.htm (accessed 25 August 2026).
• Autor, D. (2024) Applying AI to Rebuild Middle Class Jobs. NBER Working Paper 32140. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w32140 (accessed 25 August 2026).
• Bailey, A. (2026) Can AI make cutlery? Speech at the Cutlers' Feast, Sheffield, 21 May 2026. London: Bank of England. Available at: https://www.bankofengland.co.uk/speech/2026/may/andrew-bailey-speech-at-cutlers-feast-sheffield (accessed 25 August 2026).
• Bick, A., Blandin, A. and Deming, D. (2024) The Rapid Adoption of Generative AI. NBER Working Paper 32966. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w32966 (accessed 25 August 2026).
• Brynjolfsson, E., Chandar, B. and Chen, R. (2025) Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab, November 2025. Available at: https://digitaleconomy.stanford.edu/app/uploads/2025/11/CanariesintheCoalMine_Nov25.pdf (accessed 25 August 2026).
• Brynjolfsson, E., Chandar, B. and Chen, R. (2026) Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (revised 12 August 2026). Stanford Digital Economy Lab. Available at: https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ (accessed 25 August 2026).
• Brynjolfsson, E., Li, D. and Raymond, L.R. (2025) 'Generative AI at Work', The Quarterly Journal of Economics, 140(2), pp. 889–942. Working paper record available at: https://www.nber.org/papers/w31161 (accessed 25 August 2026).
• Chatterji, A., Cunningham, T., Deming, D., Hitzig, Z., Ong, C., Shan, C. and Wadman, K. (2025) How People Use ChatGPT. NBER Working Paper 34255. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w34255 (accessed 25 August 2026).
• Cominetti, N. and Slaughter, H. (2025) Labour Market Outlook Q4 2025. London: Resolution Foundation, 15 December 2025. Available at: https://www.resolutionfoundation.org/publications/labour-market-outlook-q4-2025/ (accessed 25 August 2026).
• Davis, S. (2026) AI is simultaneously aiding and replacing workers, wage data suggest. Dallas: Federal Reserve Bank of Dallas, 24 February 2026. Available at: https://www.dallasfed.org/research/economics/2026/0224 (accessed 25 August 2026).
• Dell'Acqua, F., McFowland, E., Mollick, E.R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K.R. (2023) Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321 (accessed 25 August 2026).
• Department for Education (DfE) (2023) The impact of AI on UK jobs and training. 28 November 2023. Available at: https://assets.publishing.service.gov.uk/media/656856b8cc1ec500138eef49/Gov.UK_Impact_of_AI_on_UK_Jobs_and_Training.pdf (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2025) AI Opportunities Action Plan. 13 January 2025. Available at: https://www.gov.uk/government/publications/ai-opportunities-action-plan/ai-opportunities-action-plan (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2026a) AI Labour Market Survey 2025 report: executive summary. 28 January 2026. Available at: https://www.gov.uk/government/publications/ai-labour-market-survey-2025-report/ai-labour-market-survey-2025-report-executive-summary (accessed 25 August 2026).
• Department for Science, Innovation and Technology (DSIT) (2026b) Assessment of AI capabilities and the impact on the UK labour market. 28 January 2026. Available at: https://www.gov.uk/government/publications/assessment-of-ai-capabilities-and-the-impact-on-the-uk-labour-market/assessment-of-ai-capabilities-and-the-impact-on-the-uk-labour-market (accessed 25 August 2026).
• Dwan-O'Reilly, J. (2026) London's workforce exposure to generative artificial intelligence. GLA Economics Working Paper 103. London: Greater London Authority, April 2026. Available at: https://www.london.gov.uk/sites/default/files/2026-04/London%E2%80%99s%20workforce%20exposure%20to%20generative%20artificial%20intelligence.pdf (accessed 25 August 2026).
• Egan, E. (2026) Generative AI: degenerative for jobs? Bank Underground, 22 January 2026. London: Bank of England. Available at: https://bankunderground.co.uk/2026/01/22/generative-ai-degenerative-for-jobs/ (accessed 25 August 2026).
• Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2024) 'GPTs are GPTs: Labor market impact potential of LLMs', Science. doi:10.1126/science.adj0998. Preprint available at: https://arxiv.org/abs/2303.10130 (accessed 25 August 2026).
• Felten, E.W., Raj, M. and Seamans, R. (2021) 'Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses', Strategic Management Journal, 42(12), pp. 2195–2217. doi:10.1002/smj.3286. Record available at: https://collaborate.princeton.edu/en/publications/occupational-industry-and-geographic-exposure-to-artificial-intel/ (accessed 25 August 2026).
• Georgieff, A. (2024) Artificial intelligence and wage inequality. OECD Artificial Intelligence Papers No. 13. Paris: OECD Publishing. Record available at: https://ideas.repec.org/p/oec/comaaa/13-en.html (accessed 25 August 2026).
• Gimbel, M., Kinder, M., Kendall, J. and Lee, M. (2025) Evaluating the Impact of AI on the Labor Market: Current State of Affairs. New Haven, CT: The Budget Lab at Yale, 1 October 2025. Available at: https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs (accessed 25 August 2026).
• Giupponi, G. and Machin, S. (2022) Labour market inequality. IFS Deaton Review of Inequalities. London: Institute for Fiscal Studies, 15 March 2022. Available at: https://ifs.org.uk/publications/labour-market-inequality (accessed 25 August 2026).
• Gmyrek, P., Berg, J. and Bescond, D. (2023) Generative AI and jobs: A global analysis of potential effects on job quantity and quality. ILO Working Paper 96. Geneva: International Labour Organization, 21 August 2023. Available at: https://www.ilo.org/publications/generative-ai-and-jobs-global-analysis-potential-effects-job-quantity-and (accessed 25 August 2026).
• Gmyrek, P., Berg, J., Kamiński, K., Konopczyński, F., Ładna, A., Nafradi, B., Rosłaniec, K. and Troszyński, M. (2025) Generative AI and jobs: A 2025 update. ILO Research Brief, 20 May 2025. Geneva: International Labour Organization. Available at: https://www.ilo.org/sites/default/files/2025-05/Research%20brief_GenAI%202025%20Update.pdf (accessed 25 August 2026).
• Goos, M., Manning, A. and Salomons, A. (2014) 'Explaining Job Polarization: Routine-Biased Technological Change and Offshoring', American Economic Review, 104(8), pp. 2509–2526. Available at: https://www.aeaweb.org/articles?id=10.1257/aer.104.8.2509 (accessed 25 August 2026).
• Henseke, G., Davies, R., Felstead, A., Gallie, D., Green, F. and Zhou, Y. (2025) How Exposed Are UK Jobs to Generative AI? Developing and Applying a Novel Task-Based Index. arXiv:2507.22748v2, 12 August 2025. Available at: https://arxiv.org/html/2507.22748v2 (accessed 25 August 2026).
• Hui, X., Reshef, O. and Zhou, L. (2024) 'The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market', Organization Science, 35. Record available at: https://profiles.wustl.edu/en/publications/the-short-term-effects-of-generative-artificial-intelligence-on-e/ (accessed 25 August 2026).
• Humlum, A. and Vestergaard, E. (2025) Large Language Models, Small Labor Market Effects. BFI Working Paper 2025-56. Chicago: Becker Friedman Institute. Available at: https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-1.pdf (accessed 25 August 2026).
• Humlum, A. and Vestergaard, E. (2026) Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777, revised March 2026. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w33777 (accessed 25 August 2026).
• International Labour Organization (ILO) (2025) One in four jobs at risk of being transformed by GenAI, new ILO–NASK global index shows. Press release, 20 May 2025. Available at: https://www.ilo.org/resource/news/one-four-jobs-risk-being-transformed-genai-new-ilo%E2%80%93nask-global-index-shows (accessed 25 August 2026).
• International Monetary Fund (IMF) (2024) Gen-AI: Artificial Intelligence and the Future of Work. Staff Discussion Note SDN/2024/001. Washington, DC: IMF. Available at: https://www.imf.org/-/media/files/publications/sdn/2024/english/sdnea2024001.pdf (accessed 25 August 2026).
• International Monetary Fund (IMF) (2025) Global Financial Stability Report, October 2025: Shifting Ground Beneath the Calm, Chapter 1. Washington, DC: IMF. Available at: https://www.imf.org/-/media/files/publications/gfsr/2025/october/english/ch1.pdf (accessed 25 August 2026).
• International Monetary Fund (IMF) (2026a) Global Financial Stability Report, April 2026: Global Financial Markets Confront the War in the Middle East and Amplification Risks. Washington, DC: IMF. Available at: https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/text.pdf (accessed 25 August 2026).
• International Monetary Fund (IMF) (2026b) World Economic Outlook, April 2026: Global Economy in the Shadow of War, Chapter 1. Washington, DC: IMF. Available at: https://www.imf.org/-/media/files/publications/weo/2026/april/english/ch1.pdf (accessed 25 August 2026).
• Jaumotte, F., Kim, J., Koll, D., Li, E.Z., Li, L., Melina, G., Song, A. and Tavares, M.M. (2026) Bridging Skill Gaps for the Future: New Jobs Creation in the AI Age. Staff Discussion Note SDN/2026/001. Washington, DC: IMF. Available at: https://www.imf.org/-/media/files/publications/sdn/2026/english/sdnea2026001.pdf (accessed 25 August 2026).
• Karabarbounis, L. and Neiman, B. (2013) The Global Decline of the Labor Share. NBER Working Paper 19136. Cambridge, MA: National Bureau of Economic Research. Available at: https://www.nber.org/papers/w19136 (accessed 25 August 2026).
• Microsoft Research (2025) Applicability vs. job displacement: further notes on our recent research on AI and occupations. 21 August 2025. Available at: https://www.microsoft.com/en-us/research/blog/applicability-vs-job-displacement-further-notes-on-our-recent-research-on-ai-and-occupations/ (accessed 25 August 2026).
• Noy, S. and Zhang, W. (2023) 'Experimental evidence on the productivity effects of generative artificial intelligence', Science, 381(6654), pp. 187–192. doi:10.1126/science.adh2586. Record available at: https://www.mendeley.com/catalogue/35312800-4345-3022-98ab-d1bd6ff8b744/ (accessed 25 August 2026).
• Office for National Statistics (ONS) (2024) Trends in the UK labour share, 1997 to 2023. 25 November 2024. Available at: https://www.ons.gov.uk/economy/economicoutputandproductivity/output/articles/trendsintheuklabourshare1997to2023/2024-11-25 (accessed 25 August 2026).
• Office for National Statistics (ONS) (2025a) Employee earnings in the UK: 2025. 23 October 2025. Available at: https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/earningsandworkinghours/bulletins/annualsurveyofhoursandearnings/2025 (accessed 25 August 2026).
• Office for National Statistics (ONS) (2025b) Low and high pay in the UK: 2025. 23 October 2025. Available at: https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/earningsandworkinghours/bulletins/lowandhighpayuk/2025 (accessed 25 August 2026).
• Office for National Statistics (ONS) (2026a) Labour market overview, UK: August 2026. 18 August 2026. Available at: https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/employmentandemployeetypes/bulletins/uklabourmarket/august2026 (accessed 25 August 2026).
• Office for National Statistics (ONS) (2026b) Artificial intelligence in UK businesses: 2023 to 2026. 20 July 2026. Available at: https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026 (accessed 25 August 2026).
• Otis, N., Clarke, R., Delecourt, S., Holtz, D. and Koning, R. (2023) The Uneven Impact of Generative AI on Entrepreneurial Performance. Cambridge, MA: Harvard Kennedy School Center for International Development, December 2023. Available at: https://www.hks.harvard.edu/centers/cid/publications/uneven-impact-generative-ai-entrepreneurial-performance (accessed 25 August 2026).
• Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023) The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv:2302.06590. Available at: https://arxiv.org/abs/2302.06590 (accessed 25 August 2026).
• PwC (2026) Global AI Jobs Barometer 2026. Available at: https://www.pwc.com/gx/en/issues/artificial-intelligence/ai-jobs-barometer.html (accessed 25 August 2026).
• Schlicht, H. (2026) Canaries in the column? AI exposure and the UK's hiring slowdown. Bank Underground, 6 August 2026. London: Bank of England. Available at: https://bankunderground.co.uk/2026/08/06/canaries-in-the-column-ai-exposure-and-the-uks-hiring-slowdown/ (accessed 25 August 2026).
• Tamirisa, N., Barhoumi, K., de Carvalho, F., Gorbanyov, M., Kido, Y., Koll, D., Nguyen, A.D.M., Ostojic, D., Shang, B., Toms, S. and Zhao, Y. (2026) Global Economic and Financial Implications of Artificial Intelligence: Lessons from a Scenario-Planning Exercise. IMF Note 2026/002. Washington, DC: IMF. Available at: https://www.imf.org/-/media/files/publications/imf-notes/2026/english/insea2026002.pdf (accessed 25 August 2026).