--- title: "AI and the UK labour market: the evidence so far" aliases: - "CBP UK Evidence Review" - "Serôdio 2026" tags: - uk - descriptive - evidence-review - displacement - augmentation - wages - hiring source: "/research/2026-04-cbp-uk-labour-market-evidence.pdf" extract: "/research/extracts/2026-04-cbp-uk-labour-market-evidence.md" authors: "Pedro Serôdio" publisher: "Centre for British Progress (Research UK Day One)" date: "2026-04-22" --- # AI and the UK labour market: the evidence so far > [!info] Source > PDF (human): [2026-04-cbp-uk-labour-market-evidence.pdf](/research/2026-04-cbp-uk-labour-market-evidence.pdf) · Raw extract (machine): [2026-04-cbp-uk-labour-market-evidence.md](/research/extracts/2026-04-cbp-uk-labour-market-evidence.md) > Centre for British Progress — 22 April 2026 — UK evidence review, 36 pages ## TL;DR A UK-focused evidence review concluding that, three years after ChatGPT, **there is no consistent displacement signal in the UK labour market** even in the occupations most exposed on paper. Annual Population Survey data covering 412 occupations shows no employment difference between most- and least-exposed roles. Wages in high-exposure occupations have grown more slowly since 2019 — but this **predates ChatGPT** and cannot be cleanly attributed to AI. A modest increase in hours worked in AI-exposed occupations is consistent with augmentation. Programmers and finance analysts continue to grow; admin and clerical roles have contracted. Adoption is concentrated: ~20% of tasks account for the vast majority of usage. ## Key findings - **No aggregate employment displacement signal yet.** UK Annual Population Survey (412 occupations) shows no difference between most- and least-exposed. - **Wage growth is slower in high-exposure occupations since 2019**, but the trend predates ChatGPT — cannot easily be attributed to AI. - **Hours worked have risen modestly in AI-exposed occupations** relative to unexposed — consistent with augmentation raising demand. - **Aggregate hides composition.** Programmers and finance analysts: total roles continue to grow. Administrative and clerical: contracted since ChatGPT deployment. - **Adoption is narrow.** Roughly a fifth of all tasks account for the vast majority of AI usage. - **UK ranks #2 globally for breadth of task coverage** (after the US in absolute terms). When restricted to tasks adopted in 10+ countries, UK and US both cover all of them; at 5+ countries, UK is second at 94%. - **UK Claude usage intensity:** UK is 9th among the top 20 countries by per-capita Claude conversations (2.7× population share). Israel/Singapore/Australia lead at 4–7×; US at 3.6×. - **Collaboration patterns (UK):** 29% task iteration, 28% directive, 24% learning, 11% feedback loop, 6% validation. Stable across five AEI releases (Feb 2025 – Mar 2026). - **AI autonomy is flat at ~3.4/5 across all adoption quartiles.** Heavy users do not give AI more independence than light users. - **Higher-skilled workers adopt more.** Highest AI adoption clusters around upper-secondary qualifications; lowest adoption around lower-secondary — consistent with augmentation, not routine-task replacement. - **Adoption pace context:** generative AI reached ~52.6% US household adoption faster than electricity, refrigerator, radio, automobile, and most digital technologies. At-work adoption ~36%. - **Firm-level adoption:** ~40% of large UK firms reported using AI tools by late 2025. - **Investment scale (cited):** OpenAI, SoftBank, Oracle, MGX announced $100B initial Stargate funding (Jan 2025), up to $500B over four years. ## Why measurement is hard - Three years is short in technology-diffusion timescales. Electricity took ~four decades to show up clearly in productivity stats. - AI's effects may resemble the internet or trade-with-China rather than COVID — gradual, confounded by trade policy and the business cycle. - Annual Population Survey + ONS PAYE data lag the technology by months and lack task-level granularity. ## Methodology in brief Triangulates three sources: 1. **UK Annual Population Survey** (employment by 412 occupations) — checks for differential employment trends by exposure tier. 2. **ONS PAYE Real Time Information** (employer payroll data) — wage growth by exposure. 3. **Anthropic Economic Index** (CC BY 4.0) — revealed task coverage, mapped to UK SOC 2020 via O*NET-to-ISCO-to-SOC2020 crosswalk. Uses derivative measure ("revealed coverage") authored by Centre for British Progress, not endorsed by Anthropic. Compares against ILO task-based exposure index and the Eloundou et al. β score. ## Implications for AdaptAI This is one of the **most directly applicable** papers — UK-focused, recent, and explicit on which occupations are diverging. ### Calculator (`/calculator`) - **Confirms the direction of `industryModifier`** — IT, Finance, Professional Services correctly weighted high but with augmentation framing (not pure displacement). - **Programmer + Finance Analyst entries in `jobTitlesWeighted`:** consider softening their `weight` for senior roles given that total roles continue to grow. Junior/entry-level should remain weighted higher (paper notes hiring slowdown for younger workers is the displacement vector). - **Admin & clerical entries:** paper supports strong weights — these roles have actually contracted in the UK since ChatGPT. - **Time-horizon deltas (+15/+30/+45):** the paper's "narrow adoption now, long-tail effects later" framing supports keeping the deltas non-trivially positive but also supports the >0 today value. ### Big Picture (`/big-picture`) - `bigPictureData.industryExposure` — supports current ranking; UK-specific data validates the model. - `bigPictureData.skillsDemand.declining` — admin/clerical decline confirmed in UK PAYE data. - `bigPictureData.skillsDemand.growing` — programmer/finance analyst total demand still growing in UK; could be elevated. - `bigPictureData.aiAdoptionSpeed` — UK is a "broad but not frontier" adopter; the paper's UK = #9 in per-capita usage at 2.7× supports current Tech & IT high-adoption ranking with UK lag behind US. - `bigPictureData.keyStats` — narrative copy should acknowledge "no aggregate displacement signal yet in UK data" rather than asserting one. ### Industry Insights - `/industries/tech` — refresh prose: programmer total roles continue to grow in UK; entry-level hiring is the divergence vector. - `/industries/finance` — financial analysts also continue to grow in aggregate. - `/industries/admin-support` + `/industries/admin-customer-support` — actual contraction observed in UK PAYE data; cite this as evidence. - `/industries/professional-services` — UK is #2 globally for breadth of AI task coverage; relevant for the sector's prose. ### How to Adapt - `/adapt/upskilling` — narrow adoption (20% of tasks) means workers still have meaningful time to skill up. Counterweight to fatalist narratives. - `/adapt/ai-tools` — collaboration patterns (29% iteration / 28% directive / 24% learning) suggest workers are getting genuine help, not just being replaced. Useful for "use AI as a tool" framing. - `/adapt/human-advantage` — AI autonomy stuck at ~3.4/5 across all adoption levels; validation and judgement remain human strongholds. - `/adapt/career-transitions` — paper's evidence supports advising patience: aggregate displacement is not yet visible and may not arrive on hyperbolic timelines. ## Related notes - [[DSIT UK Labour Market Assessment]] — UK government framing of the same evidence base, with policy implications. - [[GLA Working Paper 103]] — London-specific complement using the ILO exposure framework. - [[Anthropic Observed Exposure]] — methodology source for the AEI-derived task coverage measure used here. ## Caveats & limitations - "No signal yet" is a statement about current data, not a forecast. - UK survey data is updated less frequently than the technology is moving. - Wages-grew-slower-since-2019 trend predates ChatGPT — likely confounded by Brexit, post-COVID adjustment, monetary tightening. - Anthropic Economic Index is Claude-only — under-counts ChatGPT/Copilot/Gemini. - Programmer growth + entry-level hiring slowdown are simultaneous and the paper acknowledges these are not contradictory. - Quoted CEO claims (Altman, Huang, Dorsey, Klarna) are cited as context for investment scale, not as evidence of impact.