uk-ai-study

Who bears the AI shock? The distributional and fiscal incidence in the UK. Who loses work is at least as uncertain as how much work is lost, so this study treats incidence — exposure-proportional, junior-concentrated, uniform, and author-designed top-loaded stress tests — as an explicit scenario axis rather than assuming the shock hits the top. The UK pairs one of the most AI-exposed workforces in the advanced economies with a tax-benefit system designed to insure shocks at the bottom of the income distribution. This study traces AI's employment, wage and capital shocks through the full UK tax-benefit system with PolicyEngine UK, using the scenario architecture of Doorley, O'Connor, O'Shea & Tuda (2026), Artificial intelligence and income inequality in Ireland, ESRI/DoF Report 16 (PDF).

The study asks: if generative AI displaces some workers, raises the wages of the rest, and raises returns to capital, what happens to the Exchequer, poverty, and income inequality — and who bears it, by income decile and age?

Method — exactly how we do it

The pipeline is C-AIOE exposure → employment/wage/capital shocks (JR16 eqs 3.4/3.5) → tax-benefit microsimulation, with these UK substitutions:

JR16 (Ireland)	This study (UK)
SILC microdata	FRS 2024-25 (frs_2024_25.h5, PolicyEngine build)
SWITCH microsimulation	PolicyEngine UK (latest release)
ISCO occupations	SOC2020 major groups from raw FRS adult.tab
C-AIOE (Pizzinelli et al. 2023)	Same measure, via the populace PR #325 UK crosswalk

1. Exposure (uk_ai_study/exposure.py)

Each FRS adult carries a 1-digit SOC2020 major group in the raw UKDA adult.tab (SOC2020, coded 1000–9000). We join it onto the PolicyEngine person table with person_id = SERNUM*1000 + PERSON — the current policyengine-uk-data convention, verified to match 100% of FRS 2024-25 adults. Each major group then gets its C-AIOE score (Felten AIOE × Pizzinelli complementarity adjustment) and complementarity θ from the packaged crosswalk (uk_ai_study/data/uk_soc2020_major_group_ai_exposure.csv, derived in populace PR #325 from open-licensed sources: Felten et al. 2021, O*NET-reconstructed θ per IMF WP/23/216, ASHE 2025 Table 14 employment weights; OGL v3 / CC BY 4.0 / MIT). Persons without a SOC code (children, non-workers) receive the mean score — they never enter the employment shock, which conditions on positive earnings.

2. Shocks (uk_ai_study/shocks.py)

• Employment (eq 3.4): aggregate displaced = displacement_rate × employees (weighted); employees without an observed SOC code form a pseudo-group with mean-imputed exposure, so the displacement and wage-uplift universes coincide. Quotas are allocated across major groups ∝ employment × mean C-AIOE and realised by systematic sampling on a random permutation within each group, with prescribed first-order inclusion probabilities that are equal for all records in a group (probability tilts such as the youth multiplier enter as tilts to these inclusion probabilities); a record's grossing weight enters only through the quota accounting, never its inclusion probability (#1, findings 6 and R2-5). Displaced workers are fully out of work: employment_income = 0, hours_worked = 0, employee pension contributions, salary sacrifice and statutory pay zeroed, employment_status = UNEMPLOYED (the shared transition constructor build_shocked_simulation, finding 4).
• Wage (eq 3.5): surviving workers get % uplifts ∝ θ, normalised by the employment-weighted mean θ over baseline workers (JR16-literal — the estimand decision on #1, finding 5; per-seed conservation tested in tests/test_shocks.py).
• Capital: interest and dividend income scaled by (1.005% + 0.4pp)/1.005% ≈ 1.398 (JR16's return-to-capital shock).

Scenario presets (all overridable; the +0.4pp capital shock is on in every preset, as in all JR16 scenarios):

Preset	Displacement	Wage	Source
central	7%	+2.6%	JR16's central calibration (§3.2), converting Briggs & Kodnani (2023) task-exposure and productivity figures into displacement and wage rates
low	1%	0%	Sensitivity case; JR16 §3.2 attributes ~1% to Acemoglu (2025), but his 0.9–1.1% is a ten-year GDP figure, not an employment effect (#1, finding 11)
high	13%	+2.6%	Brynjolfsson, Chandar & Chen — 13% per early drafts (Nov 2025 version: 16%); cohort-specific relative decline treated as economy-wide absolute, upper bound
central_youth_tilted	7%	+2.6%	+ Klein Teeselink (2025) junior/total ratio 5.8/4.5 tilting draws toward ages 16–24

The youth_displacement_multiplier extends JR16 (which draws randomly within groups) toward the seniority-biased evidence in Klein Teeselink (2025) and Hosseini & Lichtinger (2026).

3. Microsimulation (uk_ai_study/runner.py)

Baseline and shocked policyengine_uk.Microsimulation runs on the same dataset; the shocked run receives the modified employment_income, savings_interest_income, dividend_income and employment_status via set_input. Reported deltas (shocked − baseline):

• Exchequer cost — change in gov_balance
• Poverty — BHC and AHC person-weighted rates
• Gini — of equivalised household disposable income (HBAI cash concept, equiv_hbai_household_net_income, matching the poverty concept — #1, finding 2)
• By baseline income decile and by age band (16-24 … 65+) — mean HBAI-household-net-income change, plus each band's share of the displaced

Reproduce

conda create -n ukai python=3.13 -y && conda activate ukai
pip install -e .
export HUGGING_FACE_TOKEN=hf_...   # needs access to policyengine/policyengine-uk-data
python analysis/download_data.py    # FRS h5 + raw UKDA zip (adult.tab) -> data/
python analysis/run_all.py          # all presets -> results/*.json
bash analysis/regenerate_all.sh     # or: the FULL result set from empty results/
python -m pytest tests/             # shock-mechanics unit tests

Microdata is licensed (UKDS EUL) and never committed; data/ is gitignored. Results in results/ are aggregates only.

Known limitations (v0.1)

• 1-digit exposure only: within-major-group exposure variation is lost; JR16 uses finer occupations. A QRF imputation from 4-digit LFS SOC (as in populace PR #325) is the planned upgrade.
• Plain FRS weights, not the calibrated enhanced FRS (its household cloning breaks the adult.tab ID join; needs the SOC merge moved upstream of cloning).
• Displaced workers are current-period unemployed; JR16's "9+ months unemployed, contributory benefits exhausted" contract is not fully expressible in PolicyEngine UK inputs.
• Self-employed are outside all shocks (as in JR16). Decile figures average 50 seeded draws; the central preset is Monte-Carlo'd over 20 draws (analysis/robustness.py); grid cells are single-draw (seed=0).

References

• Doorley, O'Connor, O'Shea & Tuda (2026), ESRI/DoF Report No. 16.
• Pizzinelli et al. (2023), IMF WP/23/216 (C-AIOE, θ).
• Felten, Raj & Seamans (2021) (AIOE).
• Briggs & Kodnani (2023), Goldman Sachs.
• Acemoglu (2025), Economic Policy 40(121).
• Brynjolfsson, Chandar & Chen (2025), "Canaries in the Coal Mine?".
• Klein Teeselink (2025), SSRN 5516798.
• Hosseini & Lichtinger (2026), SSRN 5425555.