Pre-registration
Generative-AI tool adoption following ChatGPT's November 2022 release will produce a measurable but modest sector-level labour-productivity divergence by 2026 — high-AI-exposure white-collar sectors (information, professional services, finance) running 1-3pp/year above low-exposure sectors (construction, accommodation, transportation), consistent with Brynjolfsson's productivity-J-curve at an early-diffusion stage.
Falsification criterion — what would disprove this
This hypothesis is considered falsified if:
BLS sector-level labour-productivity (output per hour) US 2019-2026. Define AI-exposure tier from Eloundou-Manning-Mishkin (2023) GPT occupational-exposure scores aggregated to NAICS 2-digit. Compare cumulative productivity-growth divergence high-exposure vs low-exposure sectors 2022-Q4 to 2026-Q4 against pre-period (2015-Q1 to 2019-Q4) divergence. SUPPORTED if cumulative high-minus-low excess productivity growth is in [+3pp, +12pp] over the four-year post-period and exceeds the pre-period divergence by at least 2pp. REFUTED if cumulative divergence < +1pp OR is smaller than pre-period divergence (i.e. AI tracks pre-existing trend). PARTIAL if direction correct but magnitude in [+1pp, +3pp].
formal test & threshold
test: us_sector_productivity_ai_exposure_divergence_2022_2026 threshold: SUPPORTED: cumul high-minus-low productivity growth 2022-2026 in [+3pp, +12pp] AND exceeds 2015-2019 pre-trend divergence by >=2pp. REFUTED: cumul <+1pp OR < pre-trend divergence.
Method
- Template
panel_fe_decomposition- Fixed effects
sector, quarter- Clustering
sector- Sample
- 1 countries · 2015 – 2026
- Evidence type
- associational
Difference-in-differences-style decomposition: high-AI-exposure sector productivity growth post-2022Q4 vs pre-2019Q4, against low-AI-exposure sector trajectory as control. Threshold sized for a moderate-effect Brynjolfsson J-curve early-stage prediction (1-3pp/yr); strong rejection requires either no divergence or pre-existing trend dominance.
Data
| Variable | Source | Transform |
|---|---|---|
sector_labour_productivity outcome | bls:productivity_major_sector_industrytier 1 | log |
real_output_per_hour outcome | fred:OPHNFBtier 1 fred:OPHPBStier 1 | log |
ai_exposure_tier treatment | constructed:Eloundou-Manning-Mishkin 2023 GPT-exposure aggregated to NAICS 2-digit; high (>=50%), medium (20-50%), low (<20%)tier 5 | categorical |
post_chatgpt_indicator treatment | constructed:1 from 2022-12 onwardstier 5 | indicator |
ai_adoption_share treatment | us_census:Business_Trends_and_Outlook_Survey_AI_Usetier 1 | level_pct |
ict_capital_intensity control | bls:ict_capital_servicestier 1 | log |
pre_period_productivity_trend control | derived: 2015-2019 sector productivity-growth mean | level |
covid_recovery_phase control | constructed:2020-Q2 to 2022-Q4 indicatortier 5 | indicator |
● ready · ● pending · ● reconstruct-needed
Detailed result card
Result card — ai_productivity_diffusion_2023_2026_us_sectors
Verdict: INCONCLUSIVE_DATA_PENDING — insufficient observations after listwise deletion (11)
Pre-registration
- Claim: Generative-AI tool adoption following ChatGPT's November 2022 release will produce a measurable but modest sector-level labour-productivity divergence by 2026 — high-AI-exposure white-collar sectors (information, professional services, finance) running 1-3pp/year above low-exposure sectors (construction, accommodation, transportation), consistent with Brynjolfsson's productivity-J-curve at an early-diffusion stage.
- Falsification rule: BLS sector-level labour-productivity (output per hour) US 2019-2026. Define AI-exposure tier from Eloundou-Manning-Mishkin (2023) GPT occupational-exposure scores aggregated to NAICS 2-digit. Compare cumulative productivity-growth divergence high-exposure vs low-exposure sectors 2022-Q4 to 2026-Q4 against pre-period (2015-Q1 to 2019-Q4) divergence. SUPPORTED if cumulative high-minus-low excess productivity growth is in [+3pp, +12pp] over the four-year post-period and exceeds the pre-period divergence by at least 2pp. REFUTED if cumulative divergence < +1pp OR is smaller than pre-period divergence (i.e. AI tracks pre-existing trend). PARTIAL if direction correct but magnitude in [+1pp, +3pp].
- Falsification test: us_sector_productivity_ai_exposure_divergence_2022_2026
Estimate
- Error: insufficient observations after listwise deletion (11)
Variables resolved
fred:OPHNFB; fred:OPHPBS→ real_output_per_hour (outcome, publisher=fred, n=79)constructed: 1 from 2022-12 onwards→ post_chatgpt_indicator (treatment, publisher=constructed, n=12)
Variables missing data
bls:productivity_major_sector_industry(outcome, name=sector_labour_productivity) — vintage not on diskconstructed: Eloundou-Manning-Mishkin 2023 GPT-exposure aggregated to NAICS 2-digit; high (>=50%), medium (20-50%), low (<20%)(treatment, name=ai_exposure_tier) — vintage not on diskus_census:Business_Trends_and_Outlook_Survey_AI_Use(treatment, name=ai_adoption_share) — vintage not on diskbls:ict_capital_services(controls, name=ict_capital_intensity) — vintage not on diskderived: 2015-2019 sector productivity-growth mean(controls, name=pre_period_productivity_trend) — vintage not on diskconstructed: 2020-Q2 to 2022-Q4 indicator(controls, name=covid_recovery_phase) — vintage not on disk
Generated by scripts/run_panel_fe.py at 2026-06-29T17:53:26+00:00
Strongest opposing argument
Every hypothesis ships with its charitable opposing argument. The framework earns credibility by handling objections at their strongest, not weakest.
Notes
Tests the "great divergence" prediction at a stage where it should just begin to register in BLS sector productivity. By design conservative — a 1-3pp differential is small enough that null (no divergence yet) remains a plausible reading consistent with the J-curve being still in its complementary-investments phase.