H
Howardism
Plate IIEntities中文HOWARDISM

Stanford Digital Economy Lab

Erik Brynjolfsson's research center at Stanford HAI, and the vault's source for two things that must be kept apart: the 'Canaries in the Coal Mine' payroll series — ADP administrative microdata through June 2026, the sharpest measurement in the corpus of AI's age-graded employment effects — and the 'We Must Act Now' open letter, a 200-signatory call to action with no measurement in it at all. Its instrument has three standing caveats that travel with every number: a balanced ADP panel that overrepresents large firms and AI-exposed occupations, exposure defined at the occupation level (firm identifiers are anonymized), and descriptive divergence rather than causal identification

Article metadata
Publication details
Published:September 22, 2026
Filed:Entity
Domain:Entities
Tags:EconomicsWorkforceMeasurementResearch Lab
Reading:7 min
Source:AI-synthesised
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Illustration for Stanford Digital Economy Lab

Sources#

A research center at Stanford, part of the Stanford Institute for Human-Centered AI, directed by Erik Brynjolfsson. Its output reaches this vault through two very different channels, and the whole reason for this page is that they arrive under the same institutional name.

What it produces#

1. The Canaries in the Coal Mine payroll series (Brynjolfsson, Chandar & Chen). The vault's highest-resolution measurement of AI's labor-market incidence: a monthly panel of ADP administrative payroll records, first circulated August 2025, revised August 2026 with data through June 2026. Six facts, of which the load-bearing two are that there is no economy-wide displacement and that employment of 22–25-year-olds in AI-exposed occupations sits ~19% below where it would be had it kept pace with less-exposed peers. Full treatment on The Tragedy of the Cognitive Commons (the regeneration reading), Firm AI-Spend Intensity and Headcount Growth (the counter-instrument), and Codified vs Tacit Knowledge Exposure (the mechanism).

2. The AI Economic Indicators. A public dashboard series published alongside the paper, explicitly framed as data infrastructure for tracking the six facts going forward rather than as a one-time result. Not yet ingested here; it is the natural refresh target for every Canaries number the vault carries.

3. Advocacy. The July 2026 "We Must Act Now" statement (“We Must Act Now”: Sixteen Nobel Laureates Join Leading Economists and AI Researchers in Call to Prepare for AI’s Economic Transformation) was organized from the lab and announced through its news channel — 200+ economists and AI researchers including sixteen Nobel laureates. practitioner-opinion with no measurement of any kind; the vault holds it only as a dated marker of elite opinion. See Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated for why the mismatch between that consensus and the instruments underneath it is the interesting part.

The instrument, and the caveats that travel with it#

Every Canaries figure in the wiki rests on these, and they are the lab's own disclosures rather than external criticism:

  • Payroll coverage is a subset, not a census. ADP serves firms employing over 26 million US workers, but the analysis sample is a balanced panel of firms observed monthly from January 2021 (a second panel runs from January 2018 for pre-trends), comprising 3.5–5 million employees per month, restricted to full-time workers under 70 with positive earnings. Balancing conditions on firm survival, which predicts faster employment growth; the sample overrepresents manufacturing and wholesale, underrepresents retail and accommodation/food, has few firms under 10 employees, and overrepresents AI-exposed occupations relative to the ACS and CPS — especially for women.
  • Occupation codes are imputed. Job titles are missing for roughly 30% of the sample; missing occupations are filled from the worker's own most recent (failing that, next) non-missing code. The authors report the facts are largely insensitive to this.
  • Exposure is an occupational attribute, borrowed from elsewhere. Two primary measures: the GPT-4 β ratings of Eloundou et al. (2024), and the Anthropic Economic Index's automative/augmentative query shares. Both are occupation-level indices computed by other researchers and crosswalked onto ADP's 2010 SOC codes — so every criticism on Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated of those instruments is inherited here, and the lab's contribution is the outcome side, not the exposure side.
  • Occupation-level, not firm-level, and not by choice. ADP firm identifiers are anonymized, which "precludes merging external adoption measures based on job postings or earnings calls." The lab's substitute is state-level AEI usage tiers (the decline is sharpest in leading-adoption states, about −19% for the most exposed quintile). This is the structural reason Canaries and the firm-adoption panels in Firm AI-Spend Intensity and Headcount Growth cannot be averaged: they identify off different variation because one of them cannot see firms.
  • Magnitudes are sample-specific; direction is not. The lab reports its own external-validity check rather than burying it: against the 2024 ACS on a comparable 2022→2024 window, the most-vs-least-exposed gap for young workers is −0.132 in ADP against −0.022 [−0.055, +0.011] in the ACS. Within professional, information and financial services the two sources agree closely (−0.231 ADP vs −0.213 ACS); the discrepancy concentrates in education, health and public administration, partly attributable to documented 2024 ACS weighting revisions.
  • Descriptive by declaration. "Canaries in the coal mine — rather than causal estimates." The lab states pre-ChatGPT divergent trends, the attenuation under an education control, and the ADP-vs-survey magnitude gap in its own main text, and argues against them rather than around them.

Why the two channels must be kept apart#

The same director organized a 200-signatory statement about AI's economic transformation and authored one of the seven mutually-disagreeing exposure instruments that statement's subject matter rests on. Nothing improper about that — but it means "Stanford Digital Economy Lab says" can denote an administrative-microdata result with published robustness tables or a press release with a signatory count, and the vault's evidence tiers for the two are empirical and practitioner-opinion respectively. Cite the series, not the institution.

Connections#

Sources#

§ end
Cited by 7
Related articles