Erik Brynjolfsson

Confidence 0.90 · 6 sources · last confirmed 2026-08-12

Erik Brynjolfsson is a leading academic on the economics of digital technology and AI. Stanford University and NBER affiliated. Director of the Stanford Digital Economy Lab. Member of the AI Index Steering Committee at Stanford HAI (so cross-affiliated between Stanford’s two major AI-research initiatives).

Role in the wiki

Bumped to source_count: 6 on 12 August 2026 with [[2026-08-01-brynjolfsson-mckinsey-talks-talent-biggest-ai-opportunity|his McKinsey Talks Talent appearance]] (1 Aug 2026), which supplies the wiki’s most current reading of his ADP-based employment work and several new framings. The Canaries figure has moved: employment for workers aged 22–26 in the top AI-exposure quintile fell ~13% at first measurement and is “up to 16 or 17% now,” an effect “kind of growing over time,” with wages still flat. He names the pyramid-to-diamond structural shift and its consequence (“where are those middle managers going to come from?”), calls imitation-of-humans “a terrible business strategy” against complementarity, dismisses headcount-reduction-as-ROI as “a little lazy,” reframes AI as “amplifying intention,” and predicts almost everyone will manage “a fleet of agents… like the CEO of their own little entity.” He also reports fewer startups overall in the United States as part of why productivity has lagged — independent corroboration of the declining-dynamism finding.

Brynjolfsson recurs across multiple sources and is the author of two of the most-cited empirical findings in the wiki:

1. The “Equalizing Effect” customer-support study (2025 QJE / 2023 NBER WP)

Brynjolfsson, Li & Raymond (2025) “Generative AI at Work”, The Quarterly Journal of Economics 140 (2025): 889–942. Working paper predecessor: NBER 31161 (2023). Field study with 5,172 customer-support agents, 3M+ chats, fall 2020 – early 2022, at a Fortune 500 firm using a GPT-3-based AI assistant. Key findings:

• +15% productivity in resolutions per hour (preferred specification with year-month + agent + agent-tenure FE: +15.2%; +23.9% with location-only FE).
• Equalizing effect with quality nuance: low-skill workers +30% RPH and quality up; top performers small speed gains AND small quality DECLINE.
• AI-exposed workers maintain higher efficiency during AI outages — durable learning, not just real-time scaffolding.
• Treated 2-month-tenured agents perform like untreated 6-month-tenured agents — AI accelerates the experience curve ~3×.
• Convergence in communication patterns — low-skill agents begin communicating more like high-skill agents.
• Customers more polite, less likely to escalate; reduced worker attrition driven by retention of newer workers.

Note: the wiki previously cited the working-paper version (+14.2%) via AI Index 2025 §4.4. The QJE version of the paper is now the canonical primary source — slight upward revision and added top-performer-quality-decline nuance.

2. The “Canaries in the Coal Mine” employment study (2025)

Brynjolfsson, Chandar & Chen (2025), Stanford Digital Economy Lab working paper, Aug 26, 2025. Six facts using ADP payroll data showing:

• Early-career workers (22–25) in AI-exposed occupations: ~13% relative decline since late 2022.
• Older workers and less-exposed occupations stable or growing.
• Declines concentrated in automation uses, not augmentation uses.
• Adjustments visible in employment more than wages (wage stickiness).

This is the wiki’s headline empirical evidence for AI labor displacement.

3. Updated job-postings signal (cited in Thompson 2026 (NYT The Daily))

In the 14 April 2026 episode of The Daily, Clive Thompson cites a more recent Brynjolfsson analysis of job postings showing software-developer hires down 16% “in the last year or so” — extending the Canaries employment-stock signal (2022 onset, age-22-25 cohort, ~20% software-developer decline by July 2025) into the hiring-flow layer through 2025-26. Thompson reports the number verbally without naming the paper; primary-source ingest of the underlying job-postings analysis is an open target. The cited number is structurally consistent with — and somewhat sharper than — the headline numbers from Canaries, suggesting the trend is continuing rather than plateauing as AI coding tools mature.

4. Named in the entry-level-hiring-slowdown debate (MIT Sloan CIO Symposium podcast, July 2026)

Allan Tate cites Brynjolfsson (and unnamed others) as having written on the slowdown in new-graduate hiring, without the debate being resolved as to whether AI or macroeconomic factors are the driver — a passing reference, not new Brynjolfsson content, but a fifth independent wiki appearance.

Cross-paper synthesis

The two papers together describe the task-level vs. occupation-level paradox of AI’s labor impact:

• At the task level (within a job): AI raises productivity of low-skill / early-career workers more — equalizing.
• At the occupation level (across firms): AI is reducing entry-level employment in occupations where it can substitute for labor — disequalizing.

Both can be true simultaneously. The mechanism: when AI raises individual productivity, total employment in that role depends on how elastic demand is for the role’s output. See ai-employment-effects and automation-vs-augmentation.

Notable affiliations

• Stanford University — faculty
• Stanford Digital Economy Lab — director
• Stanford HAI — affiliated; member of AI Index Steering Committee
• NBER (National Bureau of Economic Research) — research associate

Open questions

• Brynjolfsson’s books (The Second Machine Age with Andrew McAfee; Race Against the Machine; Machine, Platform, Crowd) — to be filled in as more sources reference them.
• The Productivity J-Curve (Brynjolfsson, Rock, Syverson) framework — relevant to the time-lag between AI investment and measurable productivity. Mentioned only obliquely in the wiki so far.
• His earlier work on IT productivity paradox — which his AI-productivity work builds on intellectually.