How generative AI shifts research output and its distribution among researchers. The organizing question is whether AI compresses productivity gaps or widens them, and the evidence so far says the direction depends on what the task rewards.

Jiang and Wang on economics research (2026)

Tianyu Jiang and Yuqing (Claire) Wang’s job-market paper “Equalizing or Polarizing? Generative AI and Research Productivity in Economics” studies the distributional effects of AI on research productivity with three evidence sources: NBER working-paper and publication records, text-based measures of LLM-assisted writing, and an original survey of 227 academic economists measuring AI use across research tasks. ¹

• After ChatGPT’s release, productivity gaps among NBER authors narrowed: authors without research assistants published a larger share of papers in top-tier journals, and authors from smaller teams gained working-paper output and publication speed.
• The average effect of AI use on working-paper output is weak; the heterogeneity is large.
• Triple-difference estimates: greater AI exposure narrows the output gap between researchers with and without research assistants, and widens the gap between senior and junior researchers.
• Selection runs the other way from the returns: junior researchers use AI more intensively and are more likely to use agentic tools, so the widening seniority gap reflects differences in returns to AI rather than who adopts it.
• Their model has diminishing returns to execution, increasing returns to conceptual capacity, and complementarity between the two. AI substitutes for research support while amplifying the returns to accumulated expertise. They expect early convergence to give way to divergence as execution constraints recede and AI’s conceptual capabilities improve.

Labor evidence: compression on execution-heavy tasks

Where measured work is execution-heavy, generative AI compresses the productivity distribution. Brynjolfsson, Li, and Raymond study the staggered rollout of an AI assistant to 5,179 customer-support agents: average issues resolved per hour rise 14%, with 34% for novice and low-skill workers and minimal impact on experienced workers; treated agents with two months of tenure perform like untreated agents with six. ²

Gambacorta, Qiu, Shan, and Rees run the CodeFuse field experiment at Ant Group: code output rises more than 50% overall, statistically significant only among junior staff (about 67% for entry-level), with senior engagement lower — seniors request AI advice roughly 70% less often, and requests fall 13% per additional year of experience. When seniors do use the suggestions, acceptance-rate evidence shows they pay attention. ³

Read next to Jiang and Wang, the pattern lines up with the execution/conceptual split: on support tickets and junior coding, AI substitutes for execution and lifts the bottom of the distribution; in research, where output leans on conceptual capacity that accumulates with experience, AI raises the return to seniority. The synthesis is a hypothesis about task position, not a settled result — the studies measure different occupations, interventions, and outcomes.

Open questions

• Does the early convergence phase (RA-gap narrowing) persist, or does the divergence Jiang and Wang model take over as AI handles more of execution?
• Do the research-support findings transfer outside economics, where NBER replication packages and publication records give unusually clean measurement?
• The seniority gap appears in returns, not adoption. Whether juniors’ agentic-tool use closes that gap over time is untested.

Connections

• ai-economic-growth-complementarity — Cowen’s Intelligence × Polanyi-knowledge complementarity and Acemoglu’s task-based bounds model AI at the macro level; the Jiang–Wang execution/conceptual model is the same complementarity stated at the level of one profession’s tasks.
• knowledge-problem — the conceptual capacity that earns increasing returns is close to Hayek’s dispersed, tacit knowledge: the part of research that is hardest to hand to a tool.
• llm-replication-economics — machine-speed audit work is the execution side of the model at industrial scale: the same RA-substitution that narrowed support gaps now applied to 4,452 replication packages.
• saez-zucman-audit — Cowen’s status-shift claim about AI revising economic science is the distributional question applied to reputations.

Sources

• Tianyu Jiang, Yuqing (Claire) Wang 2026 — Equalizing or Polarizing? Generative AI and Research Productivity in Economics
• Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond 2023 — Generative AI at Work
• Leonardo Gambacorta, Han Qiu, Shuo Shan, Daniel M. Rees 2024 — Generative AI and labour productivity: a field experiment on coding

Footnotes

1. Tianyu Jiang, Yuqing (Claire) Wang 2026 — Equalizing or Polarizing? Generative AI and Research Productivity in Economics ↩

2. Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond 2023 — Generative AI at Work ↩

3. Leonardo Gambacorta, Han Qiu, Shuo Shan, Daniel M. Rees 2024 — Generative AI and labour productivity: a field experiment on coding ↩