Skip to content

Arin Dube Thread: LLMs Haven't Raised NBER Working Papers Above Trend

Key Ideas

  • The advent of LLMs has not raised the number of NBER working papers above trend, nor submissions to top economics journals
  • Dube's explanation: LLMs substitute for good RAs, but not for good ideas -- and RA labor supply was never the binding constraint in economic scholarship
  • Claude Code has made specific tasks faster and better, but these are components of a broader production process
  • The "hard task vs. easy task" framing (citing Chad Jones) suggests overall productivity gains from AI may be seriously constrained by bottleneck tasks
  • Respondents note SSRN and arXiv preprints have increased, suggesting effects may appear first in less-curated outlets
  • Some argue it is too early to judge since agentic capabilities only emerged three months prior; Dube cautions against over-reliance on threshold-effect thinking

Summary

Arin Dube presents data showing that LLM adoption has not yet produced a measurable increase in NBER working papers or top journal submissions above historical trends. His core argument is that LLMs are effective substitutes for research assistant labor on specific tasks, but the binding constraint in economics research has always been good ideas, not RA labor supply. He draws on Chad Jones's framework of "hard" versus "easy" tasks to argue that even substantial AI-driven productivity gains on easy tasks (coding, data cleaning, writing) may not translate into large overall productivity gains if the hard tasks (generating novel ideas, identifying causal mechanisms) remain bottlenecks.

The thread generated significant discussion. Several respondents pointed to increases in SSRN and arXiv preprints as evidence that AI effects are showing up in less-curated channels. Megan Stevenson noted that agentic capabilities had only emerged three months prior, making it too early to judge. Dube acknowledged this but pushed back on the tendency to always claim "we're just three months into a new world," arguing this reflects an AGI-influenced over-reliance on threshold effects. Other respondents noted practical limitations like restricted-access data that cannot be used with AI tools.

Relevance to Economics Research

This thread addresses a fundamental question about AI's impact on the economics profession: does making research tasks easier actually lead to more research output? The evidence so far suggests not, at least at the top-journal level. This has implications for how economists should think about AI adoption -- as a tool for task-level productivity rather than a transformative force for research output. The hard-task/easy-task framework provides a useful lens for understanding why productivity gains at the micro level may not aggregate to macro-level changes in scholarly output.