AI in Science

Summary

NBER w34953 (the PDF retrieved under this citekey is titled "AI in Science" by Agrawal, McHale and Oettl; bibliographic title-of-record differs — see Critique). The paper characterises AI as a tool "not for full automation but rather for augmentation through enhanced search over combinatorial spaces," and decomposes knowledge production into a multi-stage process to surface the "jagged frontier" of AI in science — differential returns across domains (data-rich biology vs. anomaly-sparse physics) and across workflow stages (e.g., AlphaFold-style design aids vs. subtler question generation).

Contribution

A task-based growth-style model distinguishing "ordinary" from "AI-expert" scientists, showing how exogenous improvements in AI yield nonlinear productivity gains amplified by the share of scientists who are AI-experts. The paper treats human judgement as indispensable for abductive inference, contextual nuance, and data-sparse trade-offs, motivating skills training and organisational design as essential AI complements.

Method

Conceptual / theoretical economics: a task-based model of the knowledge production function, organised around a multi-stage view of science.

Relevance to RISE

A macro-level framing for the RISE programme: it argues that gains from AI in science come not from full automation but from augmentation over combinatorial search, and that returns are jagged across stages — exactly the stage decomposition that RISE catalog projects (e.g. sakana-ai-scientist, kosmos) attempt to automate end-to-end. Implies evaluation of RISE pipelines should be per-stage rather than aggregate.

Critique / open questions

The PDF retrieved at this citekey appears to be NBER w34953 by Agrawal, McHale and Oettl ("AI in Science"), not the Cao et al. paper of the same title implied by the bibliography record — references.bib should be cross-checked. As a conceptual paper it offers no empirical calibration; the "AI-expert vs ordinary scientist" dichotomy is a useful but stylised partition.

Key quotes

"We characterize AI as a tool, not for full automation but rather for augmentation through enhanced search over combinatorial spaces. This leads to increased scientific productivity."

"We decompose knowledge production into a multi-stage process to shed light on the 'jagged frontier' of AI in science, revealing differential returns to different tools across domains … and workflow stages."