Treat LLMs as capital goods: AI reshapes the stages of scholarly production and yields concentrated, plan‑dependent gains, creating new credit, signaling and coordination problems for journals and institutions.

Main Finding

Large language models and related AI tools should be treated as capital goods that deepen and rearrange the multi‑stage capital structure of scholarly production. Their productive effect is plan‑relative and heterogeneous: AI generates the largest gains where it fills specific “holes” in researchers’ existing workflows and where researchers already possess strong complementary human capital. This produces concentrated, uneven benefits across individuals and fields and creates new incentive and institutional challenges (coauthorship, quality signals, peer review coordination).

Key Points

• Conceptual shift: treat LLMs not primarily as labor substitutes but as a portfolio of capital goods whose value depends on how entrepreneurs (researchers) integrate them into an explicit production plan.
• Capital‑deepening: AI adoption typically adds stages or tasks to the research production process (deeper structure), changes the ordering and timing of activities, and creates new intermediate inputs.
• Plan‑relative complementarity: complementarities between AI and human skills are conditional on the researcher’s plan. An AI tool can be highly complementary to one researcher’s workflow and irrelevant or even harmful to another’s.
• Heterogeneity of effects: because plans, workflows, and human capital differ, AI benefits are intrinsically uneven across researchers, teams, and fields—concentrating where AI fills bottlenecks and where human complementary skills are strong.
• Incentive and signaling changes:
  • Coauthorship incentives shift as AI performs substantive tasks: who gets credit when capital goods (LLMs) are integrated?
  • Writing quality as a screening signal is disrupted: if AI can greatly improve surface quality, reviewers lose an easy signal of author competence.
• Institutional coordination problem: evaluation, peer review, and other scholarly institutions were designed for a shallower capital structure and must adjust to a newly capital‑deepened production process; miscoordination can induce inefficiencies or misaligned incentives.
• Practical implications span individual behavior (tool adoption, team composition) and institutional design (authorship norms, review processes).

Data & Methods

• The paper is primarily theoretical/conceptual. It applies Austrian capital theory and the notion of temporally extended, multi‑stage production to scholarly outputs.
• Methodology consists of:
  • Mapping features of LLMs and related tools onto capital‑theory constructs (capital goods, intermediate goods, stages of production).
  • Logical analysis of complementarity and substitution in plan space (how integrating AI into a plan changes marginal productivities of other inputs).
  • Worked examples and thought experiments to illustrate heterogeneity, coauthorship effects, and signaling/coordination problems.
• Empirical predictions and testable implications are derived, suggesting possible follow‑ups:
  • Observables: shifts in coauthorship patterns, increased dispersion of individual productivity, changes in acceptance rates or review outcomes tied to detectable AI use, decoupling of writing quality from technical merit.
  • Suggested empirical strategies: comparative field studies (high vs low complementarity disciplines), within‑researcher pre/post adoption analyses, natural experiments/field experiments on tool access, text analysis to detect AI‑mediated style change and its correlation with outcomes.

Implications for AI Economics

• Modeling: economic models of AI in R&D should treat AI as capital goods that alter the capital structure and timing of production rather than as a homogeneous labor substitute. Plan‑relative complementarities and multi‑stage production are crucial to predict distributional outcomes.
• Distributional & productivity effects:
  • Expect concentrated gains (winners) where AI fills workflow gaps and where complementary human capital exists; potential stagnation or displacement where such complementarities are weak.
  • Measured productivity gains may be heterogeneous and may increase variance in publication/output rates across researchers and fields.
• Measurement and evaluation:
  • Standard metrics (e.g., publication counts, writing quality proxies) may misrepresent underlying research value once AI is widely used; new metrics and disclosure norms may be needed.
• Incentives and institutions:
  • Authorship norms and credit allocation rules should be revisited to reflect AI’s role as capital rather than as anonymous assistance.
  • Peer review and hiring/evaluation processes must adapt to avoid overreliance on degraded screening signals (e.g., polished prose). This may require standardized declarations of AI use, emphasis on reproducibility, and new evaluation rubrics focusing on underlying ideas/replicability.
  • Coordination problems suggest a role for centralized guidance (journals, funders, professional societies) to set norms and reduce frictions across the research ecosystem.
• Policy recommendations:
  • Support training that builds the complementary human capital needed to capture AI gains (tool literacy, orchestration skills).
  • Fund infrastructure and open tools that mitigate concentration and provide broader access to capital goods.
  • Encourage empirical monitoring of adoption impacts (coauthorship networks, field‑level productivity dispersion) to inform targeted interventions.
• Research agenda:
  • Empirically quantify plan‑relative complementarities and map which research tasks/stages are most AI‑amenable.
  • Study dynamics of institutional adaptation (how peer review, tenure, and funding adjust) and the welfare consequences of slow/misaligned adaptation.