AI and the research assistant: Evidence from economics working papers
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "AI and the research assistant".
Warren: Here's the gist of the paper. Artificial intelligence changes research-team size through scale and substitution, and their product has at most one peak The model reduces disagreement about AI and scientific employment to observables: coverage relative to s∗, the ordering of fields by task content, and a dated turning point.
Ruth: Model Framework The paper introduces a span-of-control model where team size is quasi-concave in AI capability with at most one peak.
Warren: This model reconciles the opposing effects of AI, which lowers execution cost and expands laboratory scale while automating codifiable tasks and lowering the member share of each unit.
Ruth: The model defines execution as requiring a unit measure of tasks, distinguishing between irreducibly human task share ϕ and automated task share σ(a).
Title and authors: Ruth: So, let's talk about the title "AI and the research assistant: Evidence from economics working papers" and who wrote this. It’s interesting that it’s framed as working papers, which suggests they are looking at real economic incentives rather than just pure computer science theory.
Warren: I noticed the authors list some folks like Johan Fourie, Willem Fourie, Jesse Naidoo, and Matthew Olckers. That's a pretty specific group of researchers working in this area.
Iris: From an econometrician standpoint, I wonder what kind of data these people are using to build this model. They mention their work with projects like Project STAR and deworming uptake in other papers, so we have a sense of their background.
Ruth: Right. It’s about taking those insights into economics—specifically markets and incentives—and applying them to how AI changes the structure of research teams in fields that are very codifiable.
Warren: What's the main implication for us as people who might be hiring or funding these research groups? Is it a warning sign about where we should be investing our resources?
Iris: It seems like it suggests that if you push AI adoption too far into codifiable tasks, you might hit a point where adding more people doesn't help as much, or actually hurts the efficiency.
The paper's summary: Ruth: The paper summarizes how they reconciled the opposing effects of AI—it expands laboratory scale and automates tasks, but it also lowers the individual member share.
Warren: So they’re saying that while AI makes labs bigger, it also changes how much work each person actually contributes to the final output.
Iris: They set up this mathematical framework where a leader has judgment ability C greater than zero and one unit of attention, and then they look at how execution E relates to that leader's attention.
Ruth: That framework leads them to Lemma one which defines the span of control, showing optimal attention is determined by the ratio of execution to the total effect <ref:2608.08437#pg2>.
Warren: They also introduce this concept of phi as irreducibly human task share and sigma as automated task share for a given field A.
Iris: The key insight here is that teams peak when their effective automation coverage hits a closed-form threshold, which they predict is typically near complete coverage.
Ruth: And they get a prediction about the ordering of fields: among fields with shared primitives, those with smaller irreducible human task content peak first.
The paper's improvements: Warren: Now for the parts where they suggest improvements to the original framework. They are proposing a way to predict when to stop adding new members based on an effective automation coverage metric reaching that s star threshold, which they estimate is around ninety-one percent coverage.
Iris: That sounds like a direct tool for resource allocation. If you can track your automation coverage against this threshold, you know when the benefit of adding another researcher diminishes significantly.
Ruth: Exactly. It's about giving teams a signal to stop growing when the marginal return on headcount drops off sharply due to AI taking over more tasks.
Warren: And there’s also this field prioritization engine idea, which ranks research fields based on that irreducibly human task share phi, suggesting we should focus on those with smaller phi first.
Iris: That makes sense from an identification standpoint because it targets the areas where the human judgment component is still most valuable right now.
Conclusion: Ruth: So to wrap up, the central forecast in "AI and the research assistant: Evidence from economics working papers" is that research teams in algebra and number theory, and theoretical computer science, are set to peak around two thousand twenty-seven or two thousand twenty-eight <ref:2608.08437#pg1>.
Warren: That timeline depends on whether we're looking at capability alone or if you add in realistic diffusion lags for how fast this technology actually spreads.
Iris: The paper is conditional on adoption; it forecasts a peak with about a fifty percent probability if long-run adoption hits between eighty and one hundred percent of codifiable tasks.
Ruth: So the big picture is that we're seeing a turning point based on when AI automates enough of the work that adding more people becomes less effective.
Warren: It’s a prediction about structure, telling us where those teams will naturally settle based on what they can automate versus what they have to do themselves.
Iris: It offers a concrete way to test these predictions by looking at coverage crossing that s star threshold before you assess the team's response.
Ruth: That’s it for this paper, "AI and the research assistant: Evidence from economics working papers." We’ll be looking at some other interesting stuff next time.
Johan Fourie
Department of Economics, Stellenbosch University
econ.GN, q-fin.EC
Submitted: 2026-08-09
Updated: 2026-11-02
Comments: 61 pages, 10 figures, 38 tables, including appendices
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: Artificial intelligence changes research-team size through scale and substitution, and their product has at most one peak The model reduces disagreement about AI and scientific employment to…
Key concepts
- Span-of-control model
- This model describes how the size of a research team changes based on AI capability. It accounts for AI lowering execution costs and automating tasks, which can lead to team growth at first but eventually plateauing or peaking.
- Irreducible human task share ($\phi$)
- This represents the portion of a research task that cannot be automated by AI; it requires human ingenuity. Fields with smaller $\phi$ are predicted to see their research teams peak earlier because they are more susceptible to AI augmentation.
- Effective automated task share ($\sigma(a)$)
- This measures how much of a specific research task is handled by automation. The model predicts that the fully codifiable team will peak when this share reaches a threshold $s^*$, which is predicted to be around 0.91.
- Calendar mapping
- This technique maps the capability clock to real time, incorporating realistic diffusion lags for AI adoption. This helps translate abstract capability predictions into concrete dates, such as forecasting the peak around 2028.5 when accounting for delays.
Terminology
Summary
Artificial intelligence changes research-team size through scale and substitution, and their product has at most one peak The model reduces disagreement about AI and scientific employment to observables: coverage relative to s∗, the ordering of fields by task content, and a dated turning point. p. 6
Model Framework
The paper introduces a span-of-control model where team size is quasi-concave in AI capability with at most one peak. This model reconciles the opposing effects of AI, which lowers execution cost and expands laboratory scale while automating codifiable tasks and lowering the member share of each unit. p. 0 The model defines execution as requiring a unit measure of tasks, distinguishing between irreducibly human task share ϕ and automated task share σ(a). p. 2
Key Predictions
The research yields three distinct predictions regarding the future of research teams.
-
Coverage predicts that a fully codifiable team peaks when its effective automated task share reaches a closed-form threshold s∗, with median 0. p. 191 under the priors below. p. 0
-
Ordering suggests that among fields with shared primitives, those with smaller irreducible human-task shares peak first. p. 1
-
Calendar mapping forecasts the fully codifiable peak to be around 2027. p. 42 on the capability clock, or 2028. p. 15 once realistic diffusion lags are added. p. 6
Evidence and Testing
The authors present evidence from 36,186 preprints across thirty OpenAlex subfields to test these predictions. The analysis involves two stages: Stage 1 establishes coverage by checking if effective coverage exceeds s∗, and Stage 2 reestimates the researcher-level specification to find a peak in author counts for specific fields. p. 5
Conclusion
The central forecast is that 2027 is the last year research teams grow in algebra and number theory and in theoretical computer science if capability alone governs adoption, and 2028 once realistic diffusion lags are added. The forecast is conditional on adoption, suggesting a peak occurs with probability of about one half if long-run adoption converges to a level between 80 and 100 percent of codifiable tasks. p. 6
Proofs and Details
The model relies on comparative static proofs, such as Theorem 1, which establishes the at most one peak condition based on the derivative of member employment with respect to automated task share. Monte Carlo simulations propagate priors across three scenarios—capability clock (Scenario A), diffusion-adjusted (Scenario B), and adoption ceiling (Scenario C)—to derive the dated predictions. p. 6 The test protocol requires confirmation of coverage crossing s∗ before assessing team response. p. 4
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Develop a team size scaling algorithm that predicts when to stop adding new members based on an effective automation coverage metric reaching s star, which is predicted to be near complete coverage (median 0.91). This system will inform resource allocation by signaling that further team growth is unlikely once this threshold is approached.
-
Implement a field prioritization engine that ranks research fields based on their irreducibly human task share phi, predicting that
those with smaller phi peak first.
This allows researchers to strategically choose domains where the highest impact from AI augmentation will occur earliest. -
Create a dynamic adoption forecasting model that predicts the timing of research team size changes by incorporating diffusion lags, such as the median forecast of 2028.5 when adding an adoption-and-output lag of 0.5 to 2 years is included, providing a more realistic timeline than capability clock alone (median 2027.2).
Sources
- Artificial Intelligence in Science: Returns, Reallocation, and Reorganization
- Measuring AI Ability to Complete Long Software Tasks
- OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts
- Resolution of Erd\H{o}s Problem #728: a writeup of Aristotle's Lean proof
Related papers
- Huge If True: Learning from the Size of the Surprise
- The Rebundling of Freelance Work in the AI Age
- Dynamic investment and coordination in teams: Theory and experimental evidence
- Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing
- Generative AI at Work
- What is Work For? Labor, Politics, and Meaning in the Age of AI