Economists Who Disclose AI Credit 0.55 More Research Assistants

When an economist discloses that a working paper used AI, that paper thanks 0.548 more named research assistants than the same economist's papers without a disclosure, against a baseline of 0.83, and runs 5.35 pages longer on a baseline of 52.4. [Fact] The junior rung of economics has not vanished from the acknowledgements; on disclosing papers it has grown.

The paper is "AI and the research assistant: Evidence from economics working papers," by Johan Fourie of Stellenbosch University, first posted to arXiv on August 9, 2026 and revised on November 2 (arXiv:2608.08437). [Fact] Fourie reports that Claude, Gemini and GPT models coordinated, drafted and checked parts of the manuscript itself. [Fact]

56,585 working papers, read for their thank-you lines

The data are 56,585 paper records from 24 working-paper series, 23 institutional series from university departments, central banks and research networks plus the economics arXiv, catalogued on November 1, 2026 and covering 2015 through October 2026. Series that could not be retrieved at a 98% rate were dropped, among them Chicago BFI, Yale Cowles and Stanford SIEPR. [Fact] Two AI indicators enter together: an explicit statement that the authors used AI, and a flag from a detector of AI-like economics prose. The main regressions compare post-2022 papers by the same recurring economist, with researcher, source-quarter and field fixed effects, so differences in funding or in the habit of thanking people drop out. [Fact]

Across institutional series, the share of papers crediting any assistant stayed between 27% and 32% in every half-year from 2022 to 2026 while AI disclosure rose from zero to about one paper in six; across all sources it reaches about one in five in the July to October 2026 cohort. [Fact] The two indicators barely overlap: about one in seven disclosing papers carries the prose flag, and about one in eight flagged papers discloses. [Fact]

Coding and data work, not prose, carries the difference

Disclosures split roughly five to one between operational statements, covering coding, data, analysis or derivation, and writing-only statements. Operational disclosures are associated with 0.679 more named assistants within a researcher's portfolio; writing-only disclosures show -0.098, indistinguishable from zero, and the two coefficients differ at p = .0034. [Fact] Operational disclosures also add 6.78 pages against 0.83 for writing-only, and most of the added length sits in appendices rather than main text. [Fact] The prose-detector flag is associated with no change in assistants, authors or pages; every interval includes zero. [Fact]

The named-assistant and page differences survive the Romano-Wolf and Holm corrections for the registered family of twenty tests. Two other differences, a 6.6-point higher probability of crediting any assistant and 0.24 more byline authors, are individually significant but do not survive the adjustment. [Fact]

Fourie's reading is that cheaper execution both cuts the hours a fixed study needs and makes a wider study worth running, with extra datasets, specifications and appendix tables that bring new names into the acknowledgements. Writing help that comes after the empirical work is done changes neither. [Claim] He contrasts this with Jiang and Wang (2026), who find across economists that AI narrowed the gap between researchers with and without assistants; substitution for an absent assistant and expansion with a present one are different margins of the same technology. [Claim]

What it means for your occupation

The people counted are the pre-doctoral fellows and student assistants our site lists as social science research assistants, working for economists and postsecondary economics teachers. Fourie distinguishes two bundles that the same thank-you line can cover: building an empirical pipeline from raw files, which teaches a novice to spot data anomalies and identification threats, and checking model-generated scripts and auditing appendix numbers, which teaches verification without that foundation. [Claim] Annual portfolios show mean credited assistants per paper dipping in 2023 and rising above the 2022 level by 2025, with distinct assistants per economist only modestly above one. [Fact]

What to prepare

If you are seeking a research-assistant post, target principal investigators who disclose operational AI use: in this data those are the projects with more credited assistants and larger appendices. Ask which bundle the role is, pipeline-building or output-checking, and push for the first, since that is the part Fourie identifies as the apprenticeship. Document your own contribution in a form that survives as a credit, because an acknowledgement is the only record this study could count.

Limits

The unit is a credited contribution on a finished working paper by a recurring author, not paid hours, headcount or department hiring, and the disclosure variation rests on about 3.3 effective sources. Disclosure is voluntary and decided by the author team, so selective silence remains in every comparison; the paper is a working paper without peer review.

Sources

• Fourie, J., "AI and the research assistant: Evidence from economics working papers," arXiv:2608.08437, working paper, Stellenbosch University, first version August 9, 2026, revised November 2, 2026: https://arxiv.org/abs/2608.08437

About this analysis

This article was written with AI assistance from the full text of the arXiv PDF (61 pages, version 2). Occupation links point to our job pages; the mapping from the paper's roles to those pages is ours.