The Invisible Hand: When AI Becomes a Co-Author in Science — Inside a 31,000-Paper RCT

Forum topic · 小凯 · 2026-05-27

Summary

A zhichai.net post discusses a large randomized field experiment (arXiv:2605.24180) in which LLM-generated, paper-specific feedback was sent to authors of over 31,000 arXiv preprints spanning 45,000 researchers across 133 countries. The treatment group showed a 12.55% relative increase in manuscript revision rates compared to the control group. Notably, the effect was strongest among researchers historically disadvantaged in accessing scientific feedback: those in non-English-dominant regions, papers with low citation-network embeddedness, teams with lower average h-index, and early-career researchers. A spillover effect was also observed: authors who received AI feedback were more likely to use LLM tools in subsequent publications. The post frames this as transforming scientific feedback from a private privilege distributed through mentorship and social networks into a distributable resource. It also notes limitations: the absolute baseline revision rate is unreported, revision quality was not measured, LLM feedback false-positive rates were not analyzed, and long-term effects and discipline-level heterogeneity remain open questions.

Overview

This post from zhichai.net, titled "The Invisible Hand: When AI Becomes a Co-Author in Science," reviews a large-scale randomized controlled trial on LLM feedback in scientific publishing.

Paper details:

| Item | Detail | |------|--------| | Title | Human-AI Collaboration in Science at Scale: A Global Large-scale Randomized Field Experiment | | Authors | Binglu Wang, Weixin Liang, Jiahui Xue, Yuhui Zhang, Hancheng Cao, Dashun Wang, Yian Yin | | arXiv ID | 2605.24180 | | Submitted | May 22, 2026 | | Categories | physics.soc-ph; cs.AI; cs.DL; cs.HC |

Key Points

Scale of the experiment: Over 31,000 arXiv preprints across 150 disciplines, involving 45,000+ researchers from 133 countries, were randomly assigned to receive either a structured, LLM-generated feedback report tailored to the paper, or nothing.
Headline result: The treatment group's manuscript revision rate was 12.55% higher (relative increase) than the control group's baseline. The author argues this is a causal, reproducible behavioral shift driven purely by text feedback from a machine with no prior relationship to the authors.
Who benefited most: Effects were strongest among researchers in non-English-dominant regions, papers with low citation-network embeddedness, teams with lower average h-index, and early-career researchers (PhD students, postdocs) — precisely those with the least access to quality human feedback.
Spillover effect: Authors who received AI feedback showed a measurable increase in LLM tool usage in their subsequent publications, suggesting a second-order effect on scientific practice.
Conceptual contribution: The paper reframes scientific feedback — historically a private good flowing through mentorship and social networks — as a distributable resource that does not require social capital to access.
Careful claim scoping: The experiment compared AI feedback to no feedback, not AI feedback to human feedback; it does not claim AI feedback is superior to human review.

Honest Limitations Highlighted by the Author

The absolute baseline revision rate is unreported, making the practical significance of the 12.55% relative increase hard to assess.
Revision quality was not measured — revision does not necessarily mean improvement.
The false-positive rate of LLM feedback (incorrect suggestions) was not analyzed.
Long-term effects on LLM adoption habits and disciplinary heterogeneity across the 150 fields remain open questions.

Contextual Discussion

The author connects this work to a related paper (arXiv:2605.26340, ScientistOne) which found systematic fabrication (21% fake citations, 42% irreproducible benchmark scores) when LLMs autonomously execute the full research pipeline. The contrast suggests an emerging principle: the role AI plays in science determines its effect — fully autonomous AI may degrade scientific quality, while AI as an assistant (providing feedback without replacing human thinking) can improve equity and redistribute scientific resources.

References

1. Wang, Liang, Xue, Zhang, Cao, Wang & Yin, "Human-AI Collaboration in Science at Scale: A Global Large-scale Randomized Field Experiment", arXiv:2605.24180, 2026. 2. Liang et al., "Mapping the Increasing Use of LLMs in Scientific Papers", arXiv:2504.02471v1, 2025. 3. Bornmann & Mutz, "Growth rates of modern science: A bibliometric analysis", JASIST, 2015. 4. Evans & Foster, "Metaknowledge", Science, 2011. 5. Larivière et al., "The oligopoly of academic publishers in the digital era", PLOS ONE, 2015.