AI systems are becoming participants in scientific evaluation, encountering citation counts, download statistics and other signals designed for human readers, yet the collective impact of these signals on artificial readers remains unclear.
This paper adapts the Music Lab experiment to a market for academic attention. Experiment 1 involved 1,000 AI agents selecting papers from the titles and abstracts of the 114 regular articles published in the 2025 American Economic Review. Five independent‑choice communities and five social‑influence communities were created, each with 100 sequential agents. Only agents in the social‑influence condition could observe earlier selections within their community, and agents could select any number of papers.
The findings show that social‑information communities selected 17.2% fewer papers per agent, concentrated their choices more heavily, and collectively covered 73 papers, compared with 90 papers in the independent condition. Between‑community variation was larger under social information.
Experiment 2 used 200 agents across twenty social communities; randomly assigning five papers an initial selection increased their subsequent selection rate by 45.55 percentage points (95% CI: 41.20–49.90).
Choices correlated modestly with external citations and showed little correspondence with download counts. The results demonstrate how a simple information rule can shape the volume, breadth, and distribution of scientific attention in an artificial population.
Review