Three new arXiv preprints, released within days of each other, tackle different fronts of AI’s expanding footprint: cognitive benchmarking, scientific peer review, and clinical deployment. They share a common thread—each asks what happens when AI systems are treated not just as tools but as participants in activities that have historically been human.
The first, CogGym, frames AI and cognitive science as pursuing parallel goals and proposes a large-scale comparative evaluation of human and machine cognition. The second examines a feedback loop in academic publishing: as LLM-generated reviews enter public datasets and future training corpora, the authors argue, the scientific judgment of AI reviewers could collapse, and they explore mitigation strategies. The third takes a sociotechnical view of algorithms in health systems, emphasizing that technical performance alone is insufficient and that cost and human-centered considerations must be part of the picture.
Where the papers differ is in emphasis. CogGym is primarily about measurement and scale; the peer-review paper is about systemic degradation over time; the health review is about implementation and governance. But all three converge on a similar warning: AI’s value depends on how it is embedded in human institutions, and unchecked feedback or narrow technical metrics can undermine the very processes AI is meant to improve.