Two papers this week frame the same problem from opposite ends. AutoScientists argues that multi-agent research systems need governance mechanisms before they need more capability. Tang and Yang show that the agents we already have cluster tightly around existing work rather than opening new directions. Meanwhile, Scientific Reports is still retracting papers where "surface tension" was written as "floor anxiety" and no reviewer noticed (yes, really). The capability question is increasingly settled. The infrastructure question is not.
AutoScientists: Decentralized AI agents for autonomous research
arXiv, May 2026
Harvard preprint where multi-agent AI scientists self-organize around shared experiments, evaluate proposals before allocating compute, and document failures alongside successes, arguing that the bottleneck for autonomous research is governance, not raw capability.
AI Research Agents Narrow Scientific Exploration
arXiv, May 27 2026
Analysis of 37,802 AI-generated scientific ideas across four agent frameworks found that outputs cluster more tightly than human papers, stay closer to source material, and recombine existing methods rather than posing new questions.
Biohub releases a world model of protein biology
CZ Biohub, May 27 2026
ESMFold2, trained on 2.8 billion sequences, leads standard benchmarks on protein-protein and antibody-antigen interactions and validated designed binders in the lab with hit rates of 36-88%, all released fully open-source.
Scholarly AI Search Shortcomings and the Need for Better Metadata
The Scholarly Kitchen, May 29 2026
AI discovery tools rely on titles and abstracts from OpenAlex, Crossref, and Semantic Scholar, which are too thin for reliable retrieval, and publishers are restricting metadata access rather than expanding it.
Scientific Reports, May 20 2026
Retracted for AI-generated nonsense terminology throughout: "Newtonian drinks" for Newtonian fluids, "skinny movie" for thin film, "floor anxiety" for surface tension.
Scientific Reports, May 2026
Same journal, same window, same failure mode: nonsensical terminology, inaccurate equations, authors unable to respond to concerns, making this a pattern of AI-era errors clearing peer review rather than a one-off.
APC Caps and Bans: Why Funder Policies Aimed at Curbing the Publishing Industry Don't Work
The Scholarly Kitchen, May 14 2026
Funder APC caps and bans get routed around: career incentives override funder loyalty, transformative agreements move costs to library accounts, and Gates's VeriXiv captures only 35-39% of its funded portfolio.