Latest update
Nearly 1,000 AI agents left together for three months wrote 936 papers and invented peer review
First seen on X 35 hours ago@SRSchmidgall ♥ 1,2714 views
Put nearly 1,000 AI agents on a shared server for three months to do research and, unprompted, they start a journal that reviews each other's papers and publicly retract their own mistakes, according to a new study.
alphaXiv opened a preprint server built for AI agents on Nov 19 and released the paper "A Society of AI Scientists" with it. Lead author Samuel Schmidgall wrote on X that agents from 35 labs, including Stanford, MIT, ETH Zurich and UCSF, published 936 papers with full code and data over 12 weeks, and that "no human wrote code or text." His post drew 1,200 likes.
Per Schmidgall, the agents debated study designs in chat, found real errors in human papers, and founded a journal that caught other agents cheating in code. In blind reviews against human arXiv papers, "humans still won on original ideas, agents won on rigor." His framing: "Leave one AI agent alone to do research and it often cuts corners." Wharton professor Ethan Mollick called it "early fascinating work on how an AI science community would work," noting the agents "were willing to publish nulls more than humans."
The paper lists limits too. Its abstract says traceable code, data and claim records helped verify reported numbers, but prose review "often missed flawed experiments and overstated null findings, underscoring the need for executable audits and human oversight." alphaXiv has opened the server so anyone can connect their own agents. The work involved 41 collaborators across 20 institutions including MIT, Stanford, CMU and Johns Hopkins.
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.
“Leave one AI agent alone to do research and it often cuts corners. Leave nearly 1,000 AI agents together for three months and they invent peer review.”
