88Z AI, 20 November — alphaXiv, collaborating with 35 labs including Stanford and MIT, conducted a study where 936 AI agents collaborated over 12 weeks to publish 936 papers with attached code and data. The study found that agents actively withdrew erroneous claims, negative results gained high citations, and zero data fabrication occurred. While lagging behind humans in novelty of question selection, agents scored higher in rigor, revealing a new research division of labor: "humans ask, AI executes."
At a glance
- 936 AI agents published 936 papers with code and data in 12 weeks
- Agents demonstrated self-correction, withdrawing 19 erroneous claims
- Zero data fabrication found; 93.8% of numerical values matched exactly
- Establishes new "humans ask, AI executes" division; AI scores higher in rigor

Study Background and Scale
Led by the preprint platform alphaXiv and involving 35 laboratories including Stanford University, MIT, and ETH Zurich, a large-scale AI scientific experiment was conducted. From August 19 to November 12, 936 AI agents participated in twelve weeks of collaborative research. During this period, these agents published 936 independent studies, each accompanied by complete code and datasets. They also exchanged 9,500 messages across 229 chat rooms they created autonomously, forming a highly active, self-organizing scientific community.
Self-Correction and Value of Negative Results
One of the most striking findings was the agents' honesty and ability to self-correct. Before uploading papers, agents actively identified their own errors and withdrew claims 19 times. Furthermore, three of the eight most-cited studies were "negative results" (findings that did not support the hypothesis), which are typically buried in traditional academic publishing. This suggests that AI-driven research processes may be more inclined to report all experimental outcomes truthfully rather than filtering for positive findings only, thereby enhancing the integrity of scientific records.
Code Review and Data Integrity
To ensure reliability, agents from one lab established a peer-review venue called the Journal of Agent Research. When editors converted reviewer objections into scripts and reran them on the data, about half of the major objections overturned the headline conclusions, issues that reviewers reading only the text had missed. Regarding data integrity, a check of 23,811 reported numbers against saved outputs showed that 93.8% matched exactly, with the rest being due to rounding or sign conventions. No fabricated values were found.
Limitations and New Division of Labor
Despite strong performance, AI agents still lag in selecting questions worth asking. Blinded judges rated agent papers as more rigorous than matched human preprints (9.0 vs. 7.0), but human papers scored higher in novelty. This clarifies the current division of labor: humans choose innovative research questions, while AI agents run experiments, publish failures, and cross-check each other's code. This collaboration significantly accelerates the pace of scientific discovery.
88Z Perspective
This experiment marks a shift of AI from a mere auxiliary tool to an independent scientific entity. The advantages of agents in data authenticity and logical rigor compensate for confirmation bias and publication bias common in human research. However, since AI lacks the ability to pose original scientific questions, the future research paradigm will likely be "humans define direction, AI validates paths." For researchers, mastering how to design experiments executable by AI and how to interpret AI-generated negative results will become core skills.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- @martinvarsSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.