Harvard physicist used Claude to co-author 36 papers in three months

By Splitfeed EditorialLast updated Oct 3, 2026, 09:30 UTC

On October 2, 2026, Harvard particle physicist Matthew Schwartz released 36 manuscripts he had co-authored with Anthropic's Claude over a three-month period. The papers span 18 fields, ranging from particle physics to linguistics. Schwartz used BootLoops, an open-source tool designed to help AI models perform precise scientific calculations, as the technical harness for the work. The volume alone, roughly three papers per week across disciplines far outside any single expert's training, is what drew immediate attention.

The story matters because it sits at the intersection of two questions that have been circling academic science for years: how much of the research process can AI reliably automate, and what does authorship mean when it does. Schwartz's own framing complicates a straightforwardly triumphant reading. He stated that the manuscripts only became scientifically valuable after human experts reviewed them, and he offered a pointed warning to anyone considering replicating the workflow: look at everything yourself. That caveat is not a minor footnote. It implies the raw AI output carried real risk of error, and that the 36 papers are partly an argument for human oversight rather than against it. What remains genuinely in dispute is whether this represents a net acceleration of science or a net increase in the burden placed on peer reviewers and domain experts who must now filter a much larger volume of AI-assisted submissions.

None of the three camps have published direct reactions to this story yet, so their likely positions can only be sketched from the patterns each typically follows. The Pro-AI camp would be expected to highlight the 36-paper output as evidence that AI is already functioning as a genuine research multiplier, compressing years of exploratory work into months. The Anti-AI camp would be expected to focus on Schwartz's own warning, arguing that shifting the verification burden onto human experts is not a productivity gain but a redistribution of labor and risk. The Middle Ground camp would be expected to treat BootLoops and the human-review requirement together as the relevant unit, framing this as a demonstration that AI tools are useful precisely when embedded in rigorous human oversight rather than treated as autonomous.

The argument will sharpen once the manuscripts begin moving through peer review. If a meaningful share clears scrutiny in fields where Schwartz has no direct expertise, that would support the Pro-AI reading. If reviewers surface systematic errors that Schwartz's own checking missed, the Anti-AI critique gains ground. The reception of the papers over the coming months is the concrete outcome worth watching.