Live·Open questions in longevity research
All news
Scientific ComputingScience Research

Claude Fable 5.2 chose a research question, found data, and wrote a 182-page paper: Ethan Mollick spent less than an hour on it, and no human has fully checked the results

10 November 2026· 261110010

Claude Fable 5.2 chose a research question, found data, and wrote a 182-page paper: Ethan Mollick spent less than an hour on it, and no human has fully checked the results

On November 9, Ethan Mollick, a professor at the Wharton School, described an experiment with Anthropic’s Claude Fable 5.2 model in the One Useful Thing newsletter. He asked it to choose an unresolved research question, preregister its hypothesis and methods (record them on an external platform before analyzing the data), write a paper, and submit it to other AI agents for critique.

The result was a paper titled “The One-Dollar Company,” which tests whether an increase in new company registrations means that more companies are operating. Registration counts feature in debates over registration fees: the Colorado Secretary of State’s report for the third quarter of 2014 opens with “Record growth in new registrations driven by a fee reduction”. For several months in 2012–2013 and 2014, Colorado charged 1 dollar for registration instead of 50, and registrations rose by an average of 17%. In May 2024, the United Kingdom raised its fee from 12 to 50 pounds, and registrations fell by 17%.

The paper estimates that companies whose formation depended on the fee were less likely than other companies to leave evidence of business activity: loan and collateral records in Colorado, and filings required of operating companies in the United Kingdom. If these companies had behaved like the others, the number of Colorado companies with a credit record would have risen by the same 17%. The paper instead estimates an increase of 0%, with a margin of error of ±7%. It therefore recommends that registries also publish the number of companies showing their first sign of business activity.

The model wrote the analysis plan and four amendments on October 4 and 5, with the data already available. Each document was written no more than a few hours before the calculations it specified. The documents are absent from any external platform, and their dates are supported only by the version history in the model’s own project. The paper describes them as “a dated record of our intentions” and does not claim that they constitute preregistration. The Colorado test was added to the plan when its result was already foreseeable. The UK test was added while data on companies formed from 2024 onward were still in a sealed file, so its prediction could have proved wrong.

Mollick’s initial impression was that the work exceeded what he would expect from second-year doctoral students, although it was unsuitable for publication in its current form. It would have taken a human months; Mollick spent less than an hour on it. Of the 182 pages, 31 contain the main text, and the rest contain references and appendices. According to Mollick, the model responded to every criticism with another analysis. At the AI reviewers’ request, it added a measure derived from third-party records and a test using data it had not seen. Those requests produced the two main tests.

Mollick counts the paper’s length among the model’s failures. The reviewers it assigned to almost every project found errors, but did not ask whether the work was too long or repetitive. “The ability to recognize that your work is not good enough is a large part of what makes someone an expert, and it is the part Fable handles least reliably,” Mollick concludes.

The AI reviewers were other instances of the same model. They reran the analyses and both main tests. The paper describes these as “quality checks, not independent verification,” and states in footnote 1 that no human has checked the results or rerun the code.

“AI can now produce research faster than anyone can read it, and this is creating a crisis across much of academia,” Mollick writes.
Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#claude-fable-5-2#autonomous-research#preregistration#ai-peer-review#research-verification