A Withdrawn AI Research Study and Questions About Scientific Discovery

ZUSAMMENFASSUNG

In 30 Sekunden

Matthew Berman speculates about AI making scientific discoveries and cites a study claiming productivity gains and reduced researcher satisfaction. In 2025, MIT said it had no confidence in the study’s data or research, and arXiv withdrew the preprint; its figures are not reliable evidence.

1A speculative futureBerman imagines AI making discoveries and recursively improving itself. This is a hypothetical scenario.
2Study later withdrawnThe cited preprint claimed gains in materials and patent filings, but MIT disavowed its findings and arXiv withdrew it.
3Satisfaction claims are unreliable tooThe reported satisfaction statistic comes from the same disavowed research. The video also states a product-innovation figure that differs from the preprint.

Das Wichtigste

MIT said it had no confidence in the data or findings of the preprint cited in the video, and arXiv withdrew it.
The video’s productivity and satisfaction percentages should not be treated as established results.
The video states 177% for product innovation, whereas the withdrawn preprint’s abstract stated 17%; neither figure is a reliable finding.
Berman’s proposed path from AI discovery to recursive self-improvement is speculation, not a conclusion supported by this study.

ThemenAI for science AI research Future of Work

VOLLTEXT

Den bearbeiteten Text nach Kapiteln lesen

Transkript zeigen

Nur das Transkript

AI and scientific discovery

Berman imagines AI making discoveries and improving itself. This is speculative rather than established by the cited study.

**Correction:** The study discussed in this 2024 video was later withdrawn. In May 2025, [MIT said](https://economics.mit.edu/news/assuring-accurate-research-record) it had no confidence in the data’s provenance, reliability or validity, or in the research’s veracity; [arXiv withdrew the preprint](https://arxiv.org/abs/2412.17866). Its reported results should not be relied on. I open by imagining AI making scientific discoveries and applying them to recursive self-improvement, potentially leading to an intelligence explosion. That is a speculative possibility, not an established consequence of the study.

A study later withdrawn

He cites dramatic productivity claims from a preprint. MIT later said it lacked confidence in the data and the research, and arXiv withdrew the paper.

I cite a study that claimed to examine AI tools used by more than 1,000 scientists, reporting 44% more materials discoveries and 39% more patent filings. These were claims of the withdrawn preprint, not reliable measurements of AI’s effects. The video gives product innovation as a 177% increase; the preprint abstract instead stated 17%, but that figure is also part of research MIT says should not be relied on.

Work satisfaction and human skills

Berman cites a satisfaction claim from the same unreliable preprint and asks about human creativity. The figures cannot establish those effects.

I also say that 82% of scientists were less satisfied because of reduced creativity and use of their skills. This statistic comes from the same withdrawn, disavowed study and cannot establish that outcome. I close by asking how AI handling coding and science might affect human skills and creativity; the question remains open, independent of this unreliable evidence.