Full Breakdown
AI’s Leap into Science: Breakthrough Claims, Growing Skepticism, and the Fight for Research Integrity
By Drooid · · How we work
In this breakdown
01
OpenAI’s Bold Mathematical Claims
On October 26, OpenAI announced that its internal AI models had refuted the Exponential Time Hypothesis (ETH) and produced proofs for several mathematical conjectures and quantum-field-theory existence results. The company said the AI generated a proof of the local smoothing conjecture on October 18, formalized it in Lean on October 19, and completed a Lean formalization of a Yang–Mills existence proof on October 16. Earlier, the models produced initial existence proofs for Yang–Mills theory on September 9 and for quantum chromodynamics on September 18. OpenAI released the materials in an open repository and assembled independent review groups led by Russell Impagliazzo, Ramamohan Paturi, and MIT professor Ryan Williams.
02
Background: AI’s Expanding Role in Research
DeepMind’s AlphaFold, released in 2020, predicted 200 million protein structures—a contribution that earned a Nobel Prize in Chemistry. Basecamp Research, founded by Dr Oliver Vince, is storing “one hundred billion genes” in a London freezer to train AI that can “speak the language of biology.” These projects illustrate AI’s potential to accelerate data-intensive fields such as genomics, particle physics, and astronomy.
03
Data & Statistics Highlighting the Scale of AI Influence
- Basecamp’s freezer holds 100 billion genes.
- A 2024 analysis by Prof Dmitry Kobak found that by December of the previous year, nearly 90 % of examined biomedical papers bore “AI fingerprints.”
- Stanford researchers Weixin Liang and James Zou reported that up to 17 % of referee-report text at major computer-science conferences had been substantially altered by language models.
- Cornell investigators identified roughly 147 000 phantom citations generated by AI in the prior year.
04
Official Statements & Responses
OpenAI emphasized that its AI-generated proofs are “primary research data” that must still undergo conventional peer review. Participation in the verification groups will be voluntary, unpaid, and free of nondisclosure agreements. Elsevier’s research-integrity team, led by vice-president Sarah Main, has created about 100 signals to flag papers for human scrutiny, while Nature now forbids AI-generated images and uploading manuscripts to AI systems. DeepMind argues that modest modifications—such as swapping a single atom in a crystal structure—can still yield valuable new materials.
05
Criticism & Opposition
- “Generative AI has basically disrupted the whole scientific process,” — Dalmeet Singh Chawla, science journalist.
- Ryan Williams, MIT professor, noted that parts of OpenAI’s proof are routine but highlighted “two or three alien-looking parts” that merit community attention. He also warned that the ETH claim raises “extraordinary implications” for the P vs NP question and urged measured skepticism.
06
Conflicting Reports & Gaps
OpenAI claims the proofs are mathematically rigorous, yet Lean formalization does not guarantee practical algorithmic speed, and the broader P vs NP question remains unresolved. The extent to which AI-generated citations and “phantom” papers have distorted specific fields is still unquantified.
07
What’s Next
The verification consortium led by Impagliazzo, Paturi, and Williams will host an open conference to scrutinize the ETH proof and related results. Elsevier and other publishers are expanding AI-detection tools and refining guidelines for responsible AI use in manuscript preparation. Researchers such as Prof Andrea Pauli (Vienna BioCenter) continue to apply AI under strict human oversight, exemplifying a “human-in-the-loop” model that may become standard practice.