David Gerard
Feb. 22, 2025, 6:25 p.m.
The hype cycle for Google’s fabulous new AI Co-Scientist tool, based on the Gemini LLM, includes a BBC headline about how José Penadés’ team at Imperial College asked the tool about a problem he’d been working on for years — and it solved it in less than 48 hours! [BBC; Google]
Penadés works on the evolution of drug-resistant bacteria. Co-Scientist suggested the bacteria might be hijacking fragments of DNA from bacteriophages. The team said that if they’d had this hypothesis at the start, it would have saved years of work.
Sounds almost too good to be true! Because it is. It turns out Co-Scientist had been fed a 2023 paper by Penadés’ team that included a version of the hypothesis. The BBC coverage failed to mention this bit. [New Scientist, archive]
Google’s other claimed successes for Co-Scientist follow this pattern. The system proposed new drugs for liver fibrosis — but the proposed drugs had previously been studied for this use case.
In 2023, Google loudly publicised how DeepMind had synthesized 43 “new materials” — but studies in 2024 showed that none of the materials was actually new, and that only 3 of 58 syntheses were even successful. [APS; ChemrXiv]
“Everything was already published, but in different bits,” said Penadés about Co-Scientist. “The system was able to put everything together.”
Sure. LLM-based madlibs can work as a suggestion tool. But the headline claim is not so convincing on AI scientific creativity.
