An AI matched a decade of research in two days

Researchers at Imperial College London had an unpublished hypothesis about how mobile genetic islands in bacterial cells hijack a virus's tail to jump between species, confirmed only in their own unreleased lab data and never shown outside the lab.

The team gave Google's experimental "AI co-scientist" system a short prompt describing the open problem, withholding their unpublished results. Within 48 hours, the AI returned five ranked hypotheses. The top one matched the mechanism the lab had spent roughly ten years confirming experimentally in bacterial cells.

The researchers emailed Google to ask whether the system had somehow accessed their unpublished work; the company said it had not. The underlying mechanism, that these islands build a tail by stealing parts from an infecting virus, was later confirmed again in the lab and published alongside the AI comparison. One matched case does not prove the tool works reliably on other open problems.

Ten years of lab work, reached in a weekend.