Google AI cracks a 10-year superbug mystery in 48 hours

Imagine spending ten years in a lab trying to solve a deadly medical mystery, only for an AI to spit out your exact conclusion in 48 hours.

That was the reality for Professor José R Penadés and his research team at Imperial College London. For years, they investigated how superbugs develop immunity to antibiotics. Penadés decided to test Google’s experimental tool, Co-Scientist, feeding it a short prompt outlining the problem. Two days later, the AI delivered the same hypothesis the team had spent years proving.

Penadés was so stunned he had to step away while shopping to process what happened. He even fired off an email asking Google if they had secretly accessed his computer. Google confirmed they hadn’t.

The tail that opens doors

The Imperial College team was investigating how antibiotic-resistant bacteria hop between different host species.

Their working hypothesis was that superbugs hijack tail fragments from viruses, effectively using them as biological keys to unlock and infiltrate new host species. This was a concept the team had guarded closely. Yet Co-Scientist not only ranked this viral-tail mechanism as its top hypothesis, it generated four other plausible ideas—one of which was novel enough that the team is now actively testing it in the lab.

If the story ended there, it would be the ultimate victory lap for AI in frontier science. But as Pivot to AI and New Scientist soon pointed out, the headline left out a massive catch.

A brilliant summarizer, not an Einstein

It turns out Co-Scientist was not generating novel science out of thin air. The model had been fed a 2023 paper published by Penadés' own team that already outlined a version of this very hypothesis.

Penadés later clarified the dynamic: everything was already published in different bits, and the AI was able to assemble those scattered pieces together. That is a massive difference from autonomous discovery. It fits an established pattern with Google's scientific claims, such as Co-Scientist proposing liver fibrosis treatments that were already under study, or DeepMind claiming 43 brand-new synthesized materials that independent chemists later proved were neither new nor easily replicable.

Why it matters

Google certainly stretched the hype on this one, but we still think Co-Scientist points to a real shift in lab workflows.

For working scientists, having a tool that can connect dots across thousands of disparate papers in 48 hours is a genuine superpower. The modern literature firehose makes cross-disciplinary pattern matching nearly impossible for human brains alone. Even if Co-Scientist is just an aggressive hypothesis synthesizer rather than an original thinker, trimming years off early-stage research is worth the hype.

The real danger lies in treating these models as digital colleagues rather than automated lit-review engines. LLMs remix what exists; they do not understand underlying biophysics. If teams begin pursuing AI-generated hypotheses without checking whether the model simply regurgitated obscure prior art, they risk wasting months chasing circular logic.

Co-Scientist will not replace microbiologists anytime soon. But researchers who learn how to prompt these multi-agent engines to connect the dots will run circles around those who still do literature reviews by hand.