Did AI solve a superbug problem in two days?

I consult on lab operations and AI, and this story comes up more than any other when people tell me AI is about to replace bench science. It deserves a careful read, because what happened is genuinely impressive. It just isn't what the headline says.

The headline version

The BBC piece (Tom Gerken, 20 February 2025) ran under the headline "AI cracks superbug problem in two days that took scientists years." It describes José Penadés's team at Imperial College London, who had spent years working out how certain bacterial genetic elements spread between species. Penadés gave Google's tool a short prompt about the core problem, and within 48 hours its top answer matched his team's conclusion, which had not been published.

The same article contains the sentence that should have shaped the headline: the decade "also includes the time it took to prove the research," which was itself multiple years. The team's own framing, as the BBC reports it, is that having the hypothesis at the start would have saved years of work. That is a strong claim. It is a different claim from "solved."

What the papers say happened

There are two peer-reviewed accounts. The first is the Penadés group's own paper in Cell (Penadés et al., 2025; first posted to bioRxiv on 19 February 2025). Its abstract frames the test as posing "a question that took years to resolve experimentally but remained unpublished": how capsid-forming phage-inducible chromosomal islands (cf-PICIs), mobile genetic elements that can carry virulence and antibiotic-resistance genes, spread across bacterial species. The AI's top-ranked hypothesis, that cf-PICIs hijack the tails of many different phages to expand their host range, matched the mechanism the lab had already confirmed by experiment.

The second is Google's full system paper, "Accelerating scientific discovery with Co-Scientist," published in Nature on 19 May 2026 (Gottweis et al., 51 authors). It describes the same episode in careful language. The researchers "instructed the system to explore a topic their group had independently discovered, but not yet published." With "only minimal background information," the system proposed the tail-hijacking hypothesis as its top pick, generated "in just 2 days." The section heading is "Recapitulating a breakthrough in AMR."

One detail most coverage skipped: the Cell paper is not purely an outside test. Its 13 authors include the Imperial group and several of the Google researchers who built the system. That is normal for a collaboration, and nothing suggests it changed the result. It is worth knowing when the story is told as a lab independently checking Google's tool.

Hypothesis versus proof

In a lab, the step the AI performed is the cheap one. A hypothesis is a sentence. Turning that sentence into knowledge means designing experiments, building strains, running them, failing, rerunning, and ruling out the alternatives. The BBC's own reporting says years of the decade went to that.

What the test does show is real and worth stating precisely. A system given a short prompt and the open literature ranked the correct mechanism first, on a question whose answer was not public. That suggests the clues were sitting in published work and the system connected them faster than people did. The Cell abstract adds that the authors examined its five highest-ranked hypotheses and that some "opened new research avenues" in their labs. For a scientist choosing what to test next, that's valuable.

What it does not show is that the AI could have found the answer alone. It was asked by the people who already knew the answer, about the question they had chosen, and its output was checked against results that took years to produce. A ranked list is only useful if someone can tell which entry is right, and here the answer key already existed.

The cancer-drug results have the same shape

The Nature paper's other headline result, drug repurposing for acute myeloid leukaemia (AML), shows the same division of labor. The authors constrained Co-Scientist to a curated list of 2,300 approved drugs. Its 30 top-ranked candidates went to expert oncologists, who selected five for lab testing: binimetinib, pacritinib, cerivastatin, pravastatin and dimethyl fumarate. Three of the five inhibited AML cell viability in culture. Binimetinib was the standout, with half-maximal inhibitory concentrations as low as 2 nM in three of four AML cell lines and much weaker activity in a non-AML control line.

That is a reasonable hit rate for a screen. But note who did what: the AI proposed, human experts filtered 30 down to 5, and cell-line assays did the checking. The paper itself calls these experiments "a viability check" that is "not a replacement" for preclinical and clinical assessment, and it notes that binimetinib already showed limited efficacy as a single agent in an earlier phase II study in relapsed or refractory RAS-mutant AML. (Report 022 covers how AI-found drugs pass early safety trials and then stall in Phase 2.)

What the authors say about limits

The Nature paper is candid in its discussion. Co-Scientist relies on open-access literature, so it can miss paywalled work and almost never sees negative results. It can propagate erroneous or irreproducible findings from that literature. It inherits its underlying models' "potential for hallucinations." And in the authors' own words, "the validation of Co-Scientist's hypotheses, while successful, remains preliminary."

How to read the next "AI solved it" story

Hypothesis or proof? If the AI produced an idea and humans or instruments confirmed it, the AI generated a hypothesis. That can be valuable. It is not a solution.

Who knew the answer? A test where experts already hold the answer is a fair benchmark, but it tells you about recall and ranking, not about finding something nobody knows.

How many options, and who picked? "The AI found the drug" often means the AI ranked thousands and people chose a handful to test.

Compare like with like. Two days of computing against a decade of hypothesis plus experiments is not a speed comparison. The honest comparison is how long the humans took to arrive at the hypothesis, and the reporting doesn't give that number.

Disclosure, plainly: I use AI tools in my own work, including Claude (made by Anthropic, a Google competitor) for research and drafting for this publication. I am a customer of Anthropic's products and have no other relationship with the company, Google, Google DeepMind, Imperial College London or any of the authors; none of them asked for, saw or reviewed this report. Nothing here is sponsored and no link earns a commission; here's the full policy.