Amid the noise about ever-larger models, it is worth pausing on a quieter kind of AI success. MIT's Regina Barzilay and James Collins have won the 2026 Pritzker Prize for AI in Science Research Excellence for using machine learning to find new antibiotics, a result that matters because antibiotic resistance kills and the pipeline of new drugs had nearly dried up. Their work shows AI earning its place in a field where the stakes are measured in lives rather than benchmark points.
The prize, awarded jointly by the University of Chicago and Caltech and established in 2024, recognises two landmark findings. In 2020 the pair, working with then-MIT researcher Jonathan Stokes, identified halicin, a compound picked out by a model trained on about 2,500 molecules and named after the computer HAL 9000. Halicin is a broad-spectrum antibiotic able to kill resistant bacteria including Clostridium difficile, Acinetobacter baumannii and the tuberculosis bacterium, and the model compressed a search that once took years into weeks. In 2023 the same group found abaucin, a narrow-spectrum antibiotic aimed specifically at A. baumannii, a resistant bug common on hospital surfaces.
The detail worth drawing out is that abaucin's narrowness is the point. A drug that targets a single species does less collateral damage to the beneficial bacteria in the gut and gives resistance fewer openings to develop, which is a more thoughtful use of AI than simply screening for the strongest killer. The method, careful modelling paired with real laboratory validation, is where the value sits, not in the size of the model.
The winners themselves are measured about it. Collins frames generative AI as a shift from using AI to discover molecules towards using it to design them, while Barzilay cautions that the field is still early, that AI can automate parts of the drug pipeline rather than transform the whole of it, and that realistic models of human biology remain hard to build. That candour is the right note.
The honest gaps are in the announcement itself. It names no prize amount, and more importantly it does not say how close either halicin or abaucin is to reaching patients, which is the question that ultimately decides whether this line of work saves anyone. Many promising compounds fail in the years of trials that follow discovery. This is a deserved award and a reminder that some of the strongest AI-for-science results came from modest models and patient lab work, but the drugs themselves still have to clear the slow part.