Is AI Making Science Less Innovative?
Artificial intelligence is meant to quicken the pace of discovery and push science into unexplored territory. A newstudyfrom the University of Chicago suggests it does something subtler—and more troubling. AI. . .
Artificial intelligence is meant to quicken the pace of discovery and push science into unexplored territory. A newstudyfrom the University of Chicago suggests it does something subtler—and more troubling. AI supercharges individual scientists, but nudges science as a whole towards a narrower range of questions. The researchers examined 41. 3m academic papers and found that scientists who use AI are markedly more productive and influential. They publish just over three times as many papers as their peers, attract nearly five times as many citations, and reach positions of intellectual leadership around 1. 4 years earlier. By almost any personal metric, AI is a boon. Worrying trendCollectively, however, the picture is less flattering. As AI use spreads, the overall range of scientific topics contracts. The volume of topics studied falls by 4. 6%, while interaction between researchers declines by 22%. Science becomes busier, but also more crowded—and more inward-looking. The reason is prosaic. AI works best where data are plentifu,l and progress can be benchmarked with reassuring clarity. Unsurprisingly, scientists gravitate towards such fields. Instead of widening the frontier, AI concentrates attention on data-rich domains, leaving thinner but potentially fertile areas neglected. The authors describe the result as the rise of “lonely crowds”: popular topics that attract intense interest but less genuine engagement. Papers cite the same work, pursue similar questions, and arrive at overlapping conclusions. Knowledge deepens in familiar grooves, even as its overall extent shrinks. A clear riskThis efficiency carries a risk. By funnelling effort towards a narrow set of methods and problems, AI encourages what the authors call methodological monocultures. Science, they warn, may converge too quickly on established paradigms, mistaking optimisation for discovery. Avoiding this fate may require intervention. The study points to policies that reward data collection in neglected areas and encourage AI systems designed for exploration rather than mere optimisation. The irony is that the same models prized for producing highly probable answers may also be well suited to spotting the unexpected—if they are pointed in that direction. To sustain progress, the authors argue, AI must do more than compress and analyse existing data. It must help scientists generate new kinds of evidence, opening domains that were previously inaccessible. Without that shift, AI may make science faster, but not necessarily bolder. Post navigation