Analyses of hundreds of thousands of papers in the natural sciences reveal a paradox: scientists who use AI tools produce more research but on a more confined set of topics.