AI helps individual scientists but narrows what science explores

A new study led by the sociologist James Evans, reported by IEEE Spectrum, looked at how AI is changing science by measuring what scientists actually publish. The team analyzed 41.3 million academic papers from 1980 to 2025 across six disciplines and compared researchers who adopted AI tools against those who did not. For the individual, the payoff is large: AI adopters published about three times as many papers, received nearly five times as many citations, and reached leadership roles one to two years earlier.

The catch shows up at the level of the field. The same researchers tended to cluster around popular, data-rich problems, the questions where AI helps most. Their work covered a smaller intellectual footprint, and it generated weaker follow-on citation networks, the branching lines of later research that usually grow out of genuinely new directions. So the tools reward the scientist while quietly thinning out the variety of questions being asked.

Evans frames this as a conflict between individual incentives and science as a whole, and he draws a comparison to what happened with search engines, which made finding known work easy but pulled attention toward what was already popular. AI, on his reading, is a faster version of the same homogenization. The numbers give the argument weight that a purely theoretical warning would lack.

Why it matters

If you run a lab or fund research, this is a reason to watch for concentration: the people using AI most may be the most productive and the most likely to converge on the same crowded problems. The finding suggests deliberately protecting room for odd, less data-friendly questions, because the individual incentives now point away from them.