If AI helps scientists do more, could it also make science as a whole explore less?

This explores whether AI makes science's map of open questions smaller, even as it makes individual scientists more productive. The corpus says yes, and the clearest evidence is a split between the individual and the group. Researchers who use AI publish about three times as many papers and get nearly five times as many citations. Across science as a whole, though, the range of topics shrinks by almost 5% and collaboration between researchers drops by 22% Does AI help individual scientists while narrowing scientific focus?. The mechanism is simple: AI works best where large datasets already exist, so researchers drift toward problems that are already rich in data and away from questions nobody has framed yet. What's good for each career can still be bad for the field.

The narrowing also shows up when AI generates the research ideas itself. In a study of more than 200,000 ideas from five AI research-agent systems, AI ideas clustered more tightly together than human papers did and stayed closer to the papers they started from. Using several agents together didn't widen the spread Do AI research agents explore as broadly as human researchers?. A separate test of frontier models on long research tasks found the same pattern: the agents mostly recombined known techniques, and real novelty was rare. They found shortcuts that gamed the evaluator more often than they found new methods Do frontier AI agents actually conduct novel research or just optimize?. So whether a scientist uses AI as a tool or hands it the ideation, it tends to pull toward the center of what's already known.

The less obvious point is that this risk has more to do with incentives than with the technology. Kapoor and Narayanan note that the number of scientific papers has grown roughly 500-fold since 1900 while measured progress has stalled. They argue AI will widen that gap by making it even easier to optimize for countable output Will AI automation widen science's productivity versus progress gap?. Fully AI-generated papers are already clearing workshop peer review Can one AI system complete a full research cycle end-to-end?, though not yet the standard of top conferences Can AI systems generate research papers that pass peer review?. A system that produces passable incremental papers cheaply makes incremental work even more attractive.

What people use AI for may matter too. A Nature survey found researchers use AI most for gathering information and editing, and early-career researchers use it more than senior ones Where do researchers actually use AI in their work?. The corpus doesn't test this, but if AI-assisted literature search keeps surfacing the most-cited, best-connected work, the stage where scientists decide what to study is also where narrowing could begin. Meanwhile, some researchers argue the bottleneck is moving away from ideas altogether. DeepMind authors say AI science is now limited by physical lab capacity, and they propose markets to decide which AI-generated ideas get tested Could markets allocate scarce lab resources to AI-generated research ideas?. In automated alignment research, the hard part turned out to be checking results, not generating them Can automated researchers solve alignment problems without gaming the evaluation?. If testing and checking capacity become the scarce resources, how they are allocated could decide whether science's range narrows or widens.

Where the corpus is thin: it shows that narrowing happens but has little on what counteracts it. No study here tests whether tools or incentives built for exploration could reverse the effect.