In a Comment published today in Nature Astronomy, I ask whether one of the best uses of AI in science may be to help us publish fewer, better papers.

The title of the piece, “AI can help scientists publish less”, is deliberately provocative. Scientific life is still organized around papers: how many we publish, where we publish them, how often they are cited, how quickly they accumulate into a visible record of activity. But the deeper question raised by AI is not simply whether scientists can now produce more papers. They can. The question is whether producing more papers is what science most needs.

AI is making it cheaper to produce scientific papers. It can help draft text, write code, explore calculations, map literature, polish manuscripts and assemble arguments. Much of this can be useful. Used well, these tools can remove part of the routine labour that consumes scientific time.

But the other side of science has not become cheaper. Reading a paper carefully still takes time. Refereeing it responsibly still takes judgment. Understanding whether a result really adds something to what is already known remains a slow, human, collective task.

This asymmetry may be one of the central challenges AI poses to science.

A paper does not need to be wrong to slow science down. It can be correct, polished and publishable, and still demand more attention from editors, referees, readers and colleagues than the understanding it gives back. In the Comment, I describe this danger as “negative epistemic value”.

The analogy is that of so-called negative-calorie food: food said to require more energy to digest than it provides. Whether or not the nutritional claim is literally true, the metaphor captures something important about scientific publishing. Some papers may cost the community more, in time and attention, than they return in knowledge.

This is not an argument against incremental work. Science has always advanced through correction, refinement, replication, better measurements and careful tests. A paper can be incremental and important. A null result can close a path that needed to be closed. A reanalysis can expose a hidden assumption. A careful technical note can save others years of confusion.

The problem begins when publication becomes the default form into which every contribution must be forced, and when the cost of producing papers falls much faster than the cost of judging and absorbing them.

AI could make this imbalance worse. It could flood an already strained system with more papers: more papers, more claims, more things to review, cite, classify and evaluate.

But this is not the only possible future.

We should aim for more than defending a publication system already under strain from a new technology. We can be bolder, and treat AI as a historic opportunity to correct distortions in the publication system, and to make science, and the lives of scientists, better.

In the Comment, I describe three ways in which AI could help.

First, AI can make more forms of scientific contribution visible: code, data, benchmarks, reproducibility packages, public notebooks, living syntheses. These are already central to how science works, but they often become visible to institutions only when wrapped inside a paper. AI could help make their value legible directly.

Second, AI can reduce some of the routine work that consumes researchers’ time. Literature mapping, documentation, code scaffolding, reproducibility checks, first-pass comparisons across neighbouring fields: none of this replaces scientific judgment, but it can take weight away from the parts of scientific life that have become unnecessarily burdensome.

Third, AI can help editors, referees, panels and funders focus their limited attention where human judgment matters most. It can support triage, literature comparison, consistency checks and the detection of obvious pitfalls. It cannot decide what is important. That remains our responsibility. But it can improve the conditions under which we exercise that responsibility.

The point is not to ask how AI can help us publish more.

The more interesting question is whether AI can help us publish less, and do better science.

That would mean recognizing more kinds of contribution, rewarding responsible publication, and treating fewer, better papers not as a sign of inactivity, but as a mark of maturity. It would mean giving scientists more time to think, more freedom to change direction, and more room for work whose value cannot always be measured by immediate output.

AI will not decide what matters. That remains a human task, and a collective one.

But it may help us recover some of the time, attention and judgment that science needs most.

Comment in Nature Astronomy:
https://www.nature.com/articles/s41550-026-02900-y

Free read-only version:
https://rdcu.be/fnxGo