Executive Summary

Scientists who used AI in their work published far more papers than before, and their citations rose sharply too. Yet over the same period, the range of topics science covers as a whole narrowed. While individual productivity metrics climbed, the diversity of collective knowledge shrank. This piece does not set out to prove that phenomenon again; it follows the debate that came after it.

Pebblous already covered this data in detail in our 41-million-paper analysis report. Here we turn to the follow-up commentaries that appeared in and around the pages of Nature in the month after that report. Three commentaries reach the same conclusion in different languages: the key to stopping the narrowing lies not in better models, but in what researchers, reviewers, and funders choose to reward.

For anyone who works with data, this debate is not someone else's problem. Just as good data that loses its diversity narrows the model trained on it, science that leans on AI drifts toward the same place. The question of what quietly disappears when you optimize for something has widened from data quality to science as a whole.

The Figures at a Glance

Source: Hao, Xu, Li & Evans (2026), Nature 649 · Full analysis in the Pebblous report

×3.02

Individual output

Rise in publications by researchers using AI

×4.84

Citations

Individual influence metrics rose in step

−4.63%

Topic diversity

The breadth of topics science covers shrank

−22%

Follow-on collaboration

Decline in engagement following a study

1

The Question Nature Asked Itself a Month Later

Usually, once a single paper has proven a phenomenon with data, that is where the story ends. This time was different. Barely a month after James Evans's team published its analysis of 41 million papers — showing that individuals grew stronger while science as a whole narrowed — follow-up commentaries began appearing in the pages of the same journal. If the paper proved the problem, these commentaries move on to the next question: what has to change to stop the narrowing.

The piece that put that question most directly is a Nature World View commentary by Xizhe Zhang of Nanjing Medical University. The headline is the frame: will AI spark a scientific renaissance, or a diffuse monoculture? Zhang insists that what decides which way it goes is not the performance of the model. What matters is whether institutions move researchers, reviewers, and funders to reward originality instead of speed. AI is no longer a peripheral tool but part of the research infrastructure, with a small number of AI-equipped researchers now doing the literature review, experimental design, and model building that large collaborative teams once shared among many hands.

What decides the fork AI becomes research infrastructure Reward originality & long-haul work Reward speed & productivity only Renaissance Topic diversity expands Diffuse monoculture Topics converge to one place The reward criteria decide the direction
▲ What gets rewarded decides the fork between renaissance and monoculture | Original diagram by Pebblous (reinterpreting Zhang, 2026)

A shift in framing: if the May report answered "why is science narrowing," the question Nature went on to ask itself is "how do we stop the narrowing." And the center of gravity in the answer has moved away from technology and toward institutions.

2

A Warning That Guard Rails Are Needed

Slightly ahead of Zhang's commentary, anthropologist Lisa Messeri and cognitive scientist M. J. Crockett published a more blunt comment in Nature. The title is itself a warning: the uncritical adoption of AI in science is alarming, and guard rails are urgently needed. The two ground their case not in theory but in two problems already observed. Papers that used AI tended to cluster around narrower, more conventional research questions, and some were judged to be of lower scholarly quality than papers that did not use AI.

A guardrail and 'Caution: Blind Curve Ahead, Drive Slow' warning sign on a mountain road
▲ A warning sign and guardrail on a blind mountain curve — a metaphor for building the safeguard before it is too late | Source: Wikimedia Commons (schwiki, CC BY-SA 4.0)

On the same day, Nature ran a companion editorial. Its title is a declaration that good science is impossible without humans. However well AI scans the literature, drafts text, and organizes data, the argument goes, a person has to occupy the seat that decides which questions are worth asking. What is telling is that this warning is a separate, new piece, not the 2024 paper by Messeri and Crockett that the May report cited. Where the 2024 piece was an epistemological diagnosis — that AI breeds an illusion of understanding — the 2026 comment is closer to an empirical observation: that the diagnosis has already become reality.

The core point: the narrowing is not a hypothesis about the future but a phenomenon already observed on the page. That is why the language of the debate has shifted from "there could be a risk like this" to "let's build guard rails now."

3

The Bottleneck Is the Reward System, Not the Model

The third source is a piece in the journal Communications Psychology, which explains as a mechanism why the narrowing rolls forward on its own. The generative-AI research boom itself forms a feedback loop that drives convergence in topics and methods. It is a self-referential observation: even the way we study AI converges toward looking alike. This is where the metaphor comes from — that choices which are individually rational pile up to make the whole system resemble a field planted with a single crop.

The feedback loop turns on three points. First, evaluation and funding reward speed and productivity. So individuals rationally focus on the questions AI solves well. As those choices accumulate, the topics and methods of science as a whole converge on one place, and the narrowed terrain then pushes the next person's choice in the same direction. Unless the reward criteria are touched, the loop turns by itself. The diagram below lays out those three points on a single page.

A feedback loop built from individually rational choices Reward speed & productivity Focus on solvable questions Topics & methods converge = diversity falls Change nothing in the reward criteria and the loop keeps turning
▲ Reward speed and individuals crowd toward solvable questions; those choices pile up and the group's topic diversity falls | Original diagram by Pebblous

This diagnosis overlaps with the voices of the original researchers. Evans sums up the root of the problem as a conflict between individual incentives and science as a whole. For the individual, producing more, faster, with AI is an advantage, but the sum of those advantages returns as a loss to science overall. Luís Amaral of Northwestern is more blunt. We are digging the same hole ever deeper, he says, capturing the danger of convergence.

So the prescriptions the three sources offer also gather at one point: shift the evaluation criteria toward rewarding originality, theoretical depth, and the long-haul contribution of staying with a hard problem for a long time. The danger does not lie in computers thinking uniformly. It lies in people beginning to ask only the questions computers solve well. What you reward, in the end, decides what you ask.

What the evidence says: building a better model does not resolve the narrowing. Three commentaries point to the same bottleneck by different routes. It is an institutional problem — a matter of what we reward.

4

Data Can Show What Institutions Cannot See

Here we need to be honest. Changing the criteria for evaluation and funding is not a problem that data infrastructure alone can solve. It belongs to a higher layer — the institutions and incentives of academia — and it is not a domain the tools Pebblous builds can touch directly. It is more honest to begin the story by admitting that.

Still, to change the institutions, you first have to be able to see what is narrowing. Measuring which fields are quietly emptying out, which data sits untouched by anyone, and where diversity begins to collapse is, in the end, a data problem. This is why Pebblous has been building infrastructure to diagnose and restore the diversity of data. The attempt to catch the signs of a model narrowing at the data layer first uses the same grammar as reading the signs of science narrowing.

So this debate leaves anyone who works with data with one check question. What metric is my pipeline optimizing for, and what is quietly disappearing behind that optimization? Speed and productivity are easy to measure. Disappearing diversity is hard to measure. Making the hard-to-measure side measurable is how you give the "let's stop the narrowing" conversation something to stand on.

Closing: this piece is the next chapter of the problem the May report proved with data. Proving a problem and stopping the narrowing demand different abilities. The latter is the work of institutions that decide anew what to reward — but the evidence that decision needs comes from data that measures diversity.

R

References

Academic & Commentary

Press & Related