AI Is Making Scientists More Productive. It May Be Destroying Science.

March 5, 2026

A study published last month in Nature offers one of the most striking paradoxes yet to emerge from the AI era: the same tools that are making individual scientists dramatically more productive may be quietly strangling the breadth of science itself.

The analysis covers 41 million papers published between 1980 and 2025 across biology, medicine, chemistry, physics, materials science, and geology. The findings are hard to argue with. Scientists who use AI publish three times more papers and receive nearly five times more citations over their careers. Junior researchers who adopt AI are less likely to drop out of academia and reach leadership positions almost a year and a half earlier than their peers.

And yet.

AI-driven research covers 4.6% less scientific territory than conventional studies. It generates 22% less engagement between papers — fewer studies building on one another, fewer new lines of inquiry sparked. Instead of a rich, interconnected web of ideas, AI-heavy fields are producing a hub-and-spoke structure where a handful of superstar papers collect the vast majority of attention.

"If we all climb the same mountains," says Tsinghua co-author Fengli Xu, "then there are a lot of fields we are not exploring."

Lisa Messeri of Yale puts it even more bluntly: "There needs to be some deep reckoning with what we do with a tool that benefits individuals but destroys science."

The Feedback Loop They Found

The study authors identify a self-reinforcing cycle. Popular problems attract funding, which motivates the creation of large datasets, which makes AI tools useful and appealing, which attracts more scientists to those same popular problems. Round and round.

"We're like pack animals," says co-author James Evans of the University of Chicago.

It's a clean explanation for how clustering happens at the sociological level. But it doesn't fully explain why the AI tools themselves gravitate toward the popular — why, when a researcher sits down and asks an LLM to help survey a field or generate ideas, the results tend to cluster around the same well-worn territory.

That part is worth examining.

What the Study Doesn't Say

Here is the mechanism the study doesn't discuss, and it matters.

First: all major LLMs are trained on largely the same data. The corpus of text that these models learned from — Common Crawl, academic repositories, the open web — reflects what was already prominent and well-documented at the time of training. Fields with rich online literature get rich representation. Fields that are niche, poorly digitized, or primarily published in non-English languages do not.

Second: training data has a cutoff. The most recent research — the emerging, the tentative, the not-yet-established — is largely invisible to the model. An LLM asked to survey the frontier of a field will tend to describe the frontier as it existed a year or two ago, weighted toward whatever was already generating buzz.

Third: the popular finds the popular. When a researcher uses an LLM or AI search tool to explore a literature, the model's outputs are shaped by what appears most frequently and most prominently in its training. Obscure but important work — the kind that might open a genuinely new mountain — is harder to surface. Paywalls compound this: many AI research tools cannot access subscription journals, meaning the most recent and specialized work is often simply not returned.

The feedback loop the study identifies operates at the level of scientific incentives. But underneath it, the tools themselves are built in a way that makes the popular more popular and the obscure more obscure.

A Scientific Monoculture

A companion piece in Nature Communications Psychology uses a phrase worth sitting with: scientific monoculture.

Agriculture learned this lesson the hard way. When you optimize for yield in a single crop, you get extraordinary productivity — until a blight hits, or conditions change, and there's nothing else in the ground. Diversity isn't inefficiency. It's resilience.

Science works the same way. The breakthroughs that reshape fields rarely come from the crowded center. They come from the edges, the underfunded corners, the researchers willing to climb an unfashionable mountain.

AI, as currently designed and deployed, is making that harder. Not out of malice — out of architecture.

The question the scientific community now has to answer is whether the productivity gains are worth the cost. And whether the tools can be redesigned — better datasets in underrepresented fields, broader training corpora, genuine access to specialized literature — before the monoculture takes hold.

My Adventures With Claude, a book about AI for general readers, is available on Amazon.