AI In Science Connects the Dots, But Only In Fields That Are Fragmented

The New Intelligence · Explore this series ›

Tormod Guldvog

Editor & Duowriter

July 3, 2026

Key Takeaways

• AI lifts novelty in rough fields like astronomy, reverses to automation in codified ones like cell biology.
• AI's citation premium reflects social rewards, not discovery, independent of knowledge structure.
• AI papers are 30–40% more likely than field-mates to introduce conceptual or linguistic novelty.

✓ Fact-checked · See verification details

AI assistant: Research, Drafting, QA

On 9 October 2024 the Royal Swedish Academy of Sciences gave the Nobel Prize in Chemistry to three men, two of whom had built a machine. Demis Hassabis and John Jumper of Google DeepMind had taught a neural network, AlphaFold, to predict the folded shapes of proteins. Marking the field's grandest honour, DeepMind distilled the optimist's creed into a single sentence: AI "will make science faster and ultimately help to understand disease and develop therapeutics."

DeepMind's policy team had already gone further, announcing in a manifesto "a new golden age of discovery." A machine had, it seemed, originated something, and the establishment had blessed it.

Not everyone was nodding. Seven months earlier, Lisa Messeri of Yale and M. J. Crockett of Princeton had set down in Nature a sentence that has hung over the field since: the proliferation of AI tools "risks introducing a phase of scientific enquiry in which we produce more but understand less." Their deeper worry was monoculture: "scientific monocultures," they wrote, "in which some types of methods, questions and viewpoints come to dominate," the peripheral vision of a whole field quietly narrowing.

For two years the argument between the two camps ran almost entirely on anecdote. Each side pointed at the same few triumphs and the same few terrors. Almost nobody had measured what AI does to scientific creativity across the whole of science.

Then, in April 2026, someone counted. Stefano Bianchini, an economist at BETA in Strasbourg, with Valentina Di Girolamo, Julien Ravet and David Arranz of the European Commission's research directorate, took more than eighty million papers from the OpenAlex database, spanning 2005 to 2023 and 172 fields. They asked a narrower, answerable question than "is AI good for science." When a paper uses AI, is it more novel (does it coin new words, new phrases, new combinations of ideas) and more impactful (does it get cited)?

Their headline answer was probably what the AI optimists wanted to hear. AI papers are roughly 30 to 40 percent more likely than their field-mates to introduce linguistic or conceptual novelty, and they crowd the ranks of the most-cited. "The use of AI," the authors write, "is associated with more novel and highly cited research." Then the sentence turns, in the dry register of a results section: "However, these benefits are far from uniform."

Whether AI lifts novelty, it turns out, depends almost entirely on where it is used. Picture a field's knowledge as a landscape. Where the terrain is rugged, fragmented into disconnected peaks, combinatorially complex and hard to traverse, AI behaves like a pathfinder and lifts novelty sharply. Where the terrain is a smooth plain, well-trodden and well-connected, AI merely lets researchers walk the familiar paths faster, and novelty barely moves at all.

The effect strengthens with two things at once: a field's roughness, and its "AI exposure," how deeply the tools have penetrated it. Astronomy, bioinformatics, climatology, geodesy, medical physics and radiology sit at the strong end. In fields with standardised, codified protocols, cell biology, molecular biology, traditional medicine, the association runs the other way: AI shows its strongest negative novelty, working as technical automation, classifying and predicting without ever troubling the conceptual frontier.

The paper separates two things the debate routinely fuses: novelty and impact. Novelty, the saying of new things, tracks the structure of knowledge exactly as you would expect. Citations are a different story altogether: the terrain that amplifies novelty leaves the citation premium untouched, knowledge structure explains barely any of it, and the average effect on citation counts is, in the authors' own word, "modest."

The authors are careful about what they will and will not say: impact "may rather be shaped by social and institutional mechanisms (e.g., visibility, collaboration networks, reputational effects), whose dynamics operate independently of the cognitive structures captured by knowledge-space roughness."

My take on this is that the citations AI brings may be something older at work, the social machinery of science rewarding whoever picks up the shiny new instrument first. What the paper claims is narrower than it sounds: the reward and the discovery have come apart. It does not claim that AI has stopped aiding discovery.

...the reward and the discovery have come apart.

AI does not only impact analysis and research. In a 2026 Nature paper, James Evans of Chicago, Fengli Xu of Tsinghua and their co-authors tracked what happens to the scientists themselves. Those who adopt AI, who differ from their peers in much else besides, publish 3.02 times more, are cited 4.84 times more, and become project leaders 1.37 years earlier than those who do not. The same adoption shrinks the collective volume of topics science studies by 4.63 percent and cuts scientist-to-scientist engagement by 22 percent, because AI work "moves collectively toward areas richest in data."

Their verdict echoes Messeri and Crockett: "AI tools appear to automate established fields rather than explore new ones, highlighting a tension between personal advancement and collective scientific progress."

Jian Gao and Dashun Wang of Northwestern, measuring the same citation premium in 2024, added a further discomfort: its benefits skew away from fields with more women and Black scientists. It echoes the tension Evans had named, between personal advancement and collective progress.

The optimists are not left empty-handed, and honesty requires saying so. In a blind study run by Chenglei Si, Diyi Yang and Tatsunori Hashimoto at Stanford in 2024, some seventy-nine expert reviewers judged research ideas generated by a large language model to be more novel than those of human experts (p < 0.05). The same machine ideas scored slightly weaker on feasibility, a caveat the authors themselves press.

And AlphaFold, the prize-winner, turns out on closer inspection to have redirected science rather than accelerated it. Ryan Hill and Carolyn Stein, in a 2026 working paper (not yet peer-reviewed, so hold it lightly), find that after AlphaFold2's 2021 release the rate of experimental structure determination barely changed. Basic research on previously unstructured proteins rose by 15 to 40 percent, with no measurable shift yet into early-stage drug development. AlphaFold simply pointed scientists toward the proteins they had long ignored.

There is a nice, ironic twist in our tale. Four years earlier, Bianchini himself, with Müller and Pelletier, had argued that AI was diffusing as a "general method of invention" that tracked well-defined research trajectories and was associated with less novelty. The 2026 paper reverses that picture, conditionally; its lead author was honest enough to revise himself in public.

The reversal is not really a contradiction. The data set grew, the transformer and the large language model arrived in the interval, and the measurement followed them. The field's verdict on AI is young enough that even its own measurers are still revising it.

The sharpest cautionary tale arrived in late 2024, when a graduate student named Aidan Toner-Rodgers circulated a working paper claiming that an AI tool, deployed at a 1,018-scientist materials lab, had lifted discovery by 44 percent, patents by 39 percent, and product innovation by 17 percent. It was the single cleanest empirical proof anyone had produced that AI supercharges discovery, and it was embraced by Daron Acemoglu and David Autor, two of the most decorated economists alive. Then it dissolved.

In May 2025 MIT announced it had "no confidence in the provenance, reliability or validity of the data", the paper was withdrawn from arXiv, and its author left the institution. "There is no world where this makes any sense," Autor told the Wall Street Journal. MIT itself noted that "even in its non-published form, the paper is having an impact on discussions and projections about the effects of AI on science." The 44 percent did not survive scrutiny. What lingers is how badly the discipline's most distinguished readers seem to have wanted it to.

If AI's value depends on terrain and on depth of adoption, the binding constraint may lie in diffusion: getting AI well into the hard, fragmented fields where it earns its novelty and where adoption still lags. That, as it happens, is the bet Europe has begun to place.

More on AI in Science

AI Solves Erdős Math Problem: What's Next for AI in Mathematics?

An AI solved an 80-year-old Erdős math problem by walking a path mathematicians had collectively avoided.

→

The European Commission has read the evidence this way, launching a Strategy for AI in Science on 8 October 2025 and, a month later, the RAISE pilot in Copenhagen, on the premise that "AI has a major impact on science and the scientific process," reaching, in its own list, "from literature reviews to laboratory automation, from life sciences to humanities." The argument now has money attached.

The deeper question the data leaves open is the one science cannot outsource. The Bianchini paper begins, as such papers rarely do, with two epigraphs. In 1843 Ada Lovelace wrote that the Analytical Engine "has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform."

In 1950 Alan Turing replied, across the century, that machines "can take us by surprise." Eighty million papers later, the honest answer is that both were right, and which one holds today depends entirely on the terrain.

Sources

• Primary source: Bianchini, Stefano, Valentina Di Girolamo, Julien Ravet & David Arranz. "AI in science: When and where it makes a difference." Research Policy 55 (2026) 105478.
• Context sources:
  • Hao, Xu, Li & Evans. "AI tools expand scientists' impact but contract science's reach." Nature (2026); preprint arXiv:2412.07727.
  • Messeri, Lisa & M. J. Crockett. "Artificial intelligence and illusions of understanding in scientific research." Nature 627 (2024).
  • Gao, Jian & Dashun Wang. "Quantifying the use and potential benefits of artificial intelligence in scientific research." Nature Human Behaviour 8 (2024).
  • Si, Chenglei, Diyi Yang & Tatsunori Hashimoto. "Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study." arXiv:2409.04109 (2024).
  • Hill, Ryan & Carolyn Stein. "How Artificial Intelligence Shapes Science: Evidence from AlphaFold." Working paper (2026).
  • Bianchini, Stefano, Moritz Müller & Pierre Pelletier. "Artificial intelligence in science: An emerging general method of invention." Research Policy 51(10) (2022).
  • Google DeepMind. "A new golden age of discovery" (policy essay, 2024) and "Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry."
  • "The Nobel Prize in Chemistry 2024." Royal Swedish Academy of Sciences, October 9, 2024.
  • Toner-Rodgers, Aidan. "Artificial Intelligence, Scientific Discovery, and Product Innovation." arXiv:2412.17866 (2024, withdrawn); MIT Department of Economics, "Assuring an accurate research record" (May 16, 2025).
  • European Commission. "European Strategy for AI in Science" (October 8, 2025) and RAISE pilot launch (Copenhagen, November 2025).

Tormod Guldvog

Editor & Duowriter

Tormod Guldvog is the editor of Science Reader, covering the space where human and artificial intelligence meet. With a background in technology and science communication, he writes about how AI is changing scientific discovery — and what that means for us. More about Science Reader

Cite this article

Tormod Guldvog. (2026, July 3). AI In Science Connects the Dots, But Only In Fields That Are Fragmented. Science Reader. https://sciencereader.com/ai-in-science-connects-the-dots/