Is AI research narrowing around its biggest successes?
A preprint covering 7.18 million computer-science papers finds attention and citations became more concentrated after 2017. The pattern is substantial, but it does not prove that AI innovation is already stalling.
The short version
- What happened
- Researchers analysed 7,180,635 English-language computer-science papers from 1990–2025, 73,089 OpenReview reviews and the publication histories of 3,707,765 authors with at least two papers.
- Who it affects
- Early- and mid-career researchers may face stronger incentives to retool toward fashionable AI topics. Funders and conference organisers can respond by measuring whether less popular approaches still receive expert review, resources and a credible route to publication.
- What remains unproven
- The preprint has not been peer reviewed and the 1990–2025 corpus includes only English-language computer-science papers with abstracts.
Keep in mind
The preprint has not been peer reviewed and the 1990–2025 corpus includes only English-language computer-science papers with abstracts.
TF-IDF terms and extracted nouns are proxies for attention and ideas; they do not directly measure research quality, social value or breakthrough probability.
Observed changes around 2017 do not establish that the Transformer caused concentration or falling reviewer confidence.
The reciprocal-review result is a stylised simulation, while the authors say the post-2022 LLM period is too short for a definitive exhaustion claim.
One author is affiliated with Google; the remaining authors declare no competing interests. No external grant funding is disclosed in the paper's acknowledgements.
Source analysis · 1 linked source. Primary analysed source date: 19 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
More researchers, but attention moved in fewer directions
The study used a December 2025 OpenAlex snapshot to examine English-language computer-science papers with abstracts. It extracted the ten most distinctive terms from each abstract and measured how unevenly annual attention was distributed. The attention Gini—where a higher value means more concentration—rose from 0.26 in 2017 to 0.33 in 2025. AI's share of the corpus also increased from about 40% to more than 55%.
Author movements tell a similar story. Of 918,862 observed switches between broad subfields, 53.81% ended in AI. The estimated probability of moving into AI peaked among mid-career, moderately established researchers. For universities and funders, the practical question is whether hiring, reviewing and grants still leave room for work outside the dominant paradigm—not whether researchers should stop building on successful methods.[1]
Peer review showed strain, not a measured collapse
Across 73,089 reviews of 26,731 submissions to 45 venues from 2017–2025, reviewer confidence fell from 3.9 to 2.9 after scores were normalised to a common five-point scale. The authors interpret that as consistent with difficulty matching papers to expertise as AI spreads. It could also reflect changes in venues, scoring cultures or submission mix; the design does not isolate one cause.
The paper's strongest language concerns idea 'exhaustion', measured through nouns extracted from abstracts and fitted to logistic curves. Those are useful bibliometric signals, not direct counts of discoveries. The authors explicitly say the post-2022 LLM window is too thin for a definitive claim. More years of data, peer review and alternative measures of novelty would materially strengthen—or weaken—the conclusion.[1]
The distinction that matters
From research records to a cautious conclusion
The study combines several signals, but none alone measures whether the next breakthrough has become less likely.
Papers
Track terms, citations and subfields
Attention became more concentrated
People
Follow 918,862 field changes
More than half ended in AI
Policy model
Simulate reciprocal reviewing
A possible mechanism, not causal proof
A simulated warning about who gets to judge
To explore reciprocal reviewing, the researchers simulated 100 scientists over 120 rounds, including five prolific authors, and ran 25 random seeds per parameter setting. Stronger links between submission volume and reviewing load concentrated accepted topics around the prolific reviewers' preferences.
That model is an illustrative mechanism, not evidence that current conference rules have caused real-world narrowing. It uses a two-dimensional topic space and hand-tuned parameters. The paper nevertheless offers a testable policy question: do reviewing rules, fees and reward systems diversify accepted work over time, or reinforce the subjects already attracting the most people and money?[1]