AI research keeps growing. But researchers may be spending more of that effort pursuing the same ideas. A study of 7.2 million computer-science papers suggests that major AI breakthroughs can draw attention toward winning approaches so powerfully that the diversity of ideas available for the next discovery starts to shrink.
That is a troubling possibility for a field increasingly organized around transformers and large language models. It is still a hypothesis about scientific research, not evidence that transformers have hit a technical ceiling or that AI models are improving more slowly.
How the researchers measured AI’s growing concentration
In an October 19, 2026 preprint, Beichen Lu, Honglin Bao, Qi Huang, Qingguo Meng and James A. Evans examine computer-science research published between 1990 and 2025. Their dataset includes 7,180,635 papers, 73,089 peer reviews and the publication histories of approximately 3.7 million researchers.
The researchers measure which subjects attract attention, how scientists move between research subfields and how the vocabulary of published work changes. For each abstract, they identify prominent keywords, then calculate how unevenly attention is distributed across them. They also use nouns in abstracts as markers for potential research ideas.
One measure of attention concentration, the Gini coefficient, rose from 0.26 to 0.33 between 2017 and 2025. The higher number indicates that attention was distributed less evenly among research keywords. During roughly the same period, AI’s share of computer-science papers increased from about 40% to more than 55%.
The migration data show a similar pull. Of 918,862 recorded moves between research subfields, 53.81% ended in AI. Those are observed changes in researchers’ main subjects, not evidence that 53.81% of all computer scientists abandoned their specialties.
Peer review offers another signal. Among reviews sampled from 45 venues, average reviewer confidence fell from 3.9 to 2.9 on a venue-normalized 0-to-5 scale between 2017 and 2025. The authors suggest that matching papers to suitable experts may have become more difficult as AI expanded. Falling confidence scores alone cannot identify the cause.
Each measurement captures a different part of the research system. Taken together, they suggest that computer science is becoming more concentrated around AI, even while the field produces more papers and attracts more people.
Why breakthroughs might shorten the life of research ideas
The paper describes a recurring cycle around AlexNet in 2012, the transformer architecture in 2017 and the rise of large language models. A breakthrough creates new possibilities and attracts researchers. Fresh terminology and projects proliferate. Over time, much of the community gathers around improving the successful approach, while other lines of inquiry receive less sustained attention.
To study that pattern, the authors treat each distinct, stable noun found in paper abstracts as a potential idea. A noun becomes established after appearing in at least 10 papers in each of two consecutive years. If it fails to meet that threshold for three consecutive years, the model treats the idea as inactive. It can later return.
That method allows the authors to count the appearance, disappearance and revival of research terminology across decades. Their fitted curves suggest that successive breakthroughs raise the estimated capacity for active ideas, but are also followed by sharper declines from that modeled potential. They report a comparable cycle in structural biology after cryo-electron microscopy and AlphaFold.
The proposed mechanism is plausible. Once a powerful architecture takes hold, researchers can publish improvements using familiar tools, shared benchmarks and established assumptions. An unfamiliar approach may initially perform worse on those tests, even if it has strengths that could become useful later. It needs time, funding and reviewers willing to evaluate a different hypothesis.
The authors also simulate a peer-review system in which prolific submitters take on more reviewing work. Under the model’s assumptions, that arrangement gives established, highly active researchers more influence over which topics receive approval. It does not prove that existing conference review policies created the historical concentration.
Why the study cannot prove AI is running out of discoveries
The paper’s most important limitation is its definition of an idea. A noun can help track changes in scientific language, but it is not the same thing as a scientific concept. Researchers can develop genuinely new methods using familiar words. A fashionable term can also spread widely without representing a meaningful discovery.
The authors test alternative ways of extracting nouns and noun phrases, which strengthens their claim that the statistical pattern is present in the text. It does not resolve the harder question of whether those vocabulary patterns accurately measure the quality or originality of science.
Timing also complicates the argument. Transformer research grew during a period of increasing computing capacity, investment and commercial demand for AI. The study cannot isolate the transformer architecture as the cause of the shift in research attention, or establish that concentrating attention will delay a future breakthrough.
The authors acknowledge that the period following the newest LLM advances is too short for a definitive conclusion. Their research-idea “carrying capacity” is a fitted curve, not a physical limit on what AI systems can discover. The dataset concentrates on English-language computer-science papers with abstracts, while much unpublished industrial research remains outside its reach.
Other work raises related questions using different evidence. A 2020 American Economic Review study found that research effort rose while research productivity declined in several settings. A 2023 Nature analysis reported that scientific papers and patents became less disruptive by citation-based measures. Neither result independently establishes the causal story proposed in this new preprint.
Could transformer dominance make the next AI breakthrough harder?
It could. When funding, talent and evaluation benchmarks favor extensions of one successful architecture, researchers have weaker incentives to pursue approaches that look unpromising on today’s tests. Leading AI models might improve rapidly while less conventional ideas remain underdeveloped.
That does not mean competing architectures have disappeared. Mamba uses selective state-space models to process sequences without conventional transformer attention. It represents a genuine alternative line of research, although its existence alone says little about whether the research community gives competing approaches enough support.
There is also a gap between the productivity of a research community and the capabilities of the systems it builds. More papers, more automated experiments and faster iteration do not guarantee more useful discoveries. We have explored a related measurement problem in OpenAI’s reported use of AI-agent workdays: activity is easier to count than validated scientific progress.
For funders and labs, the question is whether promising alternatives can obtain time, computing resources and serious evaluation. Conferences face a similar challenge when assigning reviewers to unfamiliar ideas. For researchers, improving a dominant benchmark and testing a genuinely different hypothesis are both useful work, but they develop a field in different ways.
More on AI-assisted research:
What to watch before calling AI research exhausted
The new preprint makes a reasonable case for tracking the diversity of research attention alongside model performance, publication volume and research spending. But vocabulary-based measures should be checked against direct evidence of novel methods, independent replication and whether unconventional research directions survive long enough to be tested fairly.
The strongest conclusion is still conditional. Concentrating research around transformers could make the next departure from transformers harder to discover. The authors have documented a suggestive pattern, not the breakthroughs that never happened. Keeping room for experiments outside today’s winning approaches is a sensible response to that uncertainty, especially when the cost of overlooking an unusual idea may be visible only much later.
Explore more from Popular AI:
Start here | Local AI | Builds & gear | Autonomy & policy | Fixes & guides | Popular AI podcast
Could the AI industry's focus on today's winning approaches be holding back tomorrow's breakthrough, or is that focus exactly what progress needs?