A new study from Northwestern University’s Kellogg School of Management suggests that artificial intelligence is already influencing not only how scientists write grant proposals, but also which kinds of research receive public funding. Research proposals showing stronger signs of AI-assisted writing were about four percentage points more likely to receive support from the U.S. National Institutes of Health, the researchers found. Yet the same proposals tended to resemble ideas that federal agencies had funded previously, raising concerns that generative AI could make scientific funding more efficient while quietly narrowing the boundaries of discovery.
The study, led by Dashun Wang and Yifan Qian, examined the growing use of large language models in U.S. federal research funding. Its findings are among the first to investigate the role of tools such as ChatGPT before scientific work begins—at the moment when researchers compete for the resources needed to test new ideas. The researchers analyzed grant proposals and funded projects submitted to both the NIH and the National Science Foundation, looking for linguistic patterns associated with machine-assisted writing and comparing those patterns with funding outcomes and later publication records.
AI-assisted grant writing increased sharply after ChatGPT became publicly available in late 2022. Instead of spreading gradually and uniformly across the research community, its adoption appeared to divide proposals into two broad groups: those showing little evidence of language-model assistance and those showing substantially more. The researchers used computational methods to identify these signals, while also measuring how semantically distinctive each proposal was compared with recently funded research.
Semantic distinctiveness is a technical measure of how different an idea is from the existing research portfolio. A highly distinctive proposal may introduce an unfamiliar scientific question, combine fields in an unusual way or challenge assumptions embedded in previous work. A less distinctive proposal may still be valuable, but it is more closely aligned with established approaches, terminology and successful funding patterns. In the study, proposals with stronger indications of LLM involvement were consistently less distinctive, suggesting that AI-assisted writing may be associated with ideas that fit more comfortably within the agency’s recent expectations.
At the NIH, this alignment with previous funding patterns was accompanied by a measurable funding advantage. Proposals with stronger signs of AI involvement were more likely to be awarded grants, and projects that received funding went on to produce more follow-up publications acknowledging the awards. However, those projects did not generate more highly cited “hit” papers, which are often used as a rough indicator of unusually influential or breakthrough research. The pattern suggests that AI may help researchers present executable projects in a way that supports steady productivity without necessarily increasing the probability of transformative discoveries.
The results were notably different at the NSF. The researchers found no statistically significant relationship between AI use and either funding success or the number of follow-on publications. This contrast indicates that the effects of language-model assistance may depend on the culture, review criteria and disciplinary composition of each agency. NIH proposals may be evaluated within funding environments where feasibility, incremental progress and anticipated publication output carry particular weight, while NSF programs may place different emphasis on conceptual novelty or cross-disciplinary exploration.
The study does not establish that AI-assisted writing directly causes higher funding rates or less original science. The researchers emphasize that their analysis identifies associations, and they cannot determine precisely how reviewers responded to AI-generated or AI-polished language. One possibility is that human reviewers favor proposals that follow familiar rhetorical structures and clearly communicate achievable milestones. Large language models, trained on enormous collections of written material, may help applicants reproduce those established formats, sharpen their explanations and align their arguments with implicit expectations in peer review.
That possibility presents a difficult policy problem. Federal research agencies are expected to support both cumulative progress and high-risk, unconventional ideas. If AI tools are primarily trained on past successful proposals, they may encourage applicants to reproduce the language and conceptual structures that have already been rewarded. Over time, this could create a feedback loop: previously funded ideas become training material, AI systems help produce new proposals resembling those ideas, reviewers recognize the familiar patterns, and funding becomes increasingly concentrated around established research directions.
The concern is not simply that scientists are using software to improve grammar or organize complex arguments. Generative AI can influence the framing of hypotheses, the selection of supporting evidence and the description of what counts as a plausible research path. When such tools become part of a competitive funding system, they may affect the scientific portfolio before experiments are conducted and before new evidence can challenge prevailing assumptions. As Wang and Qian argue, the central question is whether AI will help science communicate better while preserving intellectual diversity—or make the future of research look increasingly like its past.
The authors say the findings have implications for transparency, research diversity and public trust in the stewardship of taxpayer-supported science. They also point to the need for further research into how language models are used, how reviewers interpret AI-assisted proposals and whether agency evaluation systems can identify genuinely unconventional ideas that do not resemble previous winners. The study, titled “The Rise of Large Language Models and the Direction and Impact of US Federal Research Funding,” is scheduled for publication in Proceedings of the National Academy of Sciences on August 11, 2026. Its broader warning is that the influence of AI on science may begin long before a paper is published: it may start with the decision about which questions society chooses to fund.
Subject of Research: The influence of large language models and AI-assisted writing on U.S. federal scientific research funding, proposal originality and research outcomes.
Article Title: The rise of large language models and the direction and impact of US federal research funding
Web References: https://doi.org/10.1073/pnas.2601439123 ; https://arxiv.org/abs/2601.15485
References: Proceedings of the National Academy of Sciences; Northwestern University’s Kellogg School of Management; Dashun Wang and Yifan Qian
Keywords: Generative AI, large language models, ChatGPT, scientific research funding, NIH, NSF, grant proposals, peer review, research diversity, semantic distinctiveness, science policy, artificial intelligence
