A new study published in the Proceedings of the National Academy of Sciences finds that scientists who use generative AI to help draft federal grant applications tend to propose research topics that more closely resemble recently funded projects. At the National Institutes of Health, submissions with greater traces of AI writing were more likely to win funding, and awards with higher AI involvement were followed by modestly more publications in the first year or two, concentrated in routine studies rather than high‑impact breakthroughs.
In the United States, much of academic science begins with a competitive grant proposal—a formal written application in which researchers pitch planned studies to agencies such as the NSF and the NIH as they compete for taxpayer funding. Over the past few years, scientists preparing grant texts have increasingly turned to large language models—AI programmes trained on vast collections of existing writing to generate and revise text by predicting probable word sequences. Because grant selection happens before experiments are funded, shifts at the proposal stage can influence which scientific questions are pursued.
Earlier research shows AI tools entered everyday scientific workflows rapidly after ChatGPT was launched in late 2022. A body of work tracked vocabulary shifts in preprints and journal articles through 2024, with a 2025 analysis estimating AI‑modified writing across more than a million papers, including about 22 percent of computer science texts. Similar tools have also begun to influence gatekeeping, with a 2024 study of computer science conference peer reviews finding substantial AI modification in texts submitted for review, especially near submission deadlines.
Because unfunded grant applications are confidential, researchers have found it difficult to study how AI writing relates to proposal topics and outcomes. The study authors—led by Yifan Qian and colleagues at Northwestern University—combined confidential university submissions with nationwide records of funded federal awards to examine patterns across four datasets, spanning start dates from 2021 to 2025. Two private datasets from large U.S. universities included 1,600 NSF proposals and 4,100 NIH proposals, while two public datasets from Dimensions captured nationwide funded awards—57,000 NSF awards and 74,000 NIH awards—with deidentified data used for the main analyses.
To estimate AI involvement in grant abstracts, the team employed a statistical detection method developed for the million‑paper study, comparing word‑choice frequencies between human and machine‑edited writing. They began with public grant abstracts from 2021 to establish human baselines and prompted OpenAI’s GPT‑3.5 to rewrite the same abstracts to map AI word distributions. Using these benchmarks, they estimated the fraction of AI‑modified sentences in rolling three‑month windows and within each abstract.
Across both private submissions and public awards at NSF and NIH, estimated AI involvement rose sharply beginning in 2023 following a flat pattern beforehand. NIH submissions and awards followed the same timeline as NSF, though at lower overall levels. Among grants starting between 2023 and 2025, AI involvement exhibited a two‑peaked distribution: some abstracts showed near‑zero AI modification, while others clustered around about 10 to 15 percent AI‑modified sentences. The patterns held when the analysis was repeated with Claude, Gemini and Llama models, and after excluding AI buzzwords and related terms.
The researchers then measured each abstract’s semantic distinctiveness—how different its topic and content were from recently funded projects—using a specialised scientific language model that maps text meaning into numerical coordinates. They calculated each 2023–2025 abstract’s average distance from all grants funded by the same agency in the preceding year, ranking each grant within its agency and year. The models adjusted for dollar amounts, start year, scientific field, and individual researcher identity, comparing each scientist’s applications only against that scientist’s other submissions to the same agency.
Across private and public data, higher AI involvement within a scientist’s work correlated with lower distinctiveness, meaning projects tended to align more closely with earlier funding patterns. Comparing a grant at the 75th percentile of AI involvement with one at the 25th percentile yielded about a 5‑point drop in distinctiveness for NSF awards and around 4 points for NIH awards.
When 2021 human abstracts were paired with AI‑rewritten versions that preserved the underlying ideas, the two sets produced nearly identical distinctiveness scores. The authors interpret this as evidence that convergence reflects substantive similarities to prior funding rather than merely wording changes, and they note that proposals with heavier AI involvement did not appear more distant from a scientist’s own earlier proposals at the same agency, suggesting the shift tracks the agencies’ recent funding patterns rather than researchers’ established lines of work.
Examining funding outcomes using confidential 2023–2025 university submissions, the study found no statistically significant association between AI involvement and winning NSF funding. By contrast, at NIH, moving from the 25th to the 75th percentile of AI involvement corresponded to roughly a 4‑percentage‑point increase in the likelihood of obtaining an award, although the baseline rate is not disclosed in the dataset.
To assess downstream productivity, the team linked publicly funded NSF and NIH awards starting in 2023–2024 to resulting preprints and journal articles by November 2025. At NSF, AI involvement did not significantly predict increases in publication counts within the initial one‑to‑two year window. At NIH, increasing AI involvement from the 25th to the 75th percentile predicted roughly 5 percent more publications, but this association lost statistical significance when limited to highly cited papers ranking in the top 5 per cent or top 1 per cent of citations in their field and year.
The tendency for AI‑assisted proposals to cluster around conventional topics echoes a 2024 short‑story experiment in which AI access boosted perceived quality and enjoyment while making stories more alike. A parallel laboratory study on verbal creativity found language models could match or exceed human scores individually while producing far more homogeneous outputs, though that work concerned controlled verbal tasks rather than grant writing.
Even within biomedical and clinical sciences, agency differences persisted, suggesting that disciplinary mix alone does not explain the NIH–NSF gap. The authors propose institutional norms may play a role, speculating that NIH panels might reward feasible, incremental projects designed to yield steady publications, with AI drafting helping applicants conform to those templates.
They caution that journal articles represent a larger share of NIH outputs than NSF, which often supports software, datasets, instruments and training, though these explanations remain untested. More broadly, the authors suggest generative AI currently helps researchers package and communicate ideas faster, without removing the slower bottlenecks of running experiments and gathering data.
Several caveats accompany the findings. While comparing each scientist against their own other proposals controls for stable individual skill, the study is observational and cannot prove that AI tools cause safer topic choices or higher NIH funding odds. Researchers may rely more on AI assistance for incremental, straightforward, or agency‑prioritised proposals.
Estimating AI involvement from abstract vocabulary captures how projects are summarised rather than detailing all uses of AI in brainstorming hypotheses, data analysis, or full technical planning. The authors themselves disclosed using GPT‑5 and Claude Opus 4 to improve their manuscript’s language and readability, illustrating the pervasiveness of AI in academic writing.
The confidential records come from only two major universities, leaving open whether patterns hold across smaller institutions. And although older cohorts suggest early publications predict longer‑run productivity, this study followed grants from 2023 and 2024 for only up to two years, so longer follow‑up will be needed to assess whether AI‑assisted grants yield high‑impact breakthroughs.
The study, “The rise of large language models and the direction and impact of US federal research funding,” is by Yifan Qian, Zhe Wen, Alexander C. Furnas, Yue Bai, Erzhuo Shao and Dashun Wang.