Scientists who use generative artificial intelligence to help write federal grant applications tend to propose research topics that more closely resemble recently funded projects, according to a new study published in PNAS. At the National Institutes of Health, submissions with greater traces of AI writing were more likely to win funding, and awards with more AI traces were followed by modestly more publications in the first year or two, concentrated in routine studies rather than highly cited 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 federal agencies such as the National Science Foundation (NSF) and the National Institutes of Health (NIH) to compete for taxpayer funding. Over the past few years, scientists preparing academic texts have increasingly turned to large language models: artificial intelligence programs trained on massive collections of existing text, which generate and revise writing by predicting probable sequences of words. Because grant selection happens before experiments are funded, staffed, or published, shifts at the proposal stage can influence which scientific questions get pursued at all.
Earlier studies of academic publishing indicate that generative AI entered everyday scientific workflows rapidly after ChatGPT launched in late 2022. Tracking vocabulary shifts across preprints and journal articles through September 2024, a 2025 analysis of more than one million scientific papers estimated a steady climb in AI-modified writing, which reached an estimated 22 percent of the text in computer science papers. Similar tools also entered scientific gatekeeping, with a 2024 study of computer science conference peer reviews estimating that between 6.5 percent and 16.9 percent of submitted review text showed signs of substantial AI modification, particularly near submission deadlines.
Although language models can reduce the effort needed to polish complex prose, their reliance on patterns drawn from published papers and public grant archives has raised questions about whether they narrow the variety of scientific ideas. Because these systems tend to generate high-probability, field-typical phrasing, using them to prepare proposals could pull new projects closer to established norms. Consistent with that tension, a 2026 study of more than 41 million natural science papers found that scientists using AI tools published roughly three times as many papers and received nearly five times as many citations as peers who did not, yet the collective range of topics studied across science shrank by nearly 5 percent.
Don’t miss the next big psychology finding
Every morning, PsyPost sends a short briefing on the most interesting new peer-reviewed research in psychology and neuroscience.
Whether that same trade-off between individual output and collective variety is playing out in government grant competitions has been difficult to study, largely because unfunded grant applications are confidential. Faster writing also does not automatically overcome the practical bottlenecks of scientific execution, such as running laboratory experiments, coordinating teams, and collecting data. To examine how AI writing relates to proposal topics, funding decisions, and follow-on research papers, Yifan Qian and colleagues at Northwestern University combined confidential university grant submissions with nationwide records of funded federal awards.
The researchers analyzed four datasets covering grant proposals and awards with start dates from 2021 to 2025. Two confidential datasets from two large U.S. research universities included funded, unfunded, and pending applications, totaling 1,600 NSF proposals and 4,100 NIH proposals. Two public datasets from the Dimensions research database captured all funded awards nationwide over the same period, totaling 57,000 NSF awards and 74,000 NIH awards, with deidentified data for the main analyses posted publicly online.
To estimate AI involvement in grant abstracts, the team used the statistical detection method developed for the million-paper publishing analysis described above, which compares word-choice frequencies between human and machine-edited writing. They took public grant abstracts from 2021, prior to ChatGPT’s release, to establish baseline human word patterns and prompted OpenAI’s GPT-3.5 model to rewrite those same abstracts to map typical AI word distributions. Using these benchmarks, the authors estimated the fraction of AI-modified sentences across rolling three-month windows and within each individual abstract.
Across both private university submissions and public funded awards at NSF and NIH, estimated AI involvement stayed flat before late 2022 and rose sharply beginning in 2023. NIH submissions and awards followed the same upward timeline as NSF, though at lower overall levels. Among individual grants starting between 2023 and 2025, AI involvement followed a two-peaked distribution: one group of abstracts showed near-zero AI modification, while a second group centered around roughly 10 to 15 percent AI-modified sentences. These patterns held when the team repeated the analysis using Claude, Gemini, and Llama, and after stripping out promotional buzzwords and AI-related technical terms.
Next, the authors measured each abstract’s semantic distinctiveness, meaning how different its topic and content were compared to recently funded projects. Using a specialized scientific language model that maps text meaning into numerical coordinates, they calculated each 2023 to 2025 abstract’s average distance from all grants funded by the same agency during the previous year, ranking each grant as a percentile within its agency and year. Their statistical models adjusted for dollar amounts, start year, scientific field, and individual researcher identity, comparing each scientist’s applications only against that same scientist’s other applications.
Across both private submissions and public awards at both agencies, higher AI involvement within a scientist’s work was consistently associated with lower distinctiveness, positioning projects closer to what the agency had funded the prior year. Comparing a grant at a higher level of AI involvement (the 75th percentile) with one at a lower level (the 25th percentile) corresponded to roughly a 5-percentile-point drop in distinctiveness for NSF awards, and about a 4-percentile-point drop for NIH awards.
When the team tested 2021 human abstracts alongside AI-rewritten versions that kept the underlying scientific ideas unchanged, the two sets showed nearly identical distinctiveness scores, which the authors read as evidence that the convergence reflects where projects sit substantively, not just how they are worded. At the same time, proposals with heavier AI involvement were not more distant from a scientist’s own earlier proposals at the same agency, suggesting the shift was toward the agency’s recent funding patterns rather than away from researchers’ established lines of work.
Using the confidential university submissions from 2023 to 2025, the researchers next tested whether AI involvement was linked to winning funding, again comparing within individual scientists and adjusting for field, year, and requested funding. At NSF, the association between AI involvement and proposal success was not statistically significant. At NIH, by contrast, moving from the 25th percentile to the 75th percentile of AI involvement corresponded to roughly a 4-percentage-point increase in the likelihood of winning an award, though the paper does not report the baseline funding rate in that university sample.
To examine downstream research output, the team linked public NSF and NIH awards starting in 2023 and 2024 to resulting preprints and published journal articles recorded through November 2025. At NSF, the relationship between grant-level AI involvement and resulting paper counts over this early one- to two-year window was not statistically significant. At NIH, moving from the 25th percentile to the 75th percentile of AI involvement predicted roughly 5 percent more publications, but that association weakened and was no longer statistically significant when restricted to highly cited papers ranking in the top 5 percent or top 1 percent of citations worldwide for their field and year.
The tendency for AI-assisted proposals to cluster around conventional topics aligns with a 2024 experiment with short-story writers, which found that access to AI ideas boosted ratings of individual story quality and enjoyability while making stories more similar to one another, though the new study observed lower semantic distinctiveness in real-world scientific grant proposals rather than short fiction written in an experimental setting.
A parallel pattern emerged in a laboratory study testing AI models on verbal creativity tasks, which found that language models matched or exceeded human scores individually while generating much more homogeneous responses across outputs, although that research examined laboratory verbal tasks rather than scientific grant writing.
When the researchers restricted their comparison to biomedical and clinical sciences across NSF, NIH, and the Department of Defense, agency differences persisted, indicating that disciplinary mix alone does not explain the gap between NIH and NSF. The authors suggest that institutional review norms might play a role, proposing that NIH panels may more strongly reward feasible, incremental projects designed to generate steady publications, and that AI drafting may help applicants conform to those templates.
They also note that journal papers represent a larger share of expected outputs at NIH than at NSF, which often supports software, datasets, instruments, and training, though they caution that these institutional explanations remain untested. More broadly, the authors propose that generative AI currently helps researchers package and communicate ideas faster without removing the slower physical bottlenecks of running experiments and gathering data.
There are a few things to keep in mind when interpreting these findings. Even though comparing each scientist against their own other proposals accounts for stable personal differences in skill and writing style, the study is observational and cannot establish that AI tools directly cause safer topic choices or higher NIH funding odds. As the authors note, researchers may rely more heavily on AI assistance when drafting proposals that are already incremental, straightforward to execute, or aligned with agency priorities.
In addition, estimating AI involvement from abstract vocabulary captures how projects are summarized rather than all the ways scientists might use AI to brainstorm hypotheses, analyze preliminary data, or write full technical plans. Reflecting how widespread such writing assistance has become in academia, the authors themselves disclosed using GPT-5 and Claude Opus 4 to improve the language and readability of their manuscript.
The confidential application records also came from just two major research universities, leaving open whether selection patterns look the same across smaller colleges and specialized institutes. Finally, although the authors’ checks on older grant cohorts indicate that early publication counts capture a meaningful share of eventual output and predict longer-run productivity, the current analysis only tracked the first one to two years after grants began in 2023 and 2024, so longer follow-up studies will be needed to see whether AI-assisted grants eventually yield high-impact breakthroughs.
The study, “The rise of large language models and the direction and impact of US federal research funding,” was authored by Yifan Qian, Zhe Wen, Alexander C. Furnas, Yue Bai, Erzhuo Shao, and Dashun Wang.