AI Helps Researchers Win NIH Grants. Will Science Suffer?

New research shows that scientists who rely heavily on artificial intelligence to write grant applications are more likely to get funding from the National Institutes of Health. But that tactic could come at the expense of exploring more novel scientific ideas.

That’s one of the central takeaways of a study published this month in the Proceedings of the National Academy of Sciences, which analyzed more than 125,000 grant applications—including funded, unfunded and pending proposals—submitted to the NIH and National Science Foundation from 2021 to 2025. The study’s time frame encompasses the rapid rise of generative AI tools that first became widely available in late 2022. Using word-distribution modeling to identify proposals suggestive of high levels of large language model involvement, the researchers identified a surge in AI use between 2023 and 2025.

“Federal funding is the primary mechanism through which the U.S. converts public resources into scientific knowledge,” Yifan Qian, a research assistant professor at Kellogg School of Management at Northwestern University and co-author of the new paper, told Inside Higher Ed. “Understanding the forces that influence the federal funding process is critical not only for science policy and scientific progress, but also for accountability of public investment in research.”

While numerous other studies and surveys have already established that scientists are increasingly using AI to write peer-reviewed papers, less is understood about how the technology is shaping federally funded research. That’s “because the proposal data is typically confidential,” Qian said. But in addition to publicly available data about awarded grants, “our team was able to get two full sets of confidential NIH and NSF submissions from two large research universities, which allowed us to study this question.”

The study found that the NSF and NIH—two of the largest funders of university-led scientific research—have had different responses to AI-generated grant applications.

While it didn’t find a correlation between applications suggestive of high large language model involvement and accepted NSF proposals, NIH submissions with high LLM involvement corresponded to a four-percentage-point jump in funding probability compared to those with low LLM involvement.

That raised a follow-up question for researchers, who wanted to know if researchers who used an LLM to write a grant application were also more likely to publish related work.

Again, the answer was agency-dependent.

While there was no meaningful association between LLM involvement and publication output for NSF awards, successful NIH grant applications with high LLM involvement were strongly associated with more resulting publications. According to the study, NIH grants stemming from applications with high LLM involvement published 5 percent more resulting papers than those with less AI involvements. However, higher publication volume didn’t translate to high levels of influence. Among the most cited papers, researchers found no notable advantage for NIH-funded projects awarded to grant applications showing signs of heavy AI use.

The researchers drew no definitive conclusions about the differences, though they offered some possible explanations such as that “NIH funding and review norms may more strongly reward incremental, executable projects that yield multiple publications, and LLM-assisted drafting may help proposals conform to those established templates,” the paper said. “The contrast also raises questions regarding specific features of agency review.”

Regardless of whether applications were submitted to the NIH or the NSF, the proposals with high LLM involvement were more similar to projects that had already received federal funding. That convergence “is not simply a byproduct of LLM-induced surface-level rewriting, but reflects shifts in the substantive positioning of proposals and awards,” the researchers wrote.

“A portfolio that is closer to recent funding patterns may reflect improved clarity, tighter alignment with reviewer expectations, or lower transaction costs in articulating a fundable project,” they continued. “But it also implies reduced exploration in the idea landscape, which matters for public funders explicitly tasked with sustaining high-variance discovery, with implications for long-run impact and sustainability of science.”

Although the study indicates that many researchers are benefiting from letting generative AI do the heavy lifting of writing grant applications, both the NIH and NSF have AI use policies on the books that emphasize research integrity. The NSF’s 2023 policy encourages applicants to disclose if and how they used generative AI, while holding them “responsible for the accuracy and authenticity” of their submission, including content developed with generative AI. In September 2025, amid a surge in grant applications, the NIH cracked down on AI use by activating a new policy that considers “applications that are either substantially developed by AI, or contain sections substantially developed by AI” a violation of the agency’s expectations that proposals constitute an applicant’s original ideas.

But Qian said that resisting a potential LLM-induced scientific research slowdown will require agencies to pass even more detailed policies about what is and isn’t an acceptable use of AI. “Specifying whether they want people to write their own first drafts and use AI to help detect grammar errors, for example,” he said. “That would be helpful.”