AI Could Help Researchers Get Funding—But at What Cost?

While many of us are just starting to get comfortable with AI’s role in our work, scientists have actually been leveraging AI to their advantage for years—using LLMs to help them write research papers faster and increase their output.

Yet despite the apparent benefits of using AI for publishing scientific research, “we don’t know much about what’s happening upstream in the scientific pipeline,” says Yifan Qian, a research assistant professor at the Kellogg Center for Science of Science and Innovation. “We know very little about how the use of AI and LLMs has been reshaping the U.S. federal research-funding landscape.”

How, for example, does using AI for grant proposals shape the type of research that scientists choose to explore, and how does it impact their chances of getting funding? Qian and his colleagues from the Kellogg School explored this line of questioning by analyzing research proposals submitted to and awarded by the National Science Foundation (NSF) and National Institutes of Health (NIH).

They found that when scientists used AI to write their proposals, the ideas hewed more closely to ideas that were already presented in prior funded research. In other words, the more researchers used AI to help with proposals, the less novel their proposed research ideas tended to be.    

In addition, proposals that relied more heavily on AI were more likely to get funded by the NIH and more likely to turn into published papers than proposals that had less help from AI. However, few of those papers turned out to be influential “hit” papers.   

Taken together, these findings underscore a potential downside to using AI for research proposals: it could result in fewer novel ideas (and more generic ideas) getting funded and, in turn, fewer breakthrough discoveries.   

“Science advances by exploring ideas that don’t yet look obvious,” says Dashun Wang, the Kellogg Chair of Technology and Director of the CSSI and the Northwestern Innovation Institute.  

“This gap [in what gets funded] is consequential,” adds Qian. “Funding operates upstream of almost everything else.”   

Analyzing grant-proposal language   

It was not obvious at the outset how AI would impact the way research proposals were written, explains Qian and Wang, who collaborated on this project with Zhe Wen, Alexander Furnas, Yue Bai, and Erzhuo Shao.   

On the one hand, AI is very good at making connections between disparate ideas. So Qian and his colleagues thought that perhaps AI could help scientists develop novel funding proposals by combining previous discoveries in unique ways. On the other hand, LLMs are mostly trained on proposals that previously received funding, so there was a chance that they might lean into conventional ideas that resemble previously successful proposals.   

The team started their study by examining a confidential dataset of all the research proposals submitted for federal funding from two large universities from 2021 to 2025. This allowed them to analyze both funded and unfunded proposals. (Unfunded proposals are generally filed away in researchers’ computers or university archives and are not publicly available.)   

They analyzed the text in the research proposals to distinguish between those that had more assistance from AI versus those that had less assistance and then assigned a score for that usage to each proposal. Not surprisingly, the use of AI for research proposals increased sharply in early 2023, around the time when ChatGPT became widely available. Still, there was a large percentage of proposals that either relied on AI minimally or did not appear to use AI at all.   

Next, Qian and his colleagues analyzed the wording in the proposals to see how closely it resembled the wording in previously funded projects. They found that the more AI was involved in a proposal, the less semantically distinct it was from prior funded work. Conversely, the less that AI was involved in a proposal, the more semantically distinct it was. 

The team wanted to make sure that this difference wasn’t due to the scientists’ demographics, such as gender and educational background. They found that, even when generated by the same scientist, proposals that were more heavily assisted by AI were less novel than proposals that were mostly or entirely human-written.   

AI’s impact on funding   

Collectively, the team’s findings indicate that the use of AI shaped the type of ideas that scientists pitched when asking for funding.    

This motivated the researchers to then look at what impact this had on the future outcome of the proposals. “Does this matter for the success of the proposal application, and what about the translation of the science in terms of the publication output?” Qian says.   

The answer was that it did matter, at least for the NIH-funded projects.   

Research proposals with high AI usage saw a roughly 4 percentage point increase in their likelihood of being funded by the NIH, compared with proposals with low AI usage. Similarly, proposals with high AI usage had about 5 percent more publications.    

Even though more AI use was associated with more publications, those research papers rarely turned out to become highly cited “hit” papers, which are usually more likely to cover new or influential ideas.   

However, this pattern did not occur for NSF-funded projects. To better understand why, the team performed a secondary analysis focusing on the largest field of study that both agencies fund—biomedical research—and found that the differences in funding and paper output still held true in this field.   

“This suggests that the different results are not simply due to a difference in scientific topics that the agencies fund,” Qian says. 

Figuring out what’s driving this difference requires further research, he says, but it could have to do with NIH’s preference for strong preliminary results in applications, or with NSF studies potentially taking longer to result in published papers.   

A shift to “safer” science   

AI may help researchers increase their output, but it also has the potential to change which ideas scientists propose and which of them get funded, Qian says. And that has implications for society.   

In order to produce breakthrough discoveries, agencies need to fund novel ideas that move beyond the topics that are already being funded and explored.    

“We’re not arguing that agencies should only fund novel ideas. But it makes sense to have a balanced portfolio,” Qian says. And AI, it seems, is shifting that balance toward more incremental proposals and discoveries.   

Scientists may turn to AI to help them submit more proposals that are “safe” and more likely to be funded, which could help progress their research and careers. But that could also mean less funding for high-risk, high-reward projects that lead to the types of breakthroughs that lead to scientific progress and potentially benefit society.   

“At an individual level, using AI may help you succeed in terms of funding, but collectively, there is a sort of convergence [of ideas], and this is a concern,” Qian says. 

“If AI increasingly learns from yesterday’s successful proposals,” Wang adds, “one of the questions we should ask is whether tomorrow’s scientific portfolio becomes less adventurous.”