Grant Proposals Written With AI Help Win More NIH Funding, and Break Less New Ground

There is no register showing which federal grant applications have been written with help from a chatbot. NIH and NSF applicants have not been required to say. So a Northwestern team looked for the traces instead.

They collected grant abstracts written in 2021, before ChatGPT reached the public, and treated them as a record of how humans write when asking the government for research money. Then they asked a version of GPT-3.5 to rewrite the same abstracts. That produced a second record. By comparing the words each version favored, the researchers built a ruler they could apply to later batches of grants. A Stanford group used the same approach to estimate chatbot involvement across more than a million papers and preprints.

They applied that ruler to more than 130,000 NSF and NIH awards and about 5,700 confidential proposals from two large American research universities. That second collection is unusual because it includes proposals that lost, documents that almost never leave the agencies that reject them.

Grant Writing Did Not Drift, It Split in Two

The share of machine-flavored sentences stayed level through 2021 and 2022, then rose sharply in 2023 as ChatGPT spread. That much was expected. What surprised the researchers was the shape of the change. Scores didn’t spread smoothly across a range, as they might have if everyone had started using chatbots a little. Instead, they piled up in two places. One group sat near zero. Another clustered around roughly one sentence in ten. Some scientists appeared not to use these tools in their abstracts at all. Others leaned on them heavily. There wasn’t much of a middle.

How much AI appears in a grant application is interesting. But what the application proposes can shape which science actually gets done. So the team turned each abstract into a string of numbers designed to capture its meaning rather than its wording. They then measured how far that meaning sat from everything the same agency had funded the year before.

Across all four collections of grants, public and private, NSF and NIH, more AI involvement went with less distance. The gap wasn’t huge, but it was consistent. Moving from a light AI touch to a heavy one pushed a proposal down about four or five places out of 100 in the distinctiveness rankings. The analysis compared each investigator with their own other proposals. That matters because it makes the result less likely to reflect cautious scientists simply being more likely to use chatbots. The same person, when writing with more machine help, wrote a less distinctive grant. A similar pattern has appeared outside research funding. In five brainstorming experiments at Wharton, groups using ChatGPT produced narrower sets of ideas. In one round, nine different participants independently gave their invented toy the same name.

The Sameness Is in the Ideas, Not Just the Prose

There’s an obvious objection. Maybe these tools simply smooth out everyone’s writing, making different ideas sound more alike. The team ran a test designed to separate wording from substance. They returned to the 2021 abstracts, had a language model rewrite them, and measured them again. The science had not changed at all. Only the language had. Yet the distinctiveness scores stayed essentially the same. That suggests the convergence lies in the ideas, not just the phrasing. “Science advances by exploring ideas that don’t yet look obvious,” said Dashun Wang, who directs Northwestern’s Center for Science of Science and Innovation. “If AI increasingly learns from yesterday’s successful proposals, one of the questions we should ask is whether tomorrow’s scientific portfolio becomes less adventurous.”

At NIH the Pattern Paid Off, at NSF It Did Not

Then comes the finding that makes this more than a study of writing habits. Among the confidential NIH submissions, proposals with heavier AI involvement were more likely to be funded. Across the span from light use to heavy use, the difference was roughly four percentage points. That’s substantial given how thin the odds have become. NIH funded about one in eight research project applications in 2025, down from about one in five two years earlier. Four percentage points can represent a meaningful share of an applicant’s chances. The same analysis found no such pattern among NSF submissions.

None of this shows that using a chatbot causes a grant to get funded, and the authors are careful about that. Comparing investigators with themselves rules out many possible differences between people. But it can’no’t rule out another possibility: scientists may be more likely to use these tools when the project they are pitching is already safe, conventional, and easier to build.

More Papers, Not More Important Ones

Looking at what happened after awards were made, NIH grants with heavier AI traces produced about 5 percent more publications. But when the researchers narrowed the count to papers in the top 5 percent of citations for their field and year, the advantage disappeared. In other words, the grants produced more papers without clear evidence that they produced more influential work. NSF awards showed no difference in publication output.

Several caveats matter. The confidential proposals came from only two universities, so they can”t represent the entire American research system. The analysis also examined abstracts rather than the full research plans that reviewers actually read. And the publication counts covered only the first year or two after an award, which may be too early to judge whether a grant produced lasting discoveries.

Agencies are already trying to write rules for a practice they can’no’t easily see. NIH said in July 2025 that applications substantially developed by AI will not count as an applicant’s original ideas. The agency noted that these tools had enabled some investigators to submit more than 40 distinct applications in a single submission round. NSF asks proposers to say whether and how they used generative AI. It also forbids reviewers from uploading proposals into outside chatbots.

NIH also runs programs, including the Pioneer and Transformative grants, created specifically for ideas the agency expects would not fare well in its own ordinary review. Those programs exist because the center of the portfolio is not always where the surprises live. That makes one line in this paper’s acknowledgments especially striking: the authors used GPT-5 and Claude Opus 4 to improve the language and readability of the manuscript.

Frequently Asked Questions