A strong AI signature in applications from two universities was associated with a higher probability of receiving NIH funding
A strong AI signature in applications from two universities was associated with a higher probability of receiving NIH funding
In an article published on August 11, the authors compared the texts of applications submitted to the US National Institutes of Health (NIH) and the US National Science Foundation (NSF) with the outcomes of their funding competitions. Here, the “AI signature” is a statistical estimate of the proportion of sentences that resemble text edited by a large language model (LLM). The measure describes the language of an application. Among NIH applications from two universities, moving from the bottom quartile of this measure to the top quartile was associated with a four percentage point higher probability of receiving funding.
Grant applications determine which research ideas receive funding and what work those ideas initiate in the laboratory. The authors collected confidential applications submitted by two large US research universities between 2021 and 2025, including applications that were funded, rejected, or still awaiting a decision. They supplemented these records with public abstracts from grants that had already been awarded, which allowed them to compare application texts with competition outcomes.
The authors constructed the AI signature measure from two language profiles. The first came from abstracts written in 2021, before ChatGPT became widely used. The second was produced by having GPT-3.5 rewrite the same abstracts. Comparing these profiles indicates where the text of each application falls between the original human language and the language produced after model editing.
The researchers then tested how much each description of an idea differed from the recent portfolio of the relevant agency. They compared every application with abstracts from grants awarded by the same agency whose funding had begun one year earlier. Greater textual similarity meant that the project description was more familiar within that portfolio. Across all four datasets, a higher AI signature was associated with lower distinctiveness. In the public grant data, it lowered an application’s position in the distinctiveness ranking by approximately four places for NIH and five places for NSF.
As a control, the authors had a model rewrite the 2021 abstracts while preserving their scientific content, then compared the rewritten versions with earlier grants. Their positions in this comparison remained almost unchanged. This test helps separate linguistic polish from the relationship between how a project is described and how closely it resembles work that has already received support.
The authors also compared applications from the same researcher while accounting for field, year, and requested funding. At NIH, a higher AI signature was associated with success in the competition. The NSF data showed no such association with application success. Among publicly documented NIH grants, a high AI signature was also associated with approximately 5% more publications during the first one to two years.
In a notice dated July 24, 2025, NIH reported that one principal investigator may have used AI to submit more than forty different applications in a single funding round. For applications with deadlines on September 25 or later, the agency imposed a limit of six applications per principal investigator per calendar year. The same document states the rule as follows:
“NIH will not consider applications or sections of applications substantially developed by AI to represent an applicant’s original ideas.”
In the NIH data, a high AI signature was associated both with greater similarity to projects that had already received support and with more frequent success in the funding competition.