Large language models reduce originality of research proposals

The use of large language models (LLMs) has increased dramatically in US research funding proposals since 2023 and has led to less original and more generic ideas that often replicate prior projects, new analysis finds.

A paper published in Proceedings of the National Academy of Sciences on 11 August examined 5,700 confidential grant proposal submissions and 131,000 publicly released awards for grants from the US National Science Foundation (NSF) and National Institutes of Health (NIH) between 2021 and 2025.

Using textual analysis, the authors, from the Center for Science of Science and Innovation at the Kellogg School of Management at Northwestern University, found that LLM use rose sharply from 2023 across both the point of submission and among funded awards that have passed peer review.

The researchers found that grant proposals with high LLM involvement were “less semantically distinctive” from projects that had been recently funded by both agencies.

However, the implications of LLM use in funding proposals differed. At the NIH, research proposals that involved the use of LLMs were more likely to be funded and to result in more early-stage publications. By contrast, no such associations were observed at the NSF.

But the authors note that any productivity gains associated with LLM use at the NIH were concentrated in lower-impact scientific publications rather than “the most highly cited work”.

They write that their findings suggest “LLM use is already reshaping how scientific ideas are articulated and evaluated in the federal funding system”, emphasising that this has widespread impacts on “research diversity, transparency, and public trust in the stewardship of taxpayer-supported science”.

The researchers chose to investigate LLM use in research funding because of a lack of empirical evidence about how these foundational models are employed in proposals, despite their rapid uptake.

Interrogating LLM use at integral stages in the federal funding pipeline, they wanted to understand how LLM use “relates to the positioning of scientific ideas, their selection for funding, and their translation into publicly supported research output”.

Although the paper acknowledges that LLMs represent “a potentially transformative development for scientific work”, reducing the time it takes for researchers to draft and revise text and helping them to communicate with fluency and speed, it adds that these same features “raise critical concerns” by potentially narrowing the scope and variation of ideas.

“AI can expand individual scientists’ productivity and impact while simultaneously contracting the collective focus of science, underscoring the possibility that gains at the individual level may coexist with reduced diversity at the system level,” the study explains.

“LLM use could shift the balance between exploration and exploitation in scientific funding, pulling proposals toward the centre of existing funding patterns and potentially narrowing the diversity of ideas that are submitted for or receive public support, even as it may lower the barriers to producing competitive applications,” it adds.

In recent years US federal agencies have issued guidance on how generative AI can be used in research proposal preparation.

In July 2025, the NIH said it would not consider applications “substantially developed by AI” as original work but added that proposers may use AI tools to assist in limited aspects of application preparation.

The NSF encourages but does not require applicants to disclose how generative AI was used. It says that the proposer is fully responsible for the accuracy and originality of their submission.

The authors say the move to issue guidance on AI use underscores “both the growing salience of LLMs in public funding and the need for empirical evidence on how they are being used and with what consequences”.

rosalind.skillen@timeshighereducation.com