LLMs linked to less original grant proposals

Large language model use has increased sharply in US research funding proposals since 2023, with new analysis linking the shift to less original ideas that more often resemble previous projects. The findings are based on textual analysis of 5,700 confidential grant proposal submissions and 131,000 publicly released awards from the US National Science Foundation and National Institutes of Health between 2021 and 2025.

Researchers at Northwestern University’s Kellogg School of Management found that proposals with high LLM involvement were “less semantically distinctive” from recently funded projects at both agencies. At the NIH, LLM-assisted proposals were more likely to win funding and produce early-stage publications, but those gains were concentrated in lower-impact work rather than “the most highly cited work”. No comparable association was found at the NSF.

The findings raise concerns that AI tools may improve individual productivity while narrowing the collective range of scientific ideas. Federal agencies have begun responding: in July 2025, the NIH said applications “substantially developed by AI” would not be treated as original work, while the NSF encourages but does not require disclosure of generative AI use.