Plate Nº 65 · recorded November 16, 2026

Science Policy & FundingReported finding

AI-Written Grant Proposals Win More but Follow the Herd

A PNAS study of 137,000 federal awards finds AI-assisted grant proposals cluster around conventional topics — yet win more funding at NIH, where AI traces boosted award odds by 4 percentage points.

By James Calloway4 min read749 words

In brief

  1. Northwestern researchers analyzed 1,600 NSF and 4,100 NIH university proposals plus 57,000 NSF and 74,000 NIH funded awards from 2021–2025.
  2. Higher AI involvement corresponded to a ~5-percentile-point drop in topic distinctiveness at NSF and ~4 points at NIH.
  3. At NIH, moving from the 25th to 75th percentile of AI involvement raised funding odds by roughly 4 percentage points; NSF showed no significant effect.
  4. AI traces in grants rose sharply starting 2023, with a second cluster of abstracts showing 10–15 percent AI-modified sentences.
  5. The gain in publications (~5 percent more at NIH) disappeared for highly cited papers in the top 5 or 1 percent.
Scientists who use AI propose less distinctive ideas for federal research grants
Plate Nº 65Scientists who use AI propose less distinctive ideas for federal research grants — jarmoluk / pixabay

Grant applications with heavy AI involvement showed roughly a 5-percentile-point drop in topic distinctiveness at the National Science Foundation — and at the National Institutes of Health, they were about 4 percentage points more likely to win funding. That is the core finding of a new PNAS study led by Yifan Qian and colleagues at Northwestern University, which tracked the fingerprints of generative AI in U.S. federal grant proposals from 2021 to 2025.

The trade-off sounds tidy: more funding success, less originality. But the paper carries caveats, and its own authors admit they cannot prove cause and effect. Still, the numbers deserve a hard look.

What did the researchers actually measure?

The team combined confidential submissions from two large U.S. research universities — 1,600 NSF proposals and 4,100 NIH proposals, including unfunded ones — with public records of every funded award nationwide: 57,000 NSF awards and 74,000 NIH awards.

To estimate AI involvement, they compared word-choice frequencies between human and machine-edited writing. They built a baseline from public grant abstracts written in 2021, before ChatGPT existed, then had OpenAI's GPT-3.5 rewrite those same abstracts to map typical AI vocabulary. The results: AI traces stayed flat before late 2022, rose sharply in 2023, and split into two camps among 2023–2025 grants — one with near-zero AI modification, another centered around 10 to 15 percent AI-modified sentences. The pattern held when tested against Claude, Gemini, and Llama.

What does this change for scientists seeking funding?

Here is the practical question. If you write grants, does AI help or hurt? At NIH, apparently help — moving from the 25th to the 75th percentile of AI involvement meant roughly a 4-percentage-point higher chance of an award. At NSF, no statistically significant effect. NIH awards with more AI traces also produced about 5 percent more publications in the first year or two, though that gain vanished when restricted to highly cited papers in the top 5 percent or top 1 percent.

The distinctiveness finding ran in one direction everywhere: more AI, fewer novel topics. The drop was about 5 percentile points at NSF and 4 at NIH, measured by how far each abstract sat from everything the agency funded the previous year.

One clever check: when the researchers had AI rewrite 2021 human abstracts without changing the underlying science, the distinctiveness scores barely moved. The convergence, they argue, reflects where projects sit substantively — not just polished wording.

Why would NIH and NSF differ?

The gap persisted even when the comparison was restricted to biomedical and clinical sciences across NSF, NIH, and the Department of Defense, so disciplinary mix alone cannot explain it. The authors suggest NIH review panels may reward feasible, incremental projects that generate steady publications, and AI drafting may help applicants hit that template. Journal papers also matter more at NIH, while NSF often funds software, datasets, instruments, and training.

Careful, though. The authors themselves call these institutional explanations untested. It is a hypothesis, not a finding.

How solid are the numbers?

Not bulletproof. The study is observational. Scientists may reach for AI precisely when their proposals are already incremental, straightforward, or aligned with agency priorities — the direction of causation remains open. Vocabulary analysis also captures only how abstracts are summarized, not whether researchers used AI to brainstorm hypotheses or analyze data. And the confidential records came from just two universities, leaving smaller colleges unexamined.

The publication window is short, too: one to two years for grants starting in 2023 and 2024. Whether AI-assisted grants eventually produce breakthroughs — or just routine papers — will take longer follow-up studies to answer.

The broader picture already shows tension. A 2026 study of more than 41 million natural science papers found AI users published roughly three times as many papers and drew nearly five times as many citations, while the collective range of topics across science shrank by nearly 5 percent. A 2024 experiment with short-story writers found the same pattern: AI access improved individual stories while making them more alike.

One detail worth savoring: the authors disclosed using GPT-5 and Claude Opus 4 to polish their own manuscript's language. AI studying AI. The irony writes itself — though presumably with human review.

The study, "The rise of large language models and the direction and impact of US federal research funding," was authored by Yifan Qian, Zhe Wen, Alexander C. Furnas, Yue Bai, Erzhuo Shao, and Dashun Wang.

via psypost.org (Source)

Filed under

  • generative-ai
  • research-grants
  • nsf
  • nih
  • funding-outcomes
Share this article:

More from James Calloway

James Calloway

Show full bio

Staff writer covering regulation, policy and their impact on general science news at SciBeat.

411 articles

Nearby plates

« Previous article