Closer to Last Year's Winners
Heavier AI involvement in grant writing goes with proposals that sit nearer to what was funded the year before, and the pattern may be a warning for any organisation that picks its ideas with AI's help

In a large study of US federal research grants, proposals that showed more traces of AI writing tended to sit closer to the projects that the funding agency had backed in the previous year, and at one of the two agencies they were also more likely to be funded. The study appeared in PNAS on 11 August 2026, written by Yifan Qian, Zhe Wen, Alexander Furnas, Yue Bai, Erzhuo Shao and Dashun Wang of Northwestern University, and PsyPost reported it on 16 November. It points to a possible trade between winning the next round and proposing something new, which is a trade that any team writing strategy papers, pitches or product roadmaps with AI’s help may recognise, and the details matter because the evidence is strong in some places and thin in others.
What the researchers measured
The team combined confidential submissions from two large US research universities, 1,600 to the National Science Foundation (NSF) and 4,100 to the National Institutes of Health (NIH), covering funded, unfunded and pending proposals, with every publicly released NSF and NIH award over the same period (57,000 and 74,000), all with start dates from 2021 to 2025. They estimated AI involvement from the abstracts, by comparing word frequencies with a baseline of human-written abstracts from 2021 and AI-rewritten versions of the same abstracts, so the figure is an estimate of how many sentences carry AI-modified wording and not a record of who used which tool. That estimate was flat before late 2022 and rose sharply from 2023, and among grants starting between 2023 and 2025 it fell into two groups, one near zero and one centred on roughly 10 to 15% of sentences.
To measure distinctiveness, they calculated how far each abstract sat from every abstract that the same agency had funded in the previous year and converted that distance into a percentile within each agency and year. They then compared each scientist only with their own other proposals, which removes stable differences between people such as skill and writing style. Higher AI involvement was consistently associated with lower distinctiveness, and moving from the 25th to the 75th percentile of AI involvement corresponded to a drop of about 5 points in the distinctiveness percentile for NSF awards and about 4 points for NIH awards. When the authors ran 2021 human abstracts alongside AI-rewritten versions that kept the scientific ideas unchanged, the two sets scored almost the same, which they read as a sign that the convergence reflects where projects sit and not only how they are worded. They also found that proposals with more AI involvement were not systematically further from a scientist’s own earlier proposals, so the movement was towards the agency’s recent pattern and not away from the researcher’s established lines of work.
The outcomes differ by agency, and at NIH the same move from low to high AI involvement corresponded to roughly a 4-percentage-point rise in the chance of winning an award, and awards starting in 2023 and 2024 went on to produce about 5% more publications in their first one to two years, but the extra output was concentrated in papers that were not highly cited, and the link with top-5% papers was not statistically significant at conventional levels, and the same holds for the top 1% in the SI Appendix. At NSF the paper found no statistically significant link between AI involvement and either success or publication output. Neither the paper nor its SI Appendix gives the baseline funding rate for the university sample, so I cannot say what four percentage points means as a proportion.
What the study cannot show
The authors are plain about the limits and describe their results as conditional correlations within investigators and not causal effects, and they note that investigators may lean more on AI for proposals that are already incremental, easy to carry out or aligned with agency priorities, in which case the AI involvement would partly reflect the kind of project and not cause the shift towards the centre. The measure captures traces in the abstract and misses other ways of using AI in preparing a proposal, which PsyPost illustrates with brainstorming hypotheses or analysing preliminary data, the confidential data come from two universities, and the publication window covers only the first one to two years after the awards began, so it says little about long-run impact. The authors also offer an explanation for the NIH and NSF contrast, that NIH review may reward feasible, incremental projects that yield steady publications, and they call that account speculative.
What I take from it
My reading, which the study does not test, is that anyone who asks a tool to improve a proposal is also asking it to make the proposal resemble the writing it has seen, and the authors raise the same worry when they note that models trained on published papers and records of funded projects may tend to produce high-probability, field-typical language. The consequence for an organisation is that the options it considers may drift towards what has already worked, and that drift would feel like improvement from the inside, because each draft reads more fluently and more fundably than the last.
The NIH result makes this uncomfortable, because winning more often and publishing about 5% more is a reasonable outcome to want, and the missing piece is the highly cited work, where the study found no clear gain. A broader paper in Nature by Qianyue Hao and colleagues, covering 41.3 million natural science papers and AI tools in general and not chat assistants specifically, found a related tension between individual gain and collective narrowing, with scientists whose work was identified as AI-augmented publishing 3.02 times as many papers and receiving 4.84 times as many citations while the collective range of topics shrank by 4.63%. That is not the same pattern as the NIH result here, where the extra output was not in highly cited work, and that analysis is also observational, so the two studies describe associations from different angles and neither shows that the tools caused the narrowing.
Why organisations drift this way
James March’s work on exploration and exploitation describes a pattern in which organisations are pulled towards refining what they already do, because the returns from refinement are nearer and more predictable than the returns from experiment. The PNAS authors use the same pair of terms when they say that AI could shift the balance between exploration and exploitation in funding, and the grant data fit that pattern without testing it.
I wrote about a related loop in Dear Algorithm, where tracking changes the behaviour that the system then learns from: “The algorithm, in other words, doesn’t just observe behaviour. It influences it. And the behaviour it influences is the behaviour it then uses to make predictions.” If tools learn from what was funded and then shape what gets proposed, next year’s funded set will carry more of their influence, but the study covers only a few years and does not test any such loop, so I offer it as a question and not as a finding.
Using AI without letting it choose
The paper’s own explanation for the NIH and NSF contrast is institutional, which suggests that review norms and incentives matter, and I would expect the same inside a company, where planning and approval habits decide whether a drift towards safe ideas is noticed or rewarded. The practical advice below is mine and not a finding of the study. First, decide which idea you are pursuing before you open the tool, and use it afterwards to sharpen the argument, the structure and the language, because the authors’ test with AI-rewritten abstracts is consistent with the convergence lying in how ideas are positioned and not only in how they are worded. Second, ask the tool what a sceptical reader would find unfamiliar in the draft and then decide deliberately whether to keep that part, since the unfamiliar part may be what makes the idea worth pursuing.
Third, compare drafts with your own past winners and notice how close they sit, which you can do by reading last year’s approved plans next to this year’s and asking what is new. Fourth, protect a deliberate share of the portfolio for ideas that look unlike anything you funded before, and judge those on different terms, because a process that scores everything against past success will keep choosing the past. Using AI to polish an idea you already hold may cost little, although the study cannot say, because it measured AI traces in the wording of abstracts and cannot separate polishing from choosing, and handing the tool the choice of which idea to pursue risks giving the next round to last year’s winners.
David Chadderton is a behavioural strategist and Chief Marketing Officer. He writes The Unoptimised Human.
References
- Chadderton, D. (2026) Dear Algorithm, It’s Not Me, It’s You: Putting the Personal Back Into Personalisation. 2nd edn. 16 May 2026.
- Dolan, E. W. (2026) “Scientists who use AI propose less distinctive ideas for federal research grants”, PsyPost, 16 November. https://www.psypost.org/scientists-who-use-ai-propose-less-distinctive-ideas-for-federal-research-grants/
- Hao, Q., Xu, F., Li, Y. and Evans, J. (2026) “Artificial intelligence tools expand scientists’ impact but contract science’s focus”, Nature, 649, pp. 1237 to 1243, published 14 January. https://doi.org/10.1038/s41586-025-09922-y
- March, J. G. (1991) “Exploration and exploitation in organizational learning”, Organization Science, 2(1), pp. 71 to 87. https://doi.org/10.1287/orsc.2.1.71
- Qian, Y., Wen, Z., Furnas, A. C., Bai, Y., Shao, E. and Wang, D. (2026) “The rise of large language models and the direction and impact of US federal research funding”, PNAS, 123(33), e2601439123, published 11 August. https://doi.org/10.1073/pnas.2601439123
Eight minutes. Thirty-three questions.
A precise picture of how you operate: motivations, decision style, and blind spots. Free to start.
- S
- T
- A
- R
The canonical source.
Essays, when there is something worth sending. Choose the subjects you want. I will ask you to confirm the address before I write.