The Rise of Large Language Models and the Direction and Impact of US Federal Research Funding

Abstract: Federal research funding shapes the direction, diversity, and impact of the US scientific enterprise. LLMs are rapidly diffusing into scientific practice, holding substantial promise while raising widespread concerns. Despite growing attention to AI use in scientific writing and evaluation, little is known about how the rise of LLMs is reshaping the public funding landscape. Here, we examine LLM involvement at key stages of the federal funding pipeline by combining two complementary data sources: confidential National Science Foundation (NSF) and National Institutes of Health (NIH) proposal submissions from two large US R1 universities, including funded, unfunded, and pending proposals, and the full population of publicly released NSF and NIH awards. We find that LLM use rises sharply beginning in 2023 and exhibits a bimodal distribution, indicating a clear split between minimal and substantive use. Across both private submissions and public awards, higher LLM involvement is consistently associated with lower semantic distinctiveness, positioning projects closer to recently funded work within the same agency. The consequences of this shift are agency-dependent. LLM use is positively associated with proposal success and higher subsequent publication output at NIH, whereas no comparable associations are observed at NSF. Notably, the productivity gains at NIH are concentrated in non-hit papers rather than the most highly cited work. Together, these findings provide large-scale evidence that the rise of LLMs is reshaping how scientific ideas are positioned, selected, and translated into publicly funded research, with implications for portfolio governance, research diversity, and the long-run impact of science.

• The paper shows that increased LLM involvement reduces semantic distinctiveness, with a 4–5 percentile drop measured via embedding distances.
• It employs a distributional quantification method comparing human-written and LLM-rewritten abstracts to analyze LLM diffusion in NSF and NIH proposals.
• The study finds that at NIH, higher LLM use boosts funding probability and publication output, whereas NSF outcomes remain largely unaffected.

The Rise of LLMs and the Direction and Impact of US Federal Research Funding

Overview and Motivation

This paper conducts a rigorous empirical analysis of the penetration and consequences of LLM adoption within the US federal research funding pipeline, specifically examining the National Science Foundation (NSF) and the National Institutes of Health (NIH) (2601.15485). The core objective is to characterize both the temporal trajectory and operational heterogeneity of LLM usage in proposal and award processes, and to ascertain how this diffusion modulates idea distinctiveness, proposal success, and research outputs. The study leverages an unprecedented combination of confidential submissions from two major R1 universities and comprehensive public data on NSF/NIH awards, enabling a uniquely granular view of both funded and unfunded research proposals.

Measurement of LLM Involvement

A distributional quantification method based on direct comparison of human-written and LLM-rewritten abstracts is deployed, utilizing GPT-3.5-turbo-0125. The resulting mixture parameter $\alpha$ at both corpus and individual-grant levels serves as an estimate for the degree of LLM involvement. This approach is robust to promotional language and controls for agency- and field-specific writing conventions.

Temporal Adoption Dynamics and Distribution Patterns

Analysis reveals a rapid inflection in LLM usage across both proposal submissions and funded awards starting in early 2023. There is a pronounced bimodal distribution at the individual grant level: a large subset of proposals exhibits negligible LLM involvement while a distinct segment demonstrates substantial use, typically clustering at $\alpha \in [10\%,\,15\%]$.

Figure 1: Rapid rise and bimodal distribution of LLM use in US federal research funding proposals and awards.

LLMs and Semantic Distinctiveness

Regression analyses incorporating fixed effects for investigator, field, and start year show that increased LLM involvement correlates consistently with diminished semantic distinctiveness, as operationalized by percentile-ranked SPECTER2 embedding distances to prior funded abstracts in the same agency and year. Across both NSF and NIH datasets, moving from low ($Q_1$) to high ($Q_3$) $\alpha$ constitutes a $\sim$4–5 percentile drop in distinctiveness. These findings persist when expanding the semantic baseline to a two-year funding window and when applying L2 metrics.

Figure 2: Higher LLM involvement is associated with reduced semantic distinctiveness in funded proposals.

Agency-Dependent Effects on Proposal Success

The study demonstrates distinct selection dynamics at NSF versus NIH. No significant association is observed between $\alpha$ and NSF funding probability. Contrastingly, higher $\alpha$ positively and significantly predicts NIH proposal success, indicating a $\sim$4 percentage-point increase in funding rate when moving from low to high LLM involvement, controlling for investigator and requested amount.

Figure 3: Positive effect of LLM use on NIH proposal funding probability, absent for NSF.

Impact on Downstream Research Outputs

Research output is quantified both as total linked publications and as field-normalized citation "hit" papers (top 5% and 1%). At NSF, LLM involvement does not influence publication volume or hit-paper counts. At NIH, there is a statistically significant increase in publication volume (\sim 5% for high $\alpha$), but no corresponding effect on hit papers. This demonstrates that LLM-induced productivity gains at NIH manifest as increased routine outputs rather than breakthrough work.

Figure 4: NIH grants with substantive LLM use generate more publications but not more high-impact work.

Field and Subagency Stratification

Full robustness checks with field and subagency stratification confirm the main results, with bimodal LLM use distributions and similar effects on semantic distinctiveness and outputs replicated across disciplinary and administrative subdivisions.

Writing Complexity and Promotional Language

Higher LLM involvement is systematically associated with increased syntactic and lexical complexity (Flesch Reading Ease metric), matching prior results in scientific preprints. Removal of promotional wording does not meaningfully alter estimates of $\alpha$, confirming that the detected LLM traces stem primarily from non-promotional text attributes.

Figure 5: Grant abstracts with high LLM use demonstrate increased writing complexity across agencies.

Figure 6: LLM-use estimation remains robust after excising promotional vocabulary.

Theoretical and Practical Implications

The paper provides compelling empirical support for the hypothesis that generative AI, when embedded into the research funding pipeline, preferentially facilitates the articulation of conventionally fundable projects, thereby dampening portfolio semantic diversity. At NIH, LLMs appear to advantage project selection and accelerate manuscript-level productivity, but do not enhance transformative output. This dichotomy implies that LLM-induced writing improvements primarily lower proposal articulation and manuscript preparation frictions, but do not circumvent operational bottlenecks in project execution or disrupt established norms for NIH-funded incremental research. NSF’s null results highlight the influence of agency-level evaluation criteria and project typology, given the prominence of non-paper outputs (software, datasets, instrumentation) and possibly higher risk tolerance.

Further, these results underscore salient governance challenges: LLMs may drive convergence towards safer, center-of-mass ideas, raising the risk of systemic exploitation bias. Policy responses—e.g., disclosure requirements and authorship clarification—must grapple with this duality: LLMs enhance individual efficiency but contract the variance necessary for high-impact discovery portfolios.

Future Directions

Empirically, extending the time horizon for hit-paper accrual and integrating non-public outputs (data, training) are critical next steps. Causally distinguishing the effects of LLMs on ideation versus narrative articulation demands direct intervention studies. Institutionally, agencies may consider reviewer retraining and adjustment of evaluation frameworks to counteract homogenization dynamics and sustain exploratory research.

Conclusion

This work presents a methodologically rigorous, large-scale assessment of LLM diffusion within US federal research funding, revealing a clear trade-off: widespread adoption improves aggregate proposal and publication productivity—especially at NIH—but systematically narrows idea distinctiveness and does not augment high-impact outcomes. These dynamics carry profound implications for the structure and efficacy of publicly funded scientific research and highlight the need for adaptive portfolio governance in the era of generative AI.