PNAS Study Warns: LLM Involvement in NIH Grant Applications Increases Funding Success but Decreases Research Novelty

August 30, 2026

Buttondown USA

Overview

A new PNAS study analyzing over 125,000 grant applications found that NIH proposals with heavy LLM involvement had a 4-percentage-point funding advantage, yet resulted in more incremental and less novel papers, raising concerns about AI narrowing the scientific idea landscape. Concurrently, a nonprofit founded by former OpenAI employees graded five major AI companies on safety practices, concluding that containment and oversight were insufficient, with Meta receiving an F and even Anthropic and OpenAI earning only C+.

In Depth

Key Findings

A study published in the Proceedings of the National Academy of Sciences (PNAS), analyzing over 125,000 grant applications, revealed that proposals to the U.S. National Institutes of Health (NIH) that extensively utilized large language models (LLMs) had a 4-percentage-point higher funding success rate. However, the study also indicated that research derived from these proposals tended to produce more incremental and less novel papers, raising concerns that AI could be narrowing the diversity of scientific ideas. This suggests that while AI may enhance the efficiency of the research process, it could inadvertently impede fundamental innovation.

Technical / Clinical Details

The research implies that LLMs are adept at optimizing proposal writing and structure to align with reviewer criteria, potentially enhancing the ability to articulate ideas efficiently within existing research themes and methodologies.
Nevertheless, the challenge of generating truly novel concepts or interdisciplinary insights became apparent, largely because AI’s generative capabilities heavily rely on existing knowledge patterns.
In related news, a nonprofit organization founded by former OpenAI employees evaluated five major AI companies on their safety practices, concluding that containment and oversight measures were insufficient. Meta received an ‘F’ grade, while even Anthropic and OpenAI managed only a ‘C+’, underscorining the significant challenges in safe AI development and deployment.
These evaluations reflect a growing concern about the safety and societal impact of AI technologies, even as their capabilities continue to advance.

Background & Context

In an increasingly competitive landscape for research funding, AI tools present an attractive means for researchers to boost the efficiency of paper writing and grant application preparation. However, the impact of widespread AI adoption on the quality and direction of research remains underexplored. The core of scientific progress lies in originality and novelty, making AI’s influence on these aspects a critical concern for the academic community.

Strategic Significance & Outlook

The PNAS findings prompt the research community to reconsider its approach to AI tools. Beyond mere efficiency gains, there’s a need to explore how AI and human creativity can optimally converge to foster genuinely innovative research. Furthermore, the low safety ratings for AI companies highlight the urgent need for industry-wide ethical guidelines, regulatory compliance, and more stringent governance over AI systems. Investors are likely to increasingly value companies demonstrating a strong commitment to ethical and responsible AI development, alongside their technological prowess.

Source: https://buttondown.com/pollak/archive/ai-intelligence-briefing-august-20-2026/

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.