Are LLMs shifting the balance between ‘exploration and exploitation’ in research funding?

The meteoric rise of large-language AI is quietly altering if not effectively upending scientific inquiry. It’s been doing so almost since the technology’s public debut in late 2022, and medical research has been anything but exempt. 

The change is perhaps most conspicuous in taxpayer-funded projects. 

The accelerated evolution comes through in a study looking at de facto receipts from the National Institutes of Health and the National Science Foundation. 

NIH, of course, supports exclusively medical research while NSF helps fund just about everything else. 

The study was conducted at Northwestern University and published Aug. 11 in PNAS, aka the Proceedings of the National Academy of Sciences. 

In the study report, Dashun Wang, PhD, Yifan Qian, PhD, and colleagues suggest their findings supply ample evidence that the fast uptake of LLMs across science is “already reshaping key stages of the U.S. federal research funding pipeline, from proposal submission to downstream scientific output.”

The ongoing development, they add, is affecting the ways in which scientific ideas are positioned, selected and translated into publicly funded research—and bringing “important implications for portfolio governance, research diversity and the long-run impact of science.”

AI gives itself away 

For the study, Wang and fellow researchers examined LLM involvement in funding proposals submitted to NIH and NSF by two large universities designated by the Carnegie organization as Research 1 institutions for maintaining “very high research spending and doctorate production.” 

The team looked at funded, unfunded and pending proposals as well as publicly released NSF and NIH awards. 

Ascertaining LLM involvement by several markers—foremost among them language distinctiveness—Wang et al. found LLM use rose sharply beginning in 2023. 

The researchers further observed that, troublingly, AI-assisted proposals tended to mimic already-funded research. 

The pattern suggests some proposal writers tend to prioritize the pursuit of hard dollars over the expansion of scientific understanding. 

“When used in proposal preparation, [LLMs] may enhance fluency and alignment with prevailing norms while reducing rhetorical and topical variance, making proposed projects more closely resemble those that have recently succeeded,” the researchers write. 

“In this way, LLM use could shift the balance between exploration and exploitation in scientific funding.” 

Predictable proposals vs. adventurous science 

In coverage of the project by Northwestern’s news division, Wang emphasizes that science advances by “exploring ideas that don’t yet look obvious.”

“If AI increasingly learns from yesterday’s successful proposals,” he warns, “one of the questions we should ask is whether tomorrow’s scientific portfolio becomes less adventurous.”

Meanwhile Qian underscores that federal research funding is the primary mechanism through which the United States converts public resources into scientific knowledge.

“Understanding forces that influence this federal funding process is essential not only for science policy and the rate and direction of scientific progress,” Qian says, “but also for the stewardship and accountability of public investment in research.”

Click here for the full, free study and here for the Northwestern news item.