Scientists Lag Behind Latest AI Models, Study Finds

A large-scale examination of how scientists have incorporated AI foundation models into their research uncovers a striking pattern: the academic community is consistently choosing smaller models than what developers are producing, and this gap is widening.

The Adoption Gap#

Researchers from across disciplines have increasingly turned to foundation models in recent years, with adoption rates climbing at nearly exponential speeds. However, the models they select tend to be substantially smaller than the cutting-edge systems becoming available. In 2015, the typical model used in scientific papers was 5.4 times larger than the average newly developed model — a sign that research was then using mature, scaled-up versions of tools. By 2024, the dynamics had inverted completely: newly created models were 6.9 times larger than those appearing in scientific work. This growing divergence suggests researchers are consistently trailing the technological frontier.

Where Adoption Is Strongest#

The uptake of foundation models varies significantly across fields. Linguistics, Computer Science, and Engineering have embraced these tools most readily, while adoption lags in other disciplines. Vision models currently dominate scientific use, accounting for the largest share of deployments. However, language models are rapidly gaining ground, expanding their presence across papers at an accelerating pace.

Open-weight models — those whose internal parameters are publicly available — have become the overwhelmingly preferred choice in the research community, suggesting scientists value transparency and the ability to customize and audit the tools they work with.

Potential Consequences#

The analysis offers preliminary evidence that scientists' reluctance to adopt larger models may exact a cost. Research that employs larger foundation models tends to achieve stronger outcomes by conventional measures: these papers are published in higher-impact journals and receive more citations. The findings raise questions about whether researchers' conservative approach to model selection may be holding back the scientific benefits AI could deliver.