LLMs More Susceptible to Spin Than Humans in Medical Literature

A study published on arXiv (2502.07963) reveals that Large Language Models (LLMs) are more susceptible to spin in medical literature abstracts than human readers. Spin refers to the presentation of study results in a more positive light than warranted, a known issue in medical publishing. The researchers evaluated 22 LLMs and found that they consistently interpreted trial results with greater bias toward positive framing. This is concerning because LLMs are increasingly used to synthesize medical evidence, potentially propagating spin into plain language summaries and influencing clinical decisions. The study highlights a critical vulnerability in AI-assisted medical research synthesis.