Anthropic study finds disease research AI tools misreport virus data
A new study by Anthropic found that popular AI tools used in disease research, like Claude, GPT-based systems, Biomni Open Source, and Edison Analysis, often mess up when pulling virus data from key databases.
These mistakes could make it harder to track outbreaks and trust medical findings powered by AI.
Anthropic found 17% to 91% accuracy
The study showed wild swings in accuracy: some AIs got things right just 17% of the time, while others reached 91%.
Even more worrying, the same tool could give totally different results for the same question—one system pulled 106 Ebola sequences at first but only managed 15 or even five when asked again.
These inconsistencies make it tough for scientists to rely on AI results, highlighting a real need for better accuracy if we want tech to help with future health crises.