A new study using Gemini 3.5 Flash found that large language models provide different medical triage recommendations based on the language of the patient's prompt, even when symptoms are identical. The model recommended emergency room visits at significantly varying rates across six languages, with English and Arabic prompts leading to higher ER recommendations than Japanese or Hindi. Adding location information, such as specifying a US location, dramatically increased ER recommendations for non-English prompts, highlighting a bias in the model's implicit geographic inference. AI
IMPACT Reveals potential for LLMs to perpetuate health disparities due to language-based geographic biases.
RANK_REASON Academic paper detailing experimental findings on LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →