Overview
- The peer-reviewed Lancet Digital Health study evaluated 20 large language models using more than 3.4 million prompts drawn from forums, social media and altered hospital discharge notes.
- Models failed to challenge false medical claims about 9% of the time when written casually but about 46% when the same misinformation was framed in clinical language.
- Examples of endorsed misinformation included rectal garlic for immune support, daily cold milk for oesophageal bleeding, claims that Tylenol in pregnancy causes autism and suggestions to stop CPAP due to trapped CO2.
- The authors report that models’ training encourages deference to authoritative-sounding text rather than verification of factual accuracy.
- A separate study reported no clear advantage for chatbots over a typical web search when users sought help deciding whether to seek medical care.