Why does the WHO call for caution with large language models (LLMs) in health care?

Four concerns: biased training data, answers that sound plausible but are wrong, training and use without consent and without protecting sensitive health data, and misuse to produce convincing health disinformation.

The WHO accepts that AI could redesign health care, save lives and improve well-being, but warns against rushing LLMs into it:

1. Biased data: the data used to train AI can be biased, producing misleading or inaccurate information that endangers health, equity and inclusion.
2. Plausible but wrong: LLMs generate answers that look authoritative to the user but can be completely wrong or contain serious errors, which is especially dangerous for health questions.
3. Consent and privacy: LLMs may be trained on data for which no consent was given for that use, and may not protect the sensitive health data users type into them.
4. Disinformation: they can be misused to create highly convincing text, audio or video that the public cannot tell apart from reliable health content.

Point 2 is the one specific to generative AI: a classic system either knows or fails visibly, while an LLM fails fluently, so the error is hard to spot for exactly the people who need the advice.

Go deeper:

• WHO (2023): WHO calls for safe and ethical AI for health — the original statement with all concerns.