CITATION:
Health information for all: do large language models bridge or widen the digital divide?
BMJ 2024; 387 doi: https://doi.org/10.1136/bmj-2024-080208 (Published 11 October 2024)
Cite this as: BMJ 2024;387:e080208
'Key messages
Large language models (LLMs) like ChatGPT could have a role in narrowing the health information digital divide, democratising access to healthcare
But evidence indicates that LLMs might exacerbate the digital disparity in health information access in low and middle income countries
Most LLMs perform badly in low resources languages like Vietnamese, resulting in the dissemination of inaccurate health information and posing potential public health risks
Coordinated effort from policy makers, research funding agencies, big technology corporations, the research community, healthcare practitioners, and linguistically underrepresented communities is crucial to bridge the gap in AI language inclusivity'
SELECTED EXTRACTS
'Imagine asking a health information chatbot for advice on atrial fibrillation symptoms and receiving information on Parkinson’s disease—a completely unrelated condition. This is not a fictional scenario; it is what currently happens when you inquire about medical information in the Vietnamese language using OpenAI’s GPT-3.5 (through ChatGPT). This mix-up, far from a simple error, illustrates a critical problem with artificial intelligence (AI) driven healthcare communication in languages like Vietnamese...'
'The robustness of LLMs and the natural interactions they provide could transform digital health communication and education,23 revolutionising the delivery of medical information, especially in underserved regions...'
COMMENT (NPW): The authors make an important point: that LLMs work better in the main languages than in 'low resources languages'. There is clearly a divide here that needs to be addressed. However, the more important question is whether LLMs widen or reduce the overall divide in access to reliable healthcare information. Before LLMs, there was already a major divide in access to information for, say, English speakers versus Vietnamese speakers. The advent of LLMs - even with their current accuracy for less-spoken languages - still has the potential to massively increase access in those languages. Also, LLMs are new: we can expect them to get better and better over time, even withless-spoken languages. Furthermore, we need a better understanding of the kinds of inaccurate information generated by LLMs. If a Vietnamese speaker enters Parkinsons and gets information on atrial fibrillation, we need to know why they obtained this result, and we need to consider the consequences. In this case, the user would doubtless reject the result and no harm would be done. What is more harmful is accidental misinformation that is plausible. While LLMs are relatively unreliable in less-spoken languages, it is important for speakers of those languages to be aware of their shortcomings.
Best wishes, Neil
HIFA profile: Neil Pakenham-Walsh is coordinator of HIFA (Healthcare Information For All), a global health community that brings all stakeholders together around the shared goal of universal access to reliable healthcare information. HIFA has 20,000 members in 180 countries, interacting in four languages and representing all parts of the global evidence ecosystem. HIFA is administered by Global Healthcare Information Network, a UK-based nonprofit in official relations with the World Health Organization. Email: neil@hifa.org
