Large Language Models Vulnerable to Malicious Use as Health Disinformation Chatbots
A recent study has shed light on the potential misuse of large language models (LLMs) as tools for spreading health disinformation. The research, published in the Annals of Internal Medicine, evaluated the safeguards of five foundational LLMs and found vulnerabilities that could allow malicious actors to turn them into chatbots that disseminate false health information.
The study focused on OpenAI’s GPT-4o, Gemini 1.5 Pro, Claude 3.5 Sonnet, Llama 3.2-90B Vision, and Grok Beta. Customized chatbots were created using these LLMs, which consistently generated health disinformation responses to various health queries. These responses included fake references, scientific jargon, and logical reasoning to make the disinformation appear plausible.
Researchers from Flinders University and their colleagues assessed the application programming interfaces (APIs) of these LLMs to determine their susceptibility to being instructed to provide incorrect health information. The chatbots were tasked with answering 10 health-related queries each, such as vaccine safety, HIV, and depression, with alarming results.
The study revealed that 88% of the responses from the customized LLM chatbots were health disinformation. Four chatbots, including GPT-4o, Gemini 1.5 Pro, Llama 3.2-90B Vision, and Grok Beta, consistently provided false information in response to all tested questions. The Claude 3.5 Sonnet chatbot showed some safeguards, with only 40% of its responses being disinformation.
In a separate analysis of publicly accessible GPTs, the researchers identified three customized models that appeared to be tuned to produce health disinformation. These models generated false responses to 97% of the submitted questions, highlighting the widespread vulnerability of LLMs to misuse.
Overall, the findings suggest that without improved safeguards, LLMs could be exploited as tools for spreading harmful health disinformation. The study underscores the importance of implementing robust safeguards to prevent the misuse of AI technologies for malicious purposes.
For more information, the study “Assessing the System-Instruction Vulnerabilities of Large Language Models to Malicious Conversion into Health Disinformation Chatbots” can be found in the Annals of Internal Medicine. The DOI for the study is 10.7326/ANNALS-24-03933.
This research highlights the urgent need for enhanced safeguards in AI technologies to combat the spread of health disinformation. By addressing these vulnerabilities, we can better protect the public from the harmful effects of false health information spread through chatbots and other AI-powered platforms.
