Researchers have uncovered startling insights into the performance of ChatGPT, a leading large language model (LLM), when it handles health-related queries. A pioneering study by Australia’s CSIRO and The University of Queensland highlights a counterintuitive finding: ChatGPT’s reliability dwindles with the infusion of additional evidence, plummeting to accuracy rates as low as 28%.
The investigation, led by CSIRO’s Principal Research Scientist and Associate Professor at UQ, Dr Bevan Koopman, traversed the digital landscape to understand how LLMs like ChatGPT fare when dispensing health advice. The study’s design mimicked the inquiries of an average, non-professional health consumer, examining the model’s performance on a variety of health-related questions. These ranged from the effectiveness of zinc against the common cold to the veracity of home remedies like drinking vinegar to dislodge a fish bone.
In an era where digital platforms are increasingly turned to for medical advice, the results of this study carry significant implications. The research unveiled that ChatGPT, when presented with straightforward questions devoid of additional evidence, demonstrated a commendable 80% accuracy level. However, the introduction of supporting or contradicting evidence caused a notable decline in its precision, dipping to 63% and further plummeting when the model expressed uncertainty in its responses.
This phenomenon, as Dr Koopman points out, defies the conventional wisdom that supplementary evidence should bolster the accuracy of responses. The paradox may lie in the possibility that additional information, regardless of its validity, could introduce “noise” that undermines the model’s performance.
The implications of these findings are vast, especially considering the rapid adoption of LLMs like ChatGPT, launched in late 2022 and swiftly becoming a staple in digital information retrieval. With LLMs being integrated into major search engines through Retrieval Augmented Generation, understanding the nuances of their interaction with search components becomes crucial, as highlighted by study co-author Professor Guido Zuccon of UQ.
As LLMs continue to shape the digital information ecosystem, this study underscores the need for ongoing research to demystify the dynamics of LLM-generated health information and to refine the accuracy of these digital advisors. The next phase of the research aims to delve into the public’s use of health information generated by LLMs, signaling a critical journey towards optimising the reliability of digital health advisories in an increasingly AI-driven world.
Support independent community journalism. Support The Indian Sun.
Follow The Indian Sun on Twitter | Instagram | Facebook
Research from Australia's CSIRO & The University of Queensland reveals #ChatGPT's declining accuracy in #health advice with additional evidence, challenging assumptions. Implications are vast as digital platforms rely on AI. 🧠💊🔍 #TheIndianSunhttps://t.co/lQKwrHYHEi
— The Indian Sun (@The_Indian_Sun) April 4, 2024
Donate To The Indian Sun
Dear Reader,The Indian Sun is an independent organisation committed to community journalism. We have, through the years, been able to reach a wide audience especially with the growth of social media, where we also have a strong presence. With platforms such as YouTube videos, we have been able to engage in different forms of storytelling. However, the past few years, like many media organisations around the world, it has not been an easy path. We have a greater challenge. We believe community journalism is very important for a multicultural country like Australia. We’re not able to do everything, but we aim for some of the most interesting stories and journalism of quality. We call upon readers like you to support us and make any contribution. Do make a DONATION NOW so we can continue with the volume and quality journalism that we are able to practice.
Thank you for your support.
Best wishes,
Team The Indian Sun
