Please stay seated – the specialist AI will see you shortly

Abstract

The adage "garbage in, garbage out" is an important reminder in the era of artificial intelligence (AI), particularly with increasing adoption in healthcare.General-purpose large language models (LLMs) such as ChatGPT, Bard, and Claude have gained popularity among clinicians due to their ability to convert prompts into meaningful, human-like responses within seconds.However, their reliance on training from a sea of unvetted internet-based data spanning journals, news articles, blogs, and social media platforms means that their outputs should be approached with caution for clinical matters.Developing a specialized LLM requires curation of high-quality, clinically relevant data.The evidence base is polluted with potential garbage in the form of flawed methodology, invalid abstracts, and even conflicting guidelines, and without expert filtering, this noise of information risks being encoded into AI outputs.Specialized development allows for expert-guided curation, emphasizing high-quality evidence from reputable sources such as guidelines and high-quality studies, including paywalled content inaccessible to general LLMs.In this issue, Simsek and colleagues present data on their gastroenterology-specific LLM termed GastroGPT, compiled using 1.2 million tokens from gastroenterology journals, guidelines, and textbooks.In a blinded comparison across 10 simulated gastroenterological scenarios, GastroGPT significantly outperformed general LLMs of ChatGPT 4, Bard, and Claude, with an overall expert rating of 8.1 out of 10 vs 5.2 to 7.0 for general LLMs.GastroGPT was particularly superior for tasks

Ähnliche Arbeiten