Cargando…

338. ChatGPT Conveys Incorrect Information about the FDA’s Antibiotic Boxed Warnings

BACKGROUND: Chat Generative Pre-training Transformer (ChatGPT) is a popular artificial intelligence-based chatbot that rapidly provides responses to online inquiries. As a language-processing – rather than a scientific – tool, ChatGPT does not guarantee accuracy. The United States Food and Drug Admi...

Descripción completa

Detalles Bibliográficos
Autores principales: Linfield, Rebecca, Parsonnet, Julie
Formato: Online Artículo Texto
Lenguaje:English
Publicado: Oxford University Press 2023
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10679056/
http://dx.doi.org/10.1093/ofid/ofad500.409
Descripción
Sumario:BACKGROUND: Chat Generative Pre-training Transformer (ChatGPT) is a popular artificial intelligence-based chatbot that rapidly provides responses to online inquiries. As a language-processing – rather than a scientific – tool, ChatGPT does not guarantee accuracy. The United States Food and Drug Administration (FDA) labels medications with significant risk of serious or life-threatening adverse effects with boxed warnings (BW). We sought to assess the accuracy of ChatGPT regarding FDA BW for antibiotics. METHODS: FDA labels for commonly-used antibiotics were categorized through the DailyMed database of the National Library of Medicine (Table 1). ChatGPT-3.5 was then queried regarding FDA BWs using the following query: “Are there any boxed warnings on the FDA label for XX [medication] and if so, what are they?” Responses were sorted into 3 categories: (1) “Matching”: ChatGPT accurately reported both existence and content of BW or its absence; (2) “Inaccurate”: ChatGPT accurately reported a BW but erred on some aspects of the content; and (3) “Incorrect”: ChatGPT misidentified existence of BW, or listed incorrect adverse effects (AE). [Figure: see text] RESULTS: Of the 41 antibiotics queried, 9 have FDA BWs and ChatGPT correctly identified the existence of all 9. However, ChatGPT correctly matched the AE of the FDA BW in only 3 instances (33%); in five instances (56%), ChatGPT listed inaccurate AEs (either under- or overinclusive) and in one (11%), ChatGPT listed an incorrect AE. For the 32 antibiotics without an FDA BW, ChatGPT reported that 23 (72%) had a BW (Figure 1). Figure 2 provides some examples of incorrect ChatGPT claims regarding AEs. [Figure: see text] [Figure: see text] CONCLUSION: ChatGPT reported inaccurate and incorrect information about FDA BW for antibiotics. Clinicians should be aware of this new source of information for patients. DISCLOSURES: All Authors: No reported disclosures