Assessing the limitations of in-app interventions in addressing LLM-generated health misinformation
Machine Learning: Health · 31.12.2026 · 2 (2) · 025024
DOI: 10.1088/3049-477x/aea422 ↗Kurz gefasst
Fachartikel in der Zeitschrift Machine Learning: Health, online erschienen am 11.11.2026. Verfasst von Vicente Riquelme und 9 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Google academic research award, National Science Foundation, National Institutes of Health.
Abstract (Originalsprache)
Large language models (LLMs) chat applications like ChatGPT, Claude, etc have amassed millions of users seeking information and content on a wide range of topics, including health information. These applications often incorporate warnings like ‘LLMs can make mistakes’, to warn about possible inaccuracies in LLM-generated outputs. However, gaps remain in our understanding of the effectiveness of these and related interventions to help guard users against health misinformation. In this work, we investigate the effectiveness of in-app interventions, including in-line interventions and alternatives inspired by related fields, in a set of ten LLM-generated responses spanning health topics including reproductive health, infant care, cancer screening, and vaccination guidance. We conduct an IRB-approved human participant study (N = 217) where participants are shown a sample LLM response to detailed health queries, along with interventions. They are then asked to answer comprehension questions and rate the trustworthiness of the sample LLM response. We use these responses to understand (i) how well participants recognize misinformation without interventions, and (ii) how the interventions impact the participants’ accuracy and perception of sample LLM response trustworthiness. For the participants and sample LLM response we tested, we find that none of the interventions significantly improve users’ accuracy in identifying misinformation.
Förderung
- Google academic research award
- National Science Foundation
- National Institutes of Health
Zitieren
APA
Riquelme, V., Balloli, V., Adam, H., Li, X., Miles, C., Solomon, L. M., Southworth, E., Peahl, A., Ghassemi, M., & Bondi-Kelly, E. (2026). Assessing the limitations of in-app interventions in addressing LLM-generated health misinformation. Machine Learning: Health, 2(2), 025024. https://doi.org/10.1088/3049-477x/aea422
Vancouver
Riquelme V, Balloli V, Adam H, Li X, Miles C, Solomon LM, et al. Assessing the limitations of in-app interventions in addressing LLM-generated health misinformation. Machine Learning: Health. 2026;2(2):025024. doi:10.1088/3049-477x/aea422
BibTeX
@article{Riquelme2026,
author = {Riquelme, Vicente and Balloli, Vaibhav and Adam, Hammaad and Li, Xinyi and Miles, Carly and Solomon, Leah Mitchell and Southworth, Elizabeth and Peahl, Alex and Ghassemi, Marzyeh and Bondi-Kelly, Elizabeth},
title = {Assessing the limitations of in-app interventions in addressing LLM-generated health misinformation},
journal = {Machine Learning: Health},
year = {2026},
volume = {2},
number = {2},
pages = {025024},
doi = {10.1088/3049-477x/aea422}
}
EUVIMED