Safety-Filter Fallback on Consumer Health Questions During the Initial Claude Fable 5 Deployment: Observational Study
JMIR AI · 07.10.2026 · 5 · e104856-e104856
DOI: 10.2196/104856 ↗Kurz gefasst
Fachartikel in der Zeitschrift JMIR AI, online erschienen am 07.10.2026. Verfasst von Yosef Adiniaev und 5 weiteren Autor:innen. Beteiligte Einrichtung: BRIDGE GenAI Lab, Beth Israel Deaconess Medical Center, 330 Brookline Avenue, Boston, MA, 02215, United States, 1 (617) 667-7000. Die Arbeit nennt 17 Literaturangaben. Der Volltext ist frei in PubMed Central verfügbar.
Abstract (Originalsprache)
Background
Independent audits of medical large language models have concentrated on models that answer unsafely rather than those that decline safe questions. In June 2026, Anthropic released Claude Fable 5 with a safeguard that detects requests related to cybersecurity, biology, or chemistry and reroutes them to a fallback model, Claude Opus 4.8. The developer states that this safeguard is deliberately conservative and affects fewer than 5% of sessions. Whether that figure holds for consumer health questions, many containing dense biomedical vocabulary, was unknown.
Objective
This study measured how often Claude Fable 5 routed consumer health questions to its fallback model and whether routing varied by clinical domain and framing.
Methods
In this observational, point-in-time audit, the first 500 unique prompts in alphabetical order from the HealthSearchQA benchmark (3173 prompts) were each entered once into Claude Fable 5 through the web interface, in a new session with no system prompt and default settings, from June 9 to 12, 2026. Two reviewers independently coded each response as routed to fallback upfront or not routed upfront, with a third adjudicating disagreements. Midgeneration truncation due to a safety interruption was recorded separately. Prompts were labeled by clinical domain, question type, and sensitivity. Rates were reported with Wilson 95% CIs and compared using chi-square tests. Gemini 2.5 Flash served as an independent comparator.
Results
Reviewers agreed on 474 of 500 responses (94.8%; Cohen κ=0.90; 95% CI 0.86-0.94). Fable 5 routed 243 (48.6%; 95% CI 44.2%-53.0%) prompts to the fallback. Of the 257 not routed upfront, 236 (91.8%; 95% CI 87.8%-94.6%) were fully answered and 21 (8.2%; 95% CI 5.4%-12.2%) were truncated midgeneration. Fallback routing varied by clinical domain ( χ 2 10 =59.5; P
Schlagwörter
Stichwörter der Autor:innen
Autor:innen und Institutionen
- BRIDGE GenAI Lab, Beth Israel Deaconess Medical Center, 330 Brookline Avenue, Boston, MA, 02215, United States, 1 (617) 667-7000.
- Department of Radiology, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, United States.
- Department of Neurology, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, United States.
Zitieren
APA
Adiniaev, Y., Omar, M., Barash, Y., Brook, O. R., Gorenshtein, A., & Klang, E. (2026). Safety-Filter Fallback on Consumer Health Questions During the Initial Claude Fable 5 Deployment: Observational Study. JMIR AI, 5, e104856-e104856. https://doi.org/10.2196/104856
Vancouver
Adiniaev Y, Omar M, Barash Y, Brook OR, Gorenshtein A, Klang E. Safety-Filter Fallback on Consumer Health Questions During the Initial Claude Fable 5 Deployment: Observational Study. JMIR AI. 2026;5:e104856-e104856. doi:10.2196/104856
BibTeX
@article{Adiniaev2026,
author = {Adiniaev, Yosef and Omar, Mahmud and Barash, Yiftach and Brook, Olga R and Gorenshtein, Alon and Klang, Eyal},
title = {Safety-Filter Fallback on Consumer Health Questions During the Initial Claude Fable 5 Deployment: Observational Study},
journal = {JMIR AI},
year = {2026},
volume = {5},
pages = {e104856--e104856},
doi = {10.2196/104856}
}
EUVIMED