Work

Voice-first generative artificial intelligence in healthcare requires end-to-end patient-safety validation

Year: 2026
Type: article
Abstract: Voice-first generative artificial intelligence (AI) may improve access to health information but introduces risks between what patients say, what systems interpret, and what patients understand and do. This Perspective proposes an end-to-end, task- and risk-specific approach to safety evaluation. Six interconnected domains address recognition of clinically consequential speech, preservation of intended meaning, local clinical grounding, risk recognition and feasible human escalation, auditory comprehension and ability to act, and cross-cutting governance. We explain the framework's conceptual derivation and distinguish it from existing conversational-model and voice-AI approaches. Rather than proposing universal numerical thresholds, we outline a decision procedure that prespecifies intended use, clinician-adjudicated critical failures, domain-specific evidence and non-compensatory safety gates. A hypothetical Somali-language postoperative follow-up case illustrates how a missed red flag could prevent deployment despite favorable average performance. This framework is an untested conceptual proposal, not a validated standard: its domain structure, feasibility, thresholds, and effects on patient outcomes require prospective investigation with intended users, clinicians and health systems.
Institution: SIMAD University
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $3,077
OpenAlex ID: W7234031247
FWCI:
Cites: 22
Cited by:
Related to: 10