Auditing Bias and Safety in Voice AI Customer Care

On September 7, 2026, the research paper “Auditing Bias and Safety in Voice AI Customer Care” was published on arXiv cs.CL. This paper proposes a verification gate control framework for voice AI customer service systems. The framework divides the system architecture into three categories: native voice-to-voice, cascaded ASR to language model and TTS, and hybrid tool mediation. The study used matched service facts for verification under controlled caller presentation conditions, covering fact invariance before reasoning, performance cues, artifacts, and acoustic measurements, and recorded material results and service burden paths. The framework defines the scope of issues, methodology, seven verification gates, six indicator families, and declaration boundaries, and presents examples of refund dispute audits. Results from production systems were not included in this release due to limitations of the verification protocol.