When Patients Cut In: Extending Clinical Conversational AI Safety to Interruptions

Clinical voice agents fail dramatically when patients interrupt during information provision tasks: When patients cut off the agent mid-sentence during FAQ scenarios, all four tested AI models failed to preserve required clinical content 100% of the time in competitive interruptions, compared to a baseline of near-zero failures in non-interrupted scenarios.

Current AI safety benchmarks don't reflect real clinical environments: Existing evaluation standards for clinical conversational AI assume patients wait politely for the agent to finish speaking, but real patients interrupt frequently—meaning these benchmarks miss critical failure modes that occur in actual deployed systems.

Different interruption types require different solutions: The researchers identified three types of interruptions (recognitional, competitive, and transitional), and found that AI models' performance varied significantly across these categories, suggesting that one-size-fits-all safety improvements won't work.

Simple fixes like apology phrases have unpredictable effects: Adding a brief apology marker ("sorry to interrupt") improved recovery by tens of percentage points for some models but actually reduced recovery for one model, highlighting that even straightforward interventions can have inconsistent and sometimes counterproductive outcomes across different AI systems.

Cascaded AI architectures create hidden vulnerabilities in clinical settings: The speech-to-text → language model → text-to-speech pipeline means that when interruptions occur, clinically necessary information can be permanently lost even if the language model itself handles the interruption reasonably well.