Real-Time Voice AI Hears but Does Not Listen

The finding isn't simply that voice AI ignores emotion: it's that these systems demonstrably possess the relevant perceptual capacity and still fail to route it into decision logic, which means the problem is architectural prioritization, not missing capability. That distinction matters enormously for how engineers would actually fix it.

This connects loosely to the RevengeBench work published the same day (arXiv cs.LG, 2026-06-24), which frames the challenge of reconstructing hidden decision logic from observed behavior. That paper treats opaque decision-making as an inverse problem worth solving from the outside. The voice AI paper essentially confirms why that kind of external probing is necessary: even when you can query a system directly and get accurate self-reports about what it perceives, those perceptions may not be wired into the actions the system actually takes. The gap between what a model 'knows' and what it 'does' is a recurring structural problem across both papers, even if their domains differ.

Watch whether OpenAI or Google publish updated safety documentation for GPT Realtime 2 or Gemini Flash Live within the next two quarters that specifically addresses paralinguistic signal weighting in action-triggering pipelines. Silence from both would suggest the architectural fix is harder than a policy update.