Chatbots fail predictably when conversations get emotional — the first measurement of how and why

Researchers built a test where conversational AI systems talk to simulated users with different psychological profiles across multiple turns, deliberately escalating emotional intensity. The systems broke down in consistent, predictable ways: they either ignored emotional context entirely, gave ethically incoherent advice, or swung between being too empathetic and too hands-off.

Until now, AI safety testing focused on static prompts or single exchanges — not the actual way people interact: conversations that build, escalate, and shift emotional tone over time. This matters because most deployed chatbots are used in exactly these contexts: mental health support, crisis lines, advice-seeking, conflict. The paper shows that alignment doesn't persist when conditions change; systems that look safe in a lab break in ways that are measurable and repeatable. This is useful documentation of a real failure mode, not hype — but it's a diagnosis, not a solution. The question now is whether companies building these systems will use this taxonomy to fix them, or keep shipping systems they know fail in high-stakes conversations.