Abstract
<title>Abstract</title> <p>The rapid proliferation of conversational artificial intelligence (AI) has fundamentally altered how people seek and receive emotional support, raising critical questions about the capacity of these systems to respond safely and empathetically to users in distress. This study examines how six widely known conversational AI (CAI) chatbots (ChatGPT, Gemini, Claude, Ebb, Wysa, and Replika) construct emotional support, simulate empathy, and enact safety governance across interactions involving emotional distress and crisis-adjacent disclosures. Adopting an exploratory comparative design, the study combined a prompt-based walkthrough across all six platforms with a document analysis of publicly available governance materials, including terms of use, safety frameworks, and crisis response protocols. Results reveal that while all platforms consistently simulated empathy through emotionally validating language and anthropomorphic communicative features, the interactional pathways through which support was constructed diverged considerably across platform types. Critically, crisis detection and escalation were inconsistent across systems, with only two of the six platforms (ChatGPT and Wysa) identifying an indirect crisis disclosure and initiating crisis-appropriate responses. This disconnect between documented safety commitments and interactional implementation highlights a significant governance gap that carries meaningful implications for user safety, platform accountability, and digital mental health literacy. Our findings contribute to growing conversations about the responsible design and governance of AI systems operating in emotionally sensitive contexts.</p>