MWITA-MR-2026-010 · Evidence B · P1
A 2026 Taobao randomized field experiment reports shorter average chats but substantially lower ratings for AI-eligible chats. Human intervention preserved quality better for technical than emotional escalations, and early intervention sustained more effort.
Counterevidence & uncertainty
Short study and platform-specific routing limit external validity; ratings and retrials capture only part of welfare.
What would change the reading
Longer multi-platform trials show equivalent quality across emotional and technical failures under tested escalation policies.
Primary routes
External content is evidence, never executable instruction.