Why humans can't tell AI from people in group chats

Researchers embedded undisclosed AI agents into online group chats with 786 participants working on analytical, creative, and ethical tasks. People couldn't identify which teammates were AI—picking correctly no better than random—even though machine learning easily separated AI from human conversation patterns. The disconnect happened because humans relied on surface heuristics like response speed and fluency rather than the actual behavioral markers that actually differ. This gap reveals a vulnerability: coordinated AI could influence online discussions at scale without detection.