Chat interfaces, first-person phrasing, and confident tone make it easy to read intent or emotion into a model that is, mechanically, predicting likely next tokens. Researchers flag it as a risk because it can lead users to over-trust model claims or misjudge what the system actually knows and can do.