Too Human or Not Enough? User Preferences for AI Error Recovery Messages
TL;DR
It is suggested that effective AI error recovery should maintain a professional baseline while adapting to the seriousness of the mistake, and value directness, clarity, accountability, and restraint over humor or highly human-like expression.
Abstract
Mistakes by AI assistants (AIAs) are inevitable, yet how these systems should communicate with users after an error occurs remains an open question. This paper examines post-error communication in AIAs, focusing specifically on the tone of voice used in recovery messages. We conducted an online survey with 50 participants, who evaluated recovery messages across three scenarios of increasing severity: a failed file conversion, an incorrect meeting time, and a wrong hotel booking. Each scenario included five tone variants: formal, casual, empathetic, neutral, and fun. Participants also explained their choices and described what makes recovery messages effective or frustrating. Results show a consistent preference for formal recovery messages across all severity levels. Across the three scenarios, participants valued directness, clarity, accountability, and restraint over humor or highly human-like expression. These findings suggest that effective AI error recovery should maintain a professional baseline while adapting to the seriousness of the mistake.
