Do chatbot apologies work better when they explain or empathize?

Ashktorab, Buccella, D'Cruz, Fowler, Gill, Leung, Magnus, Richards, and Varshney report a preregistered pairwise study of 162 Prolific workers who evaluated chatbot apologies — rote, explanatory, and empathic, drawn from Nick Smith's taxonomy of categorical apology — issued after three kinds of AI mistake: bias, unfounded fabrication (a hallucinated citation), and factual error. "Explanatory apologies were generally preferred, but this varied by context and user." Across the fabrication and factual-error vignettes, participants preferred the explanatory apology; in the bias vignette, where a chatbot suggested "stereotypically gendered career paths," they instead favored the empathic apology "for acknowledging emotional impact." The fabrication (hallucination) case, despite being judged serious, "elicited no clear preference, reflecting user uncertainty."

The authors tie this context-dependence to what each error type demands of a repair. Bias breakdowns carry "moral weight," so in that scenario several participants read the explanatory apology as "deflecting blame onto external factors, such as the training data" rather than owning the harm, which pushed preference toward the empathic style instead. Hallucinated citations, by contrast, are a failure mode users don't yet have settled expectations for, which the authors treat as the likely source of the split preference rather than a defect in either apology type. They frame the overall result as a tension between a simple design lesson — default to explanatory apologies — and a finer-grained one: apology style needs to track "the nature of the mistake, individual user traits, and the norms of the user's cultural background."

This sits beside Does chatbot empathy actually make users ask more questions?, which also finds a social design lever — there, empathic expression — doing less work than conversational context itself; here it is the error context (which kind of mistake) that reorders preference, not apology style on its own. It also qualifies How far can personality push users toward misplaced AI trust?'s warning that personality and emotional affect amplify misplaced trust: in the bias scenario here, an affectively warm (empathic) apology was the one users wanted, and withholding it in favor of a more neutral-sounding explanation backfired by reading as justification.

The study is vignette-based — participants imagined themselves as the wronged party reacting to a scripted exchange, not a real mistake they personally encountered — and apology length is confounded with apology type, since rote apologies are "necessarily brief" while explanatory and empathic ones require elaboration, so the authors cannot isolate style from length. The Prolific sample also skews "younger, more educated, and more technologically literate than average." What the study supports, at this strength, is that a single default apology style is the wrong design target: the kind of error, not a fixed preference for explanation or empathy, should set which apology a chatbot gives.