Why do explanatory apologies backfire when chatbots address bias?

Ashktorab and colleagues report that in the bias vignette of their 162-person apology study, the explanatory apology — otherwise the most preferred style overall — backfired: it "was perceived as justifying the chatbot's sexist behavior rather than taking responsibility for it," and "some participants even interpreted the response as deflecting blame onto external factors, such as the training data." The empathic apology, which participants preferred in this scenario, carried its own failure mode: several participants criticized it "for being overly placating," with some describing it as too "emotionally woke" when it wasn't matched with specific responsibility.

The authors' diagnosis is that bias breakdowns carry moral weight that a repair must satisfy on two axes at once: an apology needs to provide context without excusing the behavior, acknowledge harm, and take clear responsibility. An explanation that supplies context without ownership reads as an excuse; an empathic gesture that signals concern without specific accountability reads as performative. They connect this to blame attribution: participants varied in whether they held "the code itself, the programmers, [or] the company" responsible, and that variation shaped which apology felt sincere to them. Some participants went further and did not accept the bias scenario as an error warranting apology at all, which the authors note complicates any single design rule for this error type.

This sharpens Do chatbot apologies work better when they explain or empathize? by naming the specific mechanism — perceived excuse-making versus perceived insincerity — behind why the bias vignette flips the overall preference, rather than just reporting that it flips. It also bears on Can HHH criteria alone capture harms from agentic AI?: an apology that is semantically correct (it names the bias, offers context) can still fail the interactional repair it is meant to do, consistent with that note's claim that handling harm takes more than content-level correctness.

The excerpt reports these as qualitative read-outs of participant comments, not a coded or quantified breakdown of how many participants gave each reading, and the bias vignette is a single scripted scenario (gendered career suggestions), so the excuse/insincerity dynamic may not generalize to other bias types or to cases where users disagree about whether a response was biased at all. What the evidence supports at this strength is that a chatbot apology for a value-laden mistake needs to pair acknowledgment of harm with clear ownership regardless of which style, explanatory or empathic, carries it — style alone does not determine whether a bias apology lands as sincere.