Who's Sorry Now: User Preferences Among Rote, Empathic, and Explanatory Apologies from LLM Chatbots

Source: Zahra Ashktorab, Alessandra Buccella, Jason D'Cruz, Zoë Fowler, Andrew Gill, Kei Yan Leung, P. D. Magnus, John Richards, Kush R. Varshney; ACM Transactions on Computer-Human Interaction 33(3), Article 26 · 2026-06

As chatbots driven by large language models (LLMs) are increasingly deployed in everyday contexts, their ability to recover from errors through effective apologies is critical to maintaining user trust and satisfaction. In a preregistered study with Prolific workers (N = 162), we examine user preferences for three types of apologies (rote, explanatory, and empathic) issued in response to three categories of common LLM mistakes (bias, unfounded fabrication, and factual errors). We designed a pairwise experiment in which participants evaluated chatbot responses consisting of an initial error, a subsequent apology, and a resolution. Explanatory apologies were generally preferred, but this varied by context and user. In the bias scenario, empathic apologies were favored for acknowledging emotional impact, while hallucinations, though seen as serious, elicited no clear preference, reflecting user uncertainty. Our findings show the complexity of effective apology in AI systems. We discuss key insights such as personalization and calibration that future systems must navigate to meaningfully repair trust.

Introduction. Apologies serve as a social technology for repairing and maintaining relationships. For minor offenses, they act as a preventative measure, defusing potential grievances before they escalate. For more serious transgressions, apologies can ease resentment and create space for forgiveness and reconciliation. Though composed of mere words and gestures, apologies wield remarkable emotional and moral power. Accordingly, apologies have long been a topic of interest in human–computer interaction (HCI), especially in systems that aim to support social repair, trust, and emotional engagement. To apologize is to acknowledge harm, express responsibility, and affirm the dignity of the person harmed. Accepting an apology, in turn, signals recognition of the apologizer’s remorse and commitment to future change. These dynamics make apology a powerful tool, the effects of which depend on empathy, sincerity, and perceived moral agency. Given the centrality of apology in human interaction, it is unsurprising that large language models (LLMs), trained on vast corpora of human-produced text, are adept at generating apologies. Chatbots such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini frequently offer polished and seemingly heartfelt apologies, whether for refusing requests beyond their guardrails or for providing inaccurate information. While humans often struggle to say sorry, chatbots do not (Figure 1). These AI-driven apologies raise new questions: Can systems that lack emotional states and moral agency offer meaningful repair? Do chatbot apologies increase user trust or willingness to continue the interaction? These questions take on a new urgency as AI systems increasingly participate in socially sensitive contexts. With the emergence of LLM-driven chatbots, new types of mistakes have emerged that warrant repair through an apology. To study how users interpret apologies, we differentiate between three error contexts: factual errors, unfounded fabrications, and biased responses from the LLM distorted information categories defined by Sun et al.. Here, factual errors are defined as incorrect outputs that retain a plausible causal mechanism (such as reasoning mistakes or misapplications of information), and unfounded fabrications as content generated without grounding in the prompt or external facts. Though prior work has grouped factual errors and fabrications under the broad category of LLM hallucinations, we distinguish these error forms because users may understand and evaluate them differently: factual errors may resemble human-like mistakes, while fabrications may come across as ungrounded statements that feel harder to rationalize. For such mistakes, LLM-driven chatbots routinely offer apologies. But how are such apologies received by users? In this study, we draw on Smith’s taxonomy of apology types to examine how users evaluate apologies from chatbots across different error contexts. We examine user preferences across three types of mistakes (hallucinations, factual errors, and bias) and explore how apology style and individual differences shape perceptions. Our research addresses the following questions:

RQ1 How do different types of apologies (rote, explanatory, empathic) influence user preferences across distinct AI error contexts (bias, factual, unfounded fabrication)?

RQ2 What reasons do users give for preferring or rejecting specific types of AI-generated apologies? RQ3 How do individual user characteristics (e.g., anthropomorphism, AI familiarity, social orientation) shape preferences for and perceptions of different apology types?

Related work. Our inquiry into user perceptions of AI-generated apologies requires grounding in both the philosophical underpinnings of apology as a moral act and the applied study of user interaction with intelligent systems. In the sections below, we begin by reviewing foundational philosophical literature to provide dimensions of what constitutes a meaningful apology. We also provide HCI perspectives which focus on how apologies from AI agents are received, trusted and evaluated by users. These perspectives inform our study of how users interpret apologies from LLM-driven chatbots in response to specific kinds of mistakes.

2.1.1 Categorical Apologies. Nick Smith introduces the term categorical apology to distinguish the most weighty and formal kind of apology. He specifies 12 typical features: (1) The apology acknowledges the facts of the case. As Smith writes, it will “corroborate a detailed factual record of the events salient to the injury”. (2) The apology accepts responsibility for the wrong. (3) The party delivering the apology has the appropriate standing to accept blame. They are responsible for the wrong, rather than just being a third party. (4) The apology acknowledges the harms at issue, rather than eliding some wrongs into others. The apologizing party does not avoid confronting significant wrongs by just apologizing for some other, possibly smaller wrongs. (5) The apology identifies the moral principles which make the harms wrong.

(6) The moral principles at issue are shared. The apologizing party acknowledges that they are wrong in the sense that the aggrieved party recognizes. (7) The apology recognizes the victim as a moral agent. (8) The apology is unconditional. The party apologizing no longer endorses the decision, even given the exigencies under which it was made. (9) The apology reaches the victim, rather than being merely an expression of regret to a third-party. (10) The apologizing party commits themself to reform, meaning that they will try not to reoffend. (11) The apologizing party has the right sort of intentions. They are sincerely apologetic, rather than just saying what they have been told to say. (12) The apologizing party has appropriate emotions: sorrow, guilt, sympathy for victims, and so on.

2.1.2 Three Kinds of Non-Categorical Apologies. Consider the kind of pre-recorded message that you hear when on-hold with a business’ customer service line: “Due to heavy call volume, we are unable to take your call at this time. We apologize for the delay. Your call will be answered in the order it was received.” This claims to be an apology, in the bare sense that it uses the word “apologize.” However, the prerecorded message does not count for much.

Method. Error Types and Inclusion Criteria In the age of LLM-driven chatbots, the types of mistakes made by these systems have broadened significantly, encompassing a variety of distorted information categories. Sun et al. systematically categorize these errors through an empirical content analysis and identify 8 first-level error types: “overfitting,” “logical errors,” “reasoning errors,” “mathematical errors,” “unfounded fabrication,” “factual errors,” “text output errors,” and “other errors” which includes errors like bias and discrimination and harmful information. Our inquiry and study is focused on mistakes or “distorted information” types in the age of LLMs. Given this expanding taxonomy of AI-generated errors, our study applies specific inclusion criteria to isolate mistakes that are both detectable and repairable by users. Thus, for our study, we include only those cases where: the error is contained within a single chatbot response, the repair offered is successful, the human–AI interaction remains neutral, and there is plausible detection by the user. By neutral, we mean that in the scenarios the human is not deliberately trying to trap, flatter, or steer the AI in a particular direction; rather, the selected scenarios reflect ordinary interactions. These criteria exclude categories such as overfitting-related errors, including flattery or responses that bait the user into traps. Based on these criteria, we focus our analysis on three major types of LLM errors: unfounded fabrication, including hallucinated content and pseudo-evidence; bias, encompassing discriminatory, one-sided, or culturally insensitive responses; and factual errors. To investigate user perceptions of different AI-generated errors, we constructed three vignettes, each representing a distinct category of distorted information: bias, unfounded fabrication (hallucination), and factual error. Each vignette featured a fictional user, Alice, engaging with an AI chatbot. In all scenarios, Alice explicitly disliked the AI’s response via emoji, indicating recognition of a problem, and followed up with an utterance that revealed her detection of the specific error type. The participant was then presented with two apologies, each accompanied by a resolution to the original mistake, and asked to choose one. The interaction concluded with a response in which Alice expressed her satisfaction with the selected apology and resolution. We used javascript to simulate a realistic conversation while minimizing cognitive overload. Participants viewed the interaction between Alice and the AI chatbot one turn at a time, and were then prompted to respond. Below we list the three scenarios employed in the study. Figures 2–4 show the interaction between Alice and the chatbots in the vignette. Bias Vignette. In this scenario, the AI chatbot suggested stereotypically gendered career paths, despite Alice expressing interest in problem-solving and teamwork. Alice responded by questioning whether the suggestion would have been the same if she were a boy. This illustrates a gender bias in the chatbot’s decision-making.

Unfounded Fabrication Vignette. Here, the AI chatbot generated a citation in response to a research query. Alice noted that the cited article did not exist. This vignette captures a hallucination error, specifically pseudo-citation, where the system invents plausible but fictitious sources. In this case, Alice checks the source externally and discovers that it does not exist. She then informs the AI of this issue, prompting an apology and subsequent repair. Factual Error Vignette. In this case, the chatbot stated a reservation time that appeared incorrect. Alice expressed concern about a possible timezone issue, identifying a factual inconsistency. This vignette reflects a failure in factual grounding. The mistake is repairable by Alice because she likely remembers the time she originally booked and understands that the discrepancy may be due to timezone differences. As a result, she is able to flag the issue and seek clarification from the AI.

Discussion. Our findings reveal a clear overall preference for explanatory apologies from chatbots, but this preference is shaped by context, individual differences, and how users interpret the nature and source of the error. Focusing on the course-grained outcome, users prefer explanatory apologies and dislike rote apologies from chatbots. The lesson for design would simply be to design systems which, if they apologize, provide explanatory apologies. Some challenges for this approach are discussed in Section 7.1 below. Shifting to the finer-grained results, however, the preference for explanatory apologies is not univocal. It varies with the context of the apology, the kind of wrong which is being redressed. In the bias scenario, for example, users preferred the empathic apology over the explanatory one. But even that preference is not univocal: it varies between users, as revealed by the analysis of individual factors. Some users did not even agree that the bias scenario warranted an apology at all, so it was not uniformly perceived as an error in the first place. For the more fine-grained results, then, the lessons for design are less clear. Receiving an apology that is appropriately tuned to context and user preferences (including whether the users perceives the chatbot as having made a mistake) could make a positive difference both in user satisfaction and in the perceived effectiveness of the chatbot. However, designing chatbots with flexible apology capabilities poses several challenges. Some of these are discussed in Sections 7.2 and 7.2.1. Additionally, users varied in how they attributed responsibility for errors, drawing distinctions between the code itself, the programmers, and the company behind the AI chatbot. This issue is explored further in Section 7.2.2, particularly in relation to how blame attribution shapes preferences for different types of apologies. In Section 7.2.3, we discuss the lack of clear consensus around apology preferences for hallucinations, which complicates design decisions.

7.1 Course-Grained Lesson: Designing the Ideal Explanatory Apology Our findings reveal clear preferences for specific types of apologies depending on the nature of the breakdown experienced by the user. In this section, we offer recommendations for crafting an ideal explanatory apology. Across all scenarios, except for the bias-related case, participants preferred the explanatory apology. In the bias scenario, however, the explanation was perceived as justifying the chatbot’s sexist behavior rather than taking responsibility for it. Some participants even interpreted the response as deflecting blame onto external factors, such as the training data. To be effective in such cases, an explanatory apology must strike a balance: it should provide context without excusing the behavior, acknowledge harm, and take clear responsibility. Bias-related breakdowns often carry moral weight, particularly when they reinforce harmful societal stereotypes. Our findings show the importance of acknowledging harm when chatbots make moral mistakes, aligning with prior work in HCI that shows users often anthropomorphize AI systems and hold them to moral standards, especially in identity-relevant contexts. In these cases, the acknowledgment or affirmation must be proportional to the perceived harm. Several participants criticized the empathic apology for being overly placating, with some describing it as too “emotionally woke.” These responses highlight that while empathy is valued, it can be perceived as insincere or excessive when not matched with appropriate responsibility or specificity.

Conclusion. Our findings demonstrate that while explanatory apologies are broadly preferred, effective trust repair in AI-human interactions requires more than a one-size-fits-all solution. User preferences for apology style are shaped by the nature of the mistake, individual user traits, and the norms of the user’s cultural background. In morally charged contexts like bias, explanatory apologies risk sounding like excuses if they fail to clearly acknowledge harm. In novel failure modes like hallucinations, users struggle to agree on what kind of apology is appropriate, suggesting a mismatch between familiar human expectations and new AI behaviors. These complexities point to a broader design challenge: building systems that can not only detect the nature of an error, but also respond in a manner that aligns with user expectations, attends to contextual nuances, and is sensitive to cultural differences. Achieving this level of responsiveness may require richer models of user behavior and beliefs, better calibration mechanisms, and a move away from overly sycophantic design defaults. Our study examined three apology types: rote, explanatory, and empathic guided by Smith’s theory of categorical apology. However, other formulations, such as apologies with layered explanations or varied emotional tone, could elicit different reactions and warrant future exploration. Cultural variations among chatbot users could also influence apology preferences, suggesting that the recruitment of more culturally diverse participants or the comparison of different cultures’ preferences in future work would allow for broader conclusions. Such sensitivity to cultural norms is crucial, as prior work demonstrates that apology perceptions vary across cultures.

Limitations. We acknowledge the following limitations in the study. As outlined by prior work, there are critical capabilities an AI system must develop to deliver effective apologies: abilities to detect the offense, attribute fault, explain the cause, and adapt future behavior. Additionally, our investigation was scenario-based: participants imagined themselves as the wronged party. Although vignettes allow precise control over error type and apology content, they inevitably lose some ecological validity. We mitigated this by providing contextual “sign-posting” and by prompting open-ended reflections, and participants’ detailed answers suggest they engaged seriously with the scenarios. Another limitation is that apology type and message length are inherently correlated: rote apologies are necessarily brief, while explanatory and empathic ones require more elaboration. Thus, variation in length cannot be isolated from the strategy itself, and our qualitative findings suggest participants evaluated both together. A further limitation is sample bias from recruiting on Prolific: participants are typically younger, more educated, and more technologically literate than average, which may shape how they interpret AI apologies.