Users punish AI mistakes more harshly than human expert errors

What happens after AI gets something wrong? New research suggests that a simple apology from an AI system may do little to restore user trust after a visible error and, in some cases, may perform worse than an apology from a human expert who made the same mistake.

The study, titled "Apologizing artificial intelligence: designing and evaluating effective AI apologies after errors," was published in AI & Society. The authors used two experiments to compare how users respond when AI and human experts make noticeable advice errors, then test whether different apology styles can repair confidence and reliance.

Users punish AI errors more harshly than human mistakes

AI tools are becoming more accurate, but the study starts from a key reality of automated decision support: these systems still fail. Even one visible mistake can damage trust because users often expect AI to be efficient, objective and precise. Unlike human experts, AI systems are also widely perceived as opaque, making it harder for users to understand why an error occurred or whether it is likely to happen again.

The research focuses on trust repair, a growing issue as AI systems are used in advice-giving roles. In human relationships, apologies often help repair trust after mistakes by signaling regret, accountability or willingness to improve. The authors show that this social logic does not transfer cleanly to AI. Users do not necessarily treat an apologizing AI the same way they treat an apologizing person.

They examined both behavioral trust and perceptual trust. Behavioral trust was measured through reliance on advice, meaning how much participants adjusted their own judgment after receiving a recommendation from an AI or human expert. Perceptual trust was measured through confidence in the advice source. This distinction is important because people may say they trust a system while still refusing to follow its advice, or may use a recommendation despite having doubts about it.

Across the experiments, participants completed judgment tasks after receiving advice from either an AI system or a human expert. They first observed the advisor performing well, which allowed some initial trust to form. The advisor then made a noticeable error. After the error, the researchers tested whether an apology, and later different types of apology explanations, could restore trust.

The first major finding was that confidence in advice was positively linked to reliance on that advice. Participants were more likely to use recommendations from either AI or human experts when they had greater confidence in them. That result may seem obvious, but it establishes a key point for AI design: trust is not just an attitude. It affects whether people actually use AI outputs when making decisions.

The more consequential finding was that AI did not recover trust as easily as human experts. When the AI made a noticeable error and apologized, users were less forgiving than when a human expert made the same error and apologized. The decline in reliance on the advice was stronger after the AI error than after the human error.

This supports a broader pattern known as algorithm aversion, in which people become reluctant to use algorithmic advice after seeing it fail. The study suggests that users may hold AI to a different standard. A human mistake may be seen as understandable or correctable. An AI mistake may be read as a signal of deeper system unreliability.

That divide matters because AI is often promoted as more objective and consistent than humans. Those same expectations can backfire when the system errs. A mistake by a person may be seen as part of human fallibility. A mistake by AI may violate the expectation that automation should be stable and accurate. As a result, the trust loss can be sharper.

The authors also found that a simple apology from AI was not enough to repair trust. In both experiments, apologies without explanation did not significantly restore reliance or confidence after an AI error. This is a critical result for developers and companies designing AI interfaces. A generic apology message may appear polite, but the study shows it does not reliably solve the trust problem.

The finding also raises questions for consumer-facing AI products, digital assistants, automated advisory systems and workplace decision-support tools. If users abandon AI after errors, organizations may fail to gain the benefits of automation even when the systems are usually accurate. But if systems apologize poorly, users may feel further alienated or confused.

Why the wording of an AI apology matters

The second experiment tested whether apology explanations made a difference. The researchers compared apologies with internal attribution and external attribution. An internal attribution suggests the problem came from the AI's own limitations. An external attribution points to factors outside the AI system, such as insufficiently similar input information or limitations in the data fed into the system.

The results showed that attribution mattered, but not in the way human apology research might predict. In human interactions, taking responsibility often helps rebuild trust because it signals accountability and willingness to improve. But for non-anthropomorphized AI, the study found that external attribution could be more effective than internal attribution, especially for restoring reliance in objective tasks.

When AI admits that the error came from its own limitations, users may interpret the problem as stable and likely to recur. Unlike humans, AI systems may not be perceived as capable of learning, improving or changing in a socially meaningful way. An internal apology may therefore reinforce doubts about competence rather than repair them.

On the other hand, an external attribution can reduce the perceived blame placed directly on the AI system. If users believe the error resulted from poor input data or conditions outside the system's normal competence, they may be more willing to rely on it again. In objective tasks, that strategy appeared more effective in repairing behavioral trust.

However, the study does not offer a blanket endorsement of external-blame apologies. The authors note that such apologies raise ethical concerns if they misrepresent the true cause of failure. If the AI system itself was defective, blaming external conditions could obscure responsibility and mislead users. In real-world settings, responsibility may lie with developers, vendors, organizations, training data, interface design or deployment choices, not with the AI as an independent moral agent.

This creates a difficult design problem. The apology style that most restores user reliance may not always be the most transparent or ethically appropriate. Companies may be tempted to protect trust by framing failures as external, but doing so could shift blame away from the organizations responsible for the system. The study therefore points to the need for truthful, context-sensitive error explanations rather than formulaic apology scripts.

The research also found that apology effects were stronger for reliance than for confidence. In several cases, users changed how much they relied on advice, but their stated confidence did not recover in the same way. This suggests that behavioral trust may be more responsive to design changes than self-reported confidence. Users may be willing to use AI advice again under some conditions even if their belief in the system has not fully rebounded.

Measuring only user confidence may miss meaningful changes in behavior. Conversely, measuring only behavior may hide unresolved doubts. Trust repair in AI systems should therefore be assessed through both what users say and what they do.

The study also compared objective and subjective tasks. Objective tasks involve measurable ground truth, such as estimating weight from an image. Subjective tasks involve judgment and interpretation, such as estimating attractiveness scores. Users generally expect AI to be better suited to objective tasks because these tasks appear more rule-based and less dependent on human intuition.

That expectation shaped responses to AI failure. The study found that users were less forgiving when AI erred in objective tasks than in subjective tasks. The decline in reliance was stronger in objective contexts. This likely reflects the higher performance standard users apply to AI when the task seems factual or measurable. If AI fails at a task it is expected to handle well, users may view the failure as more serious.

In subjective tasks, expectations may already be lower. Users may not expect AI to fully capture personal taste, nuance or interpretation. As a result, a mistake in a subjective task may be less damaging to reliance because the system was not expected to be perfect in the first place.

External-attribution apologies also worked better in objective tasks than subjective tasks. This suggests that apology design must account for context. A repair strategy that helps after an AI error in a structured decision task may not work the same way in a more personal, interpretive or affective setting.

Findings raise design and ethics questions for AI systems

To sum up, AI systems cannot rely on human-style apology scripts after errors. A simple apology may fail to repair trust, and an internal admission of limitation may sometimes deepen concerns about system competence. If AI systems are going to explain failures, those explanations need to be designed carefully around the type of task, the likely user expectations and the actual cause of the error.

This has wide relevance as AI systems move into higher-stakes advisory roles. AI is increasingly used to support decisions in hiring, healthcare, finance, education, customer service, navigation, workplace productivity and legal information. In many of these contexts, errors are inevitable. The question is not only how to reduce errors, but how systems should respond when users notice them.

For companies, the findings suggest that trust repair cannot be treated as a cosmetic interface feature. A polite message after a mistake does not necessarily restore confidence or use. Users may need a clear, accurate explanation of what went wrong, why it happened, whether it is likely to happen again and what has been done to prevent recurrence.

Developers need to design a new category: apologizing AI that goes beyond explainable AI, which often focuses on how models reach outputs. Apologizing AI deals with how systems communicate after failure. The two are connected. A good apology or error response may need to explain the system's limits, input conditions, uncertainty and accountability structure.

For organizations, the ethical issue is accountability. AI does not have moral agency in the way humans do. If an AI system apologizes, users may wrongly infer that the system itself is responsible. But responsibility may rest with the firm that deployed it, the team that designed it, the data used to train it or the process that placed it in a given decision context. Poorly designed apologies could blur these lines.

The authors also warn against overusing AI apologies. Frequent automated apologies may become mechanical and meaningless. They may also reduce the value of human apology in serious situations. In minor, low-risk settings, an AI apology may improve user experience. In high-risk settings, a generic apology may be inadequate or even misleading if it masks deeper safety or accountability issues.

The findings also extend to affective AI and AI companions, which are built to simulate emotional responsiveness. If users already doubt AI's ability to handle subjective or emotional tasks, errors in those settings may require a different approach from errors in technical or objective decision support. Companies offering emotional AI tools may need to be especially transparent about system limits.

The research does not suggest that AI apologies should be abandoned. Instead, it shows that apology design must be evidence-based. The effectiveness of an apology depends on the source of advice, the nature of the error, the attribution used and the task context. AI trust repair is not simply human trust repair applied to machines.

The analysis is not without limits. The experiments used controlled online tasks involving image-based estimates, not high-risk clinical, legal or financial decisions. Participants interacted with advice sources over a limited number of rounds. Real-world trust may develop differently over longer periods, with repeated use, higher stakes and more complex systems. The findings therefore need testing in other domains and with more varied AI tools.