When Chatbots Confabulate: Human Rescue Softens Frustration but Not Doubt

When an online shopping chatbot confidently tells you that a rice cooker with a 1.6-liter capacity is actually 16 liters, or promises a free blender worth 999 yuan that no merchant ever intended to give away, the damage goes beyond a single wrong answer. A new pair of controlled experiments from Wuchang Shouyi University in Wuhan, published in Frontiers in Psychology, shows that such AI hallucinations drag down consumers' ratings of chatbot service quality through two distinct psychological channels, and that a well-timed human agent can repair one of them but not the other. The research team, led by Xi Chen, ran two scenario-based experiments with a total of 600 participants recruited through the Chinese survey platform Credamo.

A pre-experiment with 150 valid responses confirmed the manipulation worked: perceived hallucination scored 5.80 on a 7-point scale in the high condition versus 2.39 in the low condition. Analysis of covariance showed a strong main effect of hallucination on perceived service quality, and a significant interaction confirming that human recovery buffered the damage. The numbers tell a striking story.

Without human intervention, service quality ratings collapsed from 5.48 in the low-hallucination condition to 2.73 in the high-hallucination condition as errors mounted. With a human agent stepping in, the decline was much gentler, from 5.77 to 3.87. In other words, consumers have remarkably little tolerance for AI-generated misinformation when no one is available to correct it, but a brief, well-crafted human apology and correction can absorb a substantial share of the reputational cost.

Using partial least squares structural equation modeling with 5,000 bootstrap resamples, the researchers tested whether cognitive dissonance and technical frustration mediate the link between perceived hallucination and service quality. Both pathways were significant. Perceived hallucination strongly predicted cognitive dissonance (a standardized path coefficient of 0.671) and technical frustration (0.749), and each of these in turn reduced service quality evaluations.

Together the two indirect pathways accounted for 66.8 percent of the total effect, meaning that roughly two-thirds of the harm hallucination does to service ratings travels through these internal states rather than through the factual errors alone. The most consequential finding concerns the limits of human recovery. The interaction between hallucination and human agent significantly moderated the frustration pathway and the overall service quality rating, but showed no significant effect on cognitive dissonance.

The authors interpret this asymmetry through the nature of the two responses. Technical frustration is a situational emotion triggered by blocked goals, and a real-time human intervention that explains, corrects, and reassures can dissipate it. Cognitive dissonance, by contrast, reflects a deeper conflict between the consumer's prior belief that the chatbot is a reliable information source and the undeniable evidence that it lied.

Once that negative cognitive anchor forms, an apology and a corrected fact sheet do not dissolve it; the consumer still knows the machine got it wrong. The theoretical framing draws on Festinger's classic cognitive dissonance theory and on goal-obstruction accounts of frustration. Consumers approach e-commerce chatbots with pre-formed expectations of accuracy built from product pages and prior experience.

When assertive, anthropomorphic responses contradict verifiable facts, the resulting disconfirmation creates psychological discomfort that users resolve partly by devaluing the service. Frustration operates in parallel: inaccurate answers leave questions unresolved, forcing extra verification effort that spills over into negative judgments of the entire encounter. Prior work had linked hallucination to discontinuance of generative AI use and to frustration in software engineering contexts; this study extends both mechanisms to consumer service evaluations in online retail.

The authors are careful about boundaries. The chatbot was scripted rather than a live LLM, the human agent's message was pre-written, only one product category was tested, the consultation lasted under five minutes, and there was no error-free control group, since even the low-hallucination script contained two errors. Self-report measures may also be subject to common method bias, and participants' prior chatbot familiarity was not measured.

The findings therefore concern reactions to controlled service scenarios, not deployed production systems. Still, the practical implication is pointed: technical fixes like retrieval-augmented generation and data filtering reduce but never eliminate hallucination, so platforms should build agile human-AI handoff workflows, using conversational monitoring to detect user frustration in real time and route sessions to human agents for factual correction. But platforms should not expect such interventions to fully restore broken trust, because the cognitive wound of discovering an AI's confident falsehood appears to heal far more slowly than the emotional irritation it causes.

Subject of Research: Experimental study of how AI hallucination affects chatbot service quality in e-commerce and the moderating role of human agent service recovery

Article Title: The impact of AI hallucination on the chatbot service quality in e-commerce: the moderating role of human agent based service recovery

Article References: Chen, X., Liu, J., Chen, Q., & Li, L. (2026). The impact of AI hallucination on the chatbot service quality in e-commerce: the moderating role of human agent based service recovery. Frontiers in Psychology, 17, 1870991. https://doi.org/10.3389/fpsyg.2026.1870991

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1870991

Keywords: AI hallucination, chatbot, e-commerce, service quality, cognitive dissonance, technical frustration, service recovery, human agent, large language models, consumer psychology, human-AI collaboration, scenario-based experiment