Consumers shown incorrect advice from an AI shopping assistant reported a much stronger sense that the service had failed to meet their expectations, according to a new European study. The results suggest that trust, rather than the error alone, may determine whether shoppers keep using the tool or buy the recommended product.
An AI shopping assistant can compare products, summarize specifications and answer questions in seconds. It can also produce a confident answer that is invented or wrong. Researchers commonly call this type of error a hallucination.
A study published on August 17 in the Journal of Theoretical and Applied Electronic Commerce Research examined how such errors may affect consumer trust. The research was led by Sayyed Khawar Abbas of Corvinus University of Budapest, with Hafiz Muhammad Junaid of the University of Essex and Aseel Smerat of Al-Ahliyya Amman University.
The findings are relevant to retailers adding generative AI to product search and customer service. They also require careful interpretation: the researchers measured responses to shopping scenarios and stated intentions, not completed purchases or sales lost in a live store.
What the researchers tested
The project began with 25 interviews involving consumers in 15 European countries who said they had encountered incorrect AI output while shopping online. The researchers used those interviews to develop a larger survey experiment.
That second phase included 590 English-speaking consumers from the same 15 countries. Each participant read a short scenario in which an AI assistant gave either an accurate or a hallucinated product recommendation. Some also saw a transparency cue about the system.
Participants shown the hallucinated recommendation gave an average expectation-violation score of 4.48, compared with 2.18 for those shown the accurate recommendation. The statistical effect was large.
The authors then modeled what happened after this initial reaction. A stronger sense that the assistant had failed prompted more checking and skepticism. This was associated with a lower assessment of the AI’s competence and a greater adjustment of trust. Greater trust adjustment, in turn, was linked to weaker intentions to keep using the assistant or buy the product, and to a greater likelihood of negative word of mouth.
The direct path from exposure to a wrong answer to those three behavioral intentions was not significant once trust adjustment was included in the model. The commercial risk therefore appears to run through the shopper’s judgment of whether the assistant remains dependable.
Reliable product data matters more than an apology
The interviews indicated that consumers wanted the underlying problem fixed. Participants generally viewed apologies and broad explanations as less useful than changes designed to prevent another false answer.
The researchers recommend measures including retrieval-augmented generation, uncertainty disclosures and access to human help. Retrieval-augmented generation, usually shortened to RAG, lets an AI system consult approved sources before composing an answer. In retail, those sources could include current product specifications, prices, inventory records and promotion rules.
For a retailer, this means the quality of the information feeding the assistant may matter as much as the language model itself. A useful system should be able to cite the product record it used, say when reliable information is unavailable and pass sensitive or unusual questions to a person.
The study did not run a controlled comparison of those repair methods. The recommendations came mainly from the interview findings and the authors’ interpretation, so they should not be presented as proof that one technical fix will restore trust in every retail setting.
Disclosure is not an accuracy control
Transparency produced a more complicated result. Shoppers who perceived the system as more transparent tended to react differently, but that measure was based on their responses after the scenario. When the researchers retested the question using the randomly assigned transparency cue, the cue did not significantly reduce the effect of a hallucinated answer on expectation violation.
This matters because disclosure and accuracy solve separate problems. A notice can tell customers that they are dealing with AI. It cannot make an incorrect specification correct.
In the European Union, Article 50 transparency requirements under the AI Act began applying on August 2, 2026. The European Commission says providers must design systems that interact directly with people so users are informed that they are interacting with AI, unless this is already obvious from the circumstances.
Retailers may therefore need both a clear disclosure for compliance and strong information controls for reliability. Treating the label as a substitute for accurate product data would leave the main customer problem unresolved.
The sample was self-selected, recruited through online channels and limited to English-speaking respondents. Country groups were also uneven. The results should not be treated as population estimates for all European shoppers, and stated intentions do not always predict real purchasing behavior.
Even with those limits, the study identifies a practical test for AI commerce. Retailers should measure more than engagement and response speed. They also need to track unsupported claims, correction rates, source coverage and the point at which customers abandon the assistant or seek human help. A shopping tool only saves time while its answers remain worth trusting.
