Museum AI survey links responsive dialogue to visitors’ sense of heritage value

An AI museum guide can answer a visitor's question about an object. A more ambitious version can keep a conversation going, identify what is uncertain about a historical claim and give the visitor some control over how to explore an exhibition. Research on a generative-AI guide known as Jinghui suggests that these qualities are associated with different aspects of visitors' reported experience. It does not establish that chatting with the system actually made people protect cultural heritage.[1]

The study was published on 29 October in Sustainability. Researchers surveyed adults who said they had previously used Jinghui, recruiting through Xiaohongshu, WeChat and invitations at the Capital Museum between 15 May and 15 June 2026. Of 532 questionnaires, 13 were removed for screening reasons, leaving 519 for analysis. Eligibility depended on people reporting prior use; the investigators did not verify that claim against individual interaction logs. These are survey data collected after the experience, not a randomized comparison with a conventional museum guide.[1]

The team distinguished six visitor-perceived qualities: responsive dialogue, clarity about the evidence behind answers, social presence, cultural authenticity, personalised explanations and the ability to direct the exploration. It then examined their statistical relationships with perceived heritage value, immersion or flow, confidence in acting to support heritage and stated intentions to engage in heritage stewardship. The distinctions matter. Feeling moved by a display, believing one could help preserve an object and actually volunteering or donating are not equivalent outcomes.[1]

Responsive dialogue had one of the strongest fitted associations with perceived heritage-value enhancement and flow. Visitor agency was particularly prominent in the team's neural-network ranking for flow, while transparency about what the system knows and perceived cultural authenticity ranked closely for stewardship self-efficacy. These are associations estimated within a specified model of survey responses. Small differences between coefficients do not establish that one design feature is universally superior; the authors explicitly warn that neural-network importance ranks can vary across fitted runs.[1]

One proposed connection did not survive the test. The direct statistical path from perceived heritage-value enhancement to stewardship self-efficacy was unsupported. The measure concerned visitors' reported increase in appreciation attributable to the guide, not their absolute appreciation of heritage. That reported enhancement had no supported direct link to greater confidence in taking heritage-supporting action once the other model relationships were included. That is an important distinction for cultural institutions that may be tempted to equate engagement metrics with public participation. An emotionally compelling AI tour may provide meaning without supplying the opportunities or practical confidence needed to protect heritage.[1]

The researchers combined structural-equation modelling, neural networks, a Basis-KAN comparison and necessary-condition analysis. These methods answer somewhat different questions. A significant association does not automatically make a variable a good out-of-sample predictor; a strong predictive ranking does not prove causation; and statistical necessity under a particular measurement model is not a universal law about visitors. The study even found that a simpler ridge model outperformed the evaluated Basis-KAN approach on mean predictive accuracy in one comparison. Added model complexity was not synonymous with better explanation.[1]

For museums, a practical design implication is to keep the provenance of an answer visible. When a visitor asks how an object was used, what is documented, what is scholarly interpretation and what remains disputed should not blend into a polished but untraceable narrative. A guide able to sustain responsive dialogue may make a visit more personal, yet the museum remains responsible for the cultural claims it presents. More personalisation also creates decisions about accessibility, how visitors' questions are retained and whether curator-approved material can be altered by automatic generation.[1]

None of those governance outcomes was directly established by the Jinghui survey. The sample was recruited through convenience channels, concentrated among relatively young and educated respondents, and limited to one setting. Its self-reported measures do not show that actual conservation behaviour increased, and there was no randomized group using an alternative guide. The evidence supports a more modest and useful proposition: visitors perceive different qualities of an AI guide as relevant to meaning, involvement and agency. A subsequent field study could observe visitor understanding and concrete actions over time rather than inferring them from intention alone.[1]

The useful next test would compare the guide with another form of interpretation, check the accuracy of its answers and follow what visitors subsequently do. The survey offers hypotheses about dialogue, evidence and agency; it cannot decide which design changes improve heritage stewardship.[1]

Report an error or suggest a correction