This study examines how audiences form and revise perceived authenticity when encountering AI-generated images of cultural heritage. Focusing on Dunhuang heritage artificial intelligence-generated images (AIGIs), it investigates how visual fluency, prior heritage experience and source information shape authenticity perception under conditions of AI hallucination.
The study adopts a qualitative design based on image-elicitation interviews with 20 purposively sampled participants recruited via Xiaohongshu. All participants had prior exposure to Dunhuang heritage. Each interview included a pre-interview survey, an image-viewing task and a follow-up interview after source disclosure. Participants evaluated 30 images, including 15 original Dunhuang murals and 15 AIGIs, across eight thematic sets. The data were analysed through thematic analysis, with the heuristic-systematic model (HSM) used as the interpretive framework.
The findings show that perceived authenticity does not depend on visual quality alone. First, higher self-assessed heritage knowledge did not produce greater confidence. It often made participants more alert to visual inconsistencies and possible AI hallucination, which led them to question their own assessments. Second, prior heritage experience shaped how participants evaluated AIGIs. Embodied memories of site visits, exhibitions and digital tours provided reference points for recognising visual or historical inconsistency. Third, uncertainty did not always lead to sustained scrutiny. Many participants recognised the limits of their judgement, yet lacked the knowledge, confidence or willingness to examine every image in detail.
This study advances research on digital heritage authenticity by showing how perceived authenticity is formed, revised and sometimes unsettled during encounters with AIGIs. It extends HSM into a heritage context where authenticity is tied to cultural memory, institutional trust and prior experience. It also reframes AI hallucination as more than a technical defect. In heritage AIGIs, it conveys visual assumptions about the past and reshapes how audiences perceive both heritage authenticity and the reliability of generative technologies. The findings highlight the need for clearer AIGI disclosure, stronger production standards and more active institutional responsibility.
