Source accountability in AI-mediated heritage interpretation under controlled data-source contamination

AI systems transform heritage sources into museum, archival and tourism-facing accounts, but whether they preserve claims’ evidentiary status remains unclear. We introduce SOURCE-CT, a controlled data-source contamination benchmark for computational heritage science. SOURCE-CT tested 90 heritage cases from Korea, China and Japan across two evidence conditions, six prompt variants and three deployment conditions, producing 3240 outputs. In 1620 mixed-source outputs, false-claim acceptance was 0/540 for the time-stamped ChatGPT web-interface condition, 41/540 for Qwen2.5-3B and 34/540 for Llama3.2-3B. Because conditions differed in model scale, alignment, safety tuning, execution environment, infrastructure and decoding control, contrasts are interpreted at the deployment-condition level. Interpretive-causal claims showed the highest acceptance, particularly when they supplied coherent explanations of origin, memory or cultural significance. SOURCE-CT reframes AI hallucination in heritage communication as a source-boundary problem and provides a reproducible pre-deployment audit of evidentiary-boundary preservation before public release.