An evidence-based framework for evaluating generative AI in cultural heritage reconstruction: Bamiyan, Palmyra and bounded generalisation

Abstract

Generative AI can transform textual, photographic, archaeological and conservation records into visual reconstruction candidates for damaged or destroyed cultural heritage. This creates an opportunity to visualise aspects of heritage appearance that were previously difficult to reconstruct from dispersed evidence alone. The evaluative problem is acute where monuments can no longer be directly recorded and reconstruction depends on incomplete photographic, archaeological and conservation evidence. Under these conditions, a generated image may appear coherent while its evidential basis, prompt construction, interpretive assumptions and cultural acceptability remain insufficiently documented. The research asks how the evaluation of GenAI reconstruction candidates can be standardised before such outputs are treated as heritage reconstruction. To address this problem, the research develops the Heritage-Aligned Reconstruction Framework (HARF) for evaluating GenAI heritage reconstruction. HARF is formalised through the Dynamic Prompt Blueprint, the Prompt Sufficiency Index, paradata, computational pre-evaluation and structured expert review. The framework is operationalised through the Western Buddha of Bamiyan, where 245 generated candidates are assessed through computational similarity, interdisciplinary screening and specialist archaeological evaluation. The Bamiyan case is extended through a community-grounded authenticity study with 62 Afghan respondents. HARF is re-instantiated at the Temple of Bel, Palmyra, to test bounded cross-site generalisation from figural sculpture to monumental architecture. The thesis makes four contributions. It establishes an evidence-constrained workflow for GenAI heritage reconstruction; formalises HARF as an integrated AI-specific evaluation framework; shows that reconstruction validity at Bamiyan is distributed across computational, specialist and community-grounded registers; and demonstrates that HARF can be transferred to Palmyra at the level of evaluative schema without redesigning its top-level structure. The Palmyra test also shows that failure domains shift with typology, from figural micro-grammar at Bamiyan to architectural and decorative micro-grammar at Palmyra. The contribution lies not in producing definitive reconstructions of Bamiyan or Palmyra, but in establishing how generated reconstruction candidates should be evaluated before they are treated as heritage reconstruction. HARF provides a structured procedure for assessing whether such candidates are evidence-based, documented, expert-assessed, culturally accountable and bounded in transfer.