From data to decisions: human–AI loops for trustworthy cultural heritage management

Objective: The adoption of Artificial Intelligence (AI) in Cultural Heritage management raises a question that goes beyond technological performance: how can AI-supported information be transformed into reliable evidence for managing and utilising Cultural Heritage? This paper explores the role of human–AI interaction in Cultural Heritage management, focusing on the concept of trust as a property of the overall evaluation process rather than of the AI model alone. The objective is to investigate how AI can act as an intermediate layer between heterogeneous heritage data and human decision-making, while maintaining human responsibility for interpretation, validation and contextual understanding. Method: This paper develops a comparative analysis of two experimental applications of Artificial Intelligence to Cultural Heritage, carried out at different scales and for different purposes. The first case concerns the urban environment of a historical city, where a pre-existing digital documentation and management project was subsequently extended through machine-learning-based object detection. Drone images were used to identify and classify water tanks on building rooftops and to georeference the detected elements within a three-dimensional Geographic Information System. The second case explores the use of generative AI for the visual reconstruction of historical phases and for the three-dimensional simulation of missing architectural elements of a religious heritage monument. The two cases are analysed according to the relationship between data, AI-generated outputs, evidence, human validation and subsequent interpretation or decision-making. Results: The comparison highlights two complementary dimensions of AI-supported heritage processes. In the urban scale case, AI enables the automation and repetition of an information-extraction process, transforming image-based observations into structured and geographically referenced data that can contribute to the monitoring and management of the historic built environment. In the architectural scale case, generative AI provides new possibilities for visualising historical hypotheses and reconstructing missing architectural elements, but the generated outputs must remain explicitly distinguished from historical evidence, illustrating the double-edged nature of AI in heritage processes. Despite their different technological functions, both cases demonstrate that AI outputs cannot be considered autonomous knowledge. Their usefulness depends on the possibility of testing, interpreting, questioning and validating them against available evidence and contextual knowledge. The analysis therefore identifies a human–AI loop in which human intervention is not limited to the final approval of an AI-generated result. Human experts contribute to defining the problem, interpreting evidence, identifying inconsistencies, challenging outputs and providing feedback that can improve subsequent iterations of the process. This interaction is particularly relevant in Cultural Heritage, where datasets are frequently incomplete, heterogeneous and context-dependent. From a management perspective, the resulting workflow can support the transformation of large and diverse datasets into evidence that informs monitoring, evaluation and decision-making without transferring responsibility for decisions to the AI system. These findings have economic implications: automating repetitive survey tasks at urban scale can reduce the cost and time of heritage monitoring. However, the human–AI loop shifts expert work towards validation rather than eliminating it. AI-enabled heritage initiatives call for business models that go beyond one-off technology purchases, such as service-based or public–private partnership arrangements in which technology providers and heritage institutions share the costs and responsibilities of validation. In the architectural scale case, generative reconstructions may also create value for visitor engagement and cultural tourism, provided that their hypothetical status is clearly communicated. Originality: The originality of the paper lies in proposing a process-oriented interpretation of trust in AI-supported Cultural Heritage management. Existing frameworks mainly address human involvement as oversight: the Ethics Guidelines for Trustworthy AI (HLEG, 2019) describe human-in-the-loop, human-on-the-loop and human-in-command approaches, and the EU AI Act (2024) requires human oversight of high-risk AI systems. In digital heritage, the London Charter (2009) and the Seville Principles (2011) state that it should be clear to users which parts of a computer-based reconstruction are evidence-based and which are hypothetical. Rather than assessing trust exclusively through the technical accuracy of an AI model, the paper considers trust as an emergent property of the complete chain connecting data, AI analysis, evidence, human validation and decision-making. This perspective interprets the concept of the human–AI loop as a framework for understanding responsible AI adoption in heritage processes. The paper argues that a trustworthy AI-assisted system is not necessarily one that produces the fewest errors, but one in which errors and uncertainties can be detected, questioned and corrected through an iterative interaction between computational capabilities and human expertise. This approach provides a conceptual basis for evaluating digital heritage initiatives not only according to their technological innovation, but also according to their capacity to support accountable, contextualised and sustainable heritage management.