Feeding AI with heritage is not enough: the limits of data must remain interpretable

Artificial intelligence is used to describe, organise and interpret cultural heritage collections, yet discussion often focuses on what systems can produce rather than on the knowledge embodied in their data. This contribution approaches that problem through underwater archaeology, where the material traces of a shipwreck are encountered through historical documents, nautical charts, acoustic surveys, photographs, oral testimony and family memories. The identification of the Brazilian auxiliary ship Vital de Oliveira, torpedoed in 1944, offers a case for examining how evidence converge without becoming interchangeable. Historical records, bathymetric products and personal accounts arise from distinct processes and support different kinds of claims. Their integration therefore cannot be reduced to technical interoperability or persuasive reconstruction. Drawing on debates in digital archaeology, archival studies and the ethics of artificial intelligence, we argue that responsible heritage applications require attention to four interrelated conditions: provenance, uncertainty, interpretive plurality and responsibility. Provenance and paradata make the circumstances of data production traceable; uncertainty distinguishes observation from inference; plurality keeps divergent interpretations open to scrutiny; and responsibility addresses consent, rights, institutional decisions and the sensitivities surrounding human memory. These principles are proposed as a curatorial orientation, not as an implemented computational ontology or an assessment of an operational AI system. Although grounded in underwater cultural heritage, the argument extends to museums, libraries, archives and digital collections. The central challenge is not simply to make heritage legible to artificial intelligence, but to ensure that the limits, omissions and mediations of its evidentiary foundations remain legible to people as well.