As the digital transformation in the cultural heritage domain unfolds, the responsibility of museums to document and be transparent no longer applies only to human users but also to artificial users. For instance, to facilitate the return of objects to their rightful owners, museums should publish information about the ownership history of their collections (i.e., the provenance) not only as text but also as data that is machine-readable and compliant with FAIR principles (Findability, Accessibility, Interoperability, and Reusability).
Institutions face a dilemma in digitizing their collection information. Although museums have already recorded much of the information to be converted into data, it is in the form of free text and is insufficiently structured. While rerecording this information by hand in a standard, machine-readable format would require a significant investment of resources and time, fully automating the data-structuring process would call into question the quality of the data produced, with the risk of perpetuating historical biases and omissions. Focusing on museum provenance information, this paper illustrates how the use of AI models for natural language processing tasks can help institutions automatically structure provenance texts as linked open data. Finally, considering not only quantitative but also qualitative needs, the paper describes how expert users can critically intervene in data production through a human-in-the-loop approach.