Searching for trustworthiness and transparency in a world full of AI

Today, the rise of generative AI (GenAI) is creating uncertainty: which information can be trusted? How was information produced? What is real and what is not? What has been generated by humans, and what by artificial intelligence (AI)?

As a university library, we aim through the Glashelder project to explore how we can address these questions in relation to the description of (image) collections. The project is based on the glass plate collection that was digitized in 2023.

Our goal is not only to provide practical guidance for our colleagues (for example, on how to use AI in the registration and description of heritage collections), but also to support researchers and other users who access the rich materials preserved in the Boekentoren through the online catalogue. We want them to be able to clearly see where AI has been used, in whole or in part, and to enrich the collection data in an accurate and validated manner.

Metadata

The cataloguing team at the Boekentoren has already provided the 61,000 glass plates with a technical description (material, damage, donor, collection, etc.). Wherever possible, an indication of the date and/or a provisional title and description has also been added.

However, providing this extensive collection of glass plates with comprehensive metadata and detailed descriptions is too time-consuming and labour-intensive. For this reason, we are enlisting the help of AI to generate, for example, descriptions and keywords for the glass plates. These outputs are tested for usability and validated by humans before they are published in the catalogue.

But how do we document this collaboration between humans and AI? How can we do so in a transparent and clear manner? How can we make visible which elements were generated by people and which by AI?

Transparency through AI-labels?

Outside the cultural heritage sector, several initiatives have already been developed to clarify where and how AI is used through AI labels and attribution systems. Some adopt creative approaches, such as AI Nutrition Facts, which is designed in the style of a nutrition label, or the AI Ethics Label, which is based on the European energy label used for electrical appliances.

Other initiatives focus primarily on researchers and publishers, for example through AI Usage Cards or the Artificial Intelligence Disclosure (AID) Framework.

However, these approaches are less suitable in the context of a heritage library.

A particularly useful initiative for the cultural heritage sector is the AI Attribution Toolkit (AIA) developed by IBM. Bart Magnus of meemoo already highlighted this tool in May 2026 in the journal META. The toolkit is concise and, thanks to its broad formulation, can be applied to many different types of data. It also provides a clear layered structure, indicating: the degree of AI involvement; the type of output generated by AI; and the level of human review applied to that output. Its structure shows similarities to the building blocks of Creative Commons (CC) licenses, which are widely adopted within the cultural heritage sector.

The Creative Commons community is currently working on the development of “CC Signals,” which would enable machines and AI systems to determine what data they may (re)use and under what conditions. However, the IBM toolkit is not (yet) machine-readable in the way that Creative Commons licences are. For the heritage sector, machine-readability is nevertheless important, as it facilitates data exchange, interoperability, long-term preservation, and future-proofing.

Discussion panel on AI attribution

However, AI attribution guidelines or a sector-wide standard cannot be created overnight. There are many factors that need to be taken into account. As a first exploratory step towards developing (shared) guidelines or standards, we organized an interactive discussion panel on AI attribution during the Erfgoed & AI Studiedag, which took place in Antwerp on 17 June 2026. We brought together Manou de Sutter (KMSKA), Merel Geerlings (Amsterdam City Archives), Bart Magnus (meemoo), and Eva Andersen (Glashelder, Boekentoren), with the discussion moderated by Annamaria Van Ingelgem (Boekentoren).

The discussion made it clear that there is broad consensus on the importance of transparency regarding the use of artificial intelligence within the heritage sector. Opinions diverged, however, on the desirability of a sector-wide standard for AI attribution. The question of who should take the lead in developing such a standard also remained a subject of debate. Should standardization be implemented by individual archives and heritage institutions, or would a framework at the Flemish, national, or European level be more appropriate? Several participants felt that international consortia of heritage institutions might be best positioned to take on this challenge.

For the time being, many questions remain regarding terminology, feasibility, infrastructure, scalability, sector-specific requirements, and how to deal with the rapid evolution of AI technologies. For example, there is not yet a consensus on what exactly should be understood by a “standard.” Should such a standard be solely human-readable, or should it also be machine-readable? Is it realistic to aim for a single uniform standard, or would a more differentiated approach be preferable?

In addition, questions arise about the appropriate level at which AI attribution should be applied: at the collection, object, or item level. The relationship between AI attribution and other forms of transparency, such as disclaimers, also remains unclear. Furthermore, it has yet to be determined whether such a standard should be aimed primarily at end users, or whether it should also serve a function within the internal processes of cultural heritage institutions.

Finally, participants raised questions about the level of AI literacy within the sector, the impact of existing technical infrastructures and data standards, and the actual information needs of visitors and users. To what extent do they wish to know whether collection descriptions or other forms of enrichment were created with the assistance of GenAI?

These are fascinating questions for us to continue exploring together in the near future!

Although the questions surrounding AI attribution within the Boekentoren and the wider cultural heritage sector have not yet been fully answered, we aim by the end of the Glashelder project in 2028 to clearly indicate which information has been generated with the help of AI and which has not. By then, all enriched metadata will have been incorporated into the catalogue.

In doing so, we must develop an approach that is flexible enough to evolve alongside new AI technologies. At the same time, we need to take into account the limitations of the MARC21 bibliographic format on which our catalogue is based. The capabilities of the current platform infrastructure will also play an important role in shaping this approach.