‘BIAS – Towards a Bias Impact Assessment Scale for digitised cultural heritage collections’. Under this ambitious title, we – two legal scholars specialising in the interplay between law, technology and culture – started our Researcher-in-Residence (RiR) project at the National Library of The Netherlands (hereafter: KB) in August 2024 with an interdisciplinary team. Complementing our legal perspective, the team included a digital collection specialist, a computer scientist and an artificial intelligence (AI) researcher from the KB.
The project concerns an issue which is both fundamental and topical. For instance, AI is on the rise in unlocking cultural heritage collections, while cultural heritage institutions including the KB also contribute to the development of AI. That is, the KB recently agreed to make available digital public domain texts from its collections to help train the Dutch language model of GPT-NL, which aims to serve as an ethically responsible public alternative to ChatGPT. In addition, AI is central to multiple regulatory initiatives at the European Union level, where the risk of bias is repeatedly flagged (cf. Breemen and Breemen 2025). And then there is the precarious position libraries and other cultural heritage institutions increasingly find themselves in, given the current political climate in various regions of the world, regarding their traditional role to safeguard the fundamental values of diversity, equity and inclusivity (Ng et al 2025).
Against this background, our envisaged contribution to the bias discussion is twofold. First, we aim to discuss the concept of bias and different manifestations; second, together with the KB team, we explore a way to flag types of bias in selected digitised KB collections. Accompanying both aims as a central thread in the research is the perspective of positionality, i.e. awareness of both researchers’ own worldviews as shaped by their backgrounds and how this influences their research (cf. Darwin Holmes 2020). In the case of this research, the theoretical underpinnings of ‘positioning’ and criticising perceived neutrality, in particular in challenging hegemony and adding positions, are a central feature (cf. King 2024, Haraway 1988, Caswell 2021).
The remainder of this blog will introduce the project in more detail. It does so by positioning the project in the bias discussion and explaining our approach as well as some practical difficulties and limitations.
‘RiR’: a project on ‘bias’
Bias is clearly a buzz word, featuring in multiple projects in the cultural heritage sector (see for example Europeana’s DE-BIAS initiative and the Huygens Institute’s Combatting Bias research). Whereas the KB uses AI in various ways – including for efficiency reasons, such as suggesting keywords in the catalog or image and text recognition to enhance the readability of its digitised collections – the topic of ‘bias’ is also addressed. In its ethical principles on AI, the KB acknowledges that none of its (physical and digital) holdings contains datasets that are entirely bias-free, due to changing collection policies over time. On the one hand therefore, the KB aspires to know where and to what extent bias occurs; on the other hand, the KB indicates to value removing or compensating bias to safeguard inclusivity. As we will explain below, our envisaged contribution ties in with the former, i.e. raising awareness and transparency. We explicitly do not argue for removing bias.
From a legal perspective, bias is high on the agenda as well, since the term occurs often in EU policy documents on AI regulation (see for instance 25 mentions in the so-called AI Act 2024). Taking the cultural heritage and legal contexts together formed the background of our RiR project. Still, our focus is not on ‘bias and AI’ as such; rather, we take a step back to make space for reflection, endorsing the slow archives approach elaborated by Christen & Anderson. That is, examining “structures, practices, and processes of collection, cataloging, and curation to expose where cultural authority is placed, valued, and organized within archival workflows” (Christen & Anderson 2019).
Before discussing the concept of bias in more detail in the next blog, we acknowledge that there are different layers of bias in the cultural heritage context – from the composition of collections and which authors are or are not included, to the content of the works, and to the choices for metadata systems to unlock the collections. To be clear: our aim is not to de-bias collections, but to increase transparency and awareness of potential biases among users of the digitised collections, who will have different backgrounds and may not be aware of the intricacies stemming from for instance the origins, orientation or time frame of the materials. This is recognised by Europeana’s datasheets initiative: originally a machine learning concept, datasheets are currently being adapted to the cultural heritage field, aiming to provide context on the provenance of cultural data in order to encourage informed decisions on the data’s use (see Alkemade, Claeyssens e.a. 2023). Similarly, data envelopes are being developed to provide context on collections in a machine-readable way to users before they interact with the data (Luthra & Eskevich 2024).
Notably, unraveling the concept of bias and adding practical information on – in first instance, as we will explain below: gender – biases in digital collections may contribute to the quite open questions that the data envelopes structure raises regarding “known biases in the resource” (see under ‘Level 3: Data (Content and Context)’). It is not further specified how this should be assessed, nor how ‘bias’ should be understood by users – including cultural heritage professionals, researchers and the like – filling out this information. In our view therefore, if cultural heritage institutions such as the KB could contribute to developing a way to indicate which biases might be present where in the collections, this would be a valuable tool to add to datasheets or data envelopes as a concrete manifestation of the slow approach and mindful of positionality. The latter is another core concept in our research, which will be elaborated in the next blog. The current blog now turns to our approach.
Approach in theory and practice
As indicated, the aims of the project are twofold, concerning ‘bias’ in theory and practice. Therefore, our team opted for a layered approach, i.e. working from meanings, manifestations and types of bias in the form of theoretical desk research (as carried out by us legal scholars), to practically training models to detect initially one type of bias in selected collections (as performed by our KB colleagues). Some notes on our approach are the following.
To start with, we should not only acknowledge our own positions that we bring to the design of and research for this project as an interdisciplinary team, but also the exploratory character of the research and its limitations. Limitations for instance regard the available time and means and the dependency on both the current state of the technology and the availability of digitised materials in Dutch. Consequently, we had to make choices regarding the scope of the first technological experiments: we chose to center on the KB’s digitised newspaper collections – no fiction – to map gender bias over time through word embeddings (cf. Hamilton e.a. 2018). An obvious limitation regarding gender, in turn, is that we were bound to the binary categories of ‘male’ and ‘female’, since a Dutch version of the Homosaurus has not yet been developed. Although we recognise that the DE-BIAS project included gender and sexual identity in its vocabulary, these terms were not on a separate list from the other themes of migration and colonial history and ethnicity and ethno-religious identity.
And then there is the limitation that word embedding technology does not enable researchers to trace patterns to concrete texts. Nevertheless, the word embedding approach might be extended to cover other forms of bias as it indicates the context in which words appear in relation to other words, hence provides information about associations, changing meanings and potential biases over time (Hamilton e.a. 2018). However, the time and resources it takes to train the models present another limitation. Also, further research would be needed in multiple directions. One direction is further developing a way of visualising these biases to inform users, but we hope to contribute to interdisciplinary dialogue on the responsible use of digitised collections also without a visual component at this time. Another direction is historical insight to justify the terminology used to train the models on identifying potential gender biases, which are now experimental in character to illustrate the urgency of the discussion. Historical wording, or even something as basic as spelling, may be lost in the hands of researchers from other disciplines, including the present authors.
It should be mentioned that, apart from the slow archives approach mentioned above, our work is positioned among – and builds on – other initiatives and concepts as well. For instance, our research can be seen as part of the work of the Cultural AI Lab: the lab is founded by the KB and other cultural heritage and scientific instititions with the aim to study “how to deal with cultural bias in data and technology”. This aim is concretised in projects such as ‘SABIO: the SociAI BIas Observatory’ (which, as opposed to our project, focuses on bias in metadata and assesses “how collection managers and curators create and add metadata to collection objects, and how bias in these metadata can be detected using statistical models”) and ‘AI:CULT’ (which centers on the “gap between AI and our digital cultural heritage”, paying attention to “the inherent richness, subjectivity and polyvocal nature of cultural heritage data” over time in light of the current state of AI technology, which can hardly deal with such subtleties). Tying in with these projects, we focus on ‘bias’ in the KB’s digital holdings.
Outlook
After this brief introduction of the background, aims and approach of the research, the next blog will highlight some findings stemming from studying the concept of ‘bias’, positioning our research and its aims in the sphere of ‘positionality’ and, while keeping the outlined limitations in mind, linking theory to practice in view of fostering interdisciplinary dialogue.
Acknowledgements
We would like to thank the KB team (Steven Claeyssens, Willem Jan Faber and Celonie Rozema), as well as all colleagues who brainstormed with us or suggested useful sources in various stages of the project. During the project, research ideas were presented at various occasions, including at the Cultural AI Lab, the Centre of Intellectual Property Law at Utrecht University and the KB conference on historical newspapers in the AI-era.
References
Henk Alkemade, Steven Claeyssens e.a., ‘Datasheets for Digital Cultural Heritage Datasets’ (2023) 9 Journal of Open Humanities Data 1-11
Kelly Breemen and Vicky Breemen, ‘‘Slow Libraries’ and ‘Cultural AI’: Reassessing Technology Regulation in the Context of Digitised Cultural Heritage Data’ (2025) Technology and Regulation 175-193
Michelle Caswell, ‘Dusting for Fingerprints: Introducing Feminist Standpoint Appraisal’ (2021), in Elvia Arroyo-Ramirez et al, 3 ‘A Radical Empathy in Archival Practice’: Journal of Critical Library and Information Studies (special issue)
Kimberly Christen and Jane Anderson, ‘Toward Slow Archives’ (2019) 19 Archival Science 87-116
William L Hamilton, Jure Leskovec and Dan Jurafsky, ‘Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change’ (2018) Association for Computational Linguistics
Donna Haraway, ‘Situated Knowledges: The Science Question in Feminism and the Privilege of Partial Perspective’ (1988) 14 Feminist Studies 3 575-599
Andrew Gary Darwin Holmes, ‘Researcher Positionality – A Consideration of Its Influence and Place in Qualitative Research – A New Researcher Guide’ (2020) 8 International Journal of Education 1-10
Owen C. King, ‘Archival Meta-metadata: Revision History and Positionality of Finding Aids’ (2024) 24 Archival Science 509-529
Mrinalini Luthra and Maria Eskevich, ‘Data-Envelopes for Cultural Heritage: Going beyond Datasheets’ (2024) Proceedings of the Workshop on Legal and Ethical Issues in Human Language Technologies 52-65
Eddy Ng et al, ‘The Anti-DEI Agenda: Navigating the Impact of Trump's Second Term on Diversity, Equity and Inclusion’ (2025) 44 Equality, Diversity and Inclusion: An International Journal 137-150
Written by Kelly Breemen and Vicky Breemen
Image by Alina Constantin / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/