An exploration of the use of AI in performance and new media art documentation within collections and archives

This joint exploration across disciplines and with non-academic partners, with potential for wide-reaching ramifications and innovation, aims to explore how AI could help conservators and archivists to deal with the burgeoning field of performance and new media art documentation. Documentation is usually carried out by a range of stakeholders at the point of acquisition of an artwork. This may be subsequently updated if an artwork is (re-)activated over time, illustrating that documentation is a broad, varied, and often generative field benefitting multiple stakeholders. Historically, the point of view of the artist has been the primary factor considered in generating a documentation though current practice also includes documentation of the audience experience and stakeholder considerations about the maintenance of a work. Overall, documentation has been an expanding field, both in terms of the quantity of documentation generated at the point of production and acquisition, and in terms of subsequent iterations of it. However, since several parties carry out documentation, and frequently do so with different agendas, the resulting growing number of (sometimes conflicting) documents has become increasingly onerous to manage. Here we show that AI could play a key role in art documentation, including in the case of AI generated artworks, but only if the roles played by bias and positionality are addressed and utilised. In AI research, bias refers to the skewed or distorted outcomes produced by machine learning and AI algorithms. This bias stems from two main sources: 1) human preferences in selecting specific generalizations or hypotheses over others during the algorithm design process; 2) potential distortions in the training data due to human predispositions or unrepresentative sampling. These biases can lead to inaccurate results that may not be entirely dependent on or justified by the observed instances in the data. Positionality refers to the choice of a specific theoretical approach. While it is commonly presumed that AI bias can reduce AI’s accuracy and effectiveness, and so bias, whether produced by data scientists or stakeholders, is often hidden, we think that making bias and positionality explicit will improve the reliability of a documentation. Moreover, artists often introduce bias into AI artworks by subverting AI technical methodologies, including machine learning pipelines, and some researchers use positionality which AI should reflect. Just as archives declare the provenance of their collections and researchers cite the origin of their sources, making the bias and positionality more explicit will not only make art documentation more precise but also make AI more efficient in that the datasets created as this field develops could be used to enhance future training. We aim to explore how AI could play a significant role in performance and new media art documentation provided that bias and positionality are made explicit. To this extent, our project, researched with stakeholders, will investigate the advantages, challenges, modalities, and consequences of making bias and positionality visible, primarily in the museum sector but with relevance to documentation practice more widely, paving the way for future research in the use of AI for art documentation.