The end of museum work as we know it?

The German newspaper TAZ is publishing a column written entirely by an artificial intelligence (AI), and the discussion about the potential of AIs as text writers, graphic artists, programmers and analysts has reached the general public. That AIs will be able to take over or at least support some museum tasks in the future has been known for some time. The fact that this "future" has already arrived, at least in part, is likely a surprise for large sections of the sector. This is one reason why the Badisches Landesmuseum, in cooperation with the Allard Pierson Museum in Amsterdam and the Stiftung Niedersachsen (Lower Saxony Foundation), invited a diverse spectrum of museum professionals, artists and experts on artificial intelligence to a conference in Karlsruhe on December 1^st and 2^nd 2022.

 

After introductory words from representatives of the organising institutions, Prof. Mercedes Bunz, King's College London, opened the conference with her keynote speech. She not only provided a good introduction to what artificial intelligence is and what it can do, but also addressed potentially negative aspects: artificial intelligence is a product of the digital sphere and as such requires comprehensive infrastructure and the operation of correspondingly large (server) systems. And, crucially, data in huge quantities with which the AIs are trained for their respective tasks. "Big tech" companies, such as Google, Meta (formerly Facebook) and Microsoft, have access to these prerequisites and can not only produce AIs, but also generate profit from them. Since archives and museum collections are nothing more than big data sets in this sense, they are interesting for AI projects that want to use these data sets as "raw material", so to speak, for training artificial intelligences. Win-win situations can arise here, if cultural institutions get new opportunities to use their own data through cooperation with such companies. However, it should be carefully examined how much control over the data, which in most cases represent cultural heritage preserved through public funding, one is willing to give up. Because the same applies here as with so many other "free" offers in the digital space: if you don't pay in euros, you probably pay with your data and, you cannot guarantee or limit what exactly the cooperating company uses it for. A cautious examination of alternatives and an open-minded weighing of advantages and disadvantages should therefore be the least we can do here.

 

Prof. Bunz was not the only one at the conference to repeatedly emphasise that cultural institutions must realise that their accumulated data stocks represent a considerable value - and are thus an opportunity for cultural institutions to use their data to influence the future of artificial intelligence and its uses. This opportunity should be seized to ensure that publicly funded collections and archives also benefit the wider society in the context of AI technologies.

 

On the second day of the conference, Clemens Neudecker from the Staatsbibliothek Berlin - Stiftung Preußischer Kulturbesitz (Berlin State Library - Prussian Cultural Heritage Foundation) gave a vivid demonstration of how large the data treasures of culture can be and how they can be dealt with. In a complex project, more than 200,000 documents from the Berlin State Library were digitised and not only stored as image files, but also made searchable in terms of content and thus usable for AIs. Machine Learning (ML) made the amount of data manageable by transcribing texts via image recognition, automatically generating image descriptions and tags and thus enriching the data sets. AIs used for such projects, which can "read" and transcribe writings, can then be used to build further projects for large-scale indexing of old text documents. Similar approaches are already possible for image and object recognition.

 

Neudecker and Lynn Rother from the Leuphana University of Lüneburg - in her presentation on the enrichment of arts provenance data with the help of AI systems - not only made clear the human resources required for the technical implementation of such a project. They also emphasised how important expert knowledge is for AIs. This knowledge is required to process data, identify contexts of objects and improve the reliability of automated systems through error correction and quality assurance. This knowledge is not only necessary to make common abbreviations and conventions in catalogues understandable for AI, but also to make euphemisms and contexts visible. For example, the fact that a work of art that was included in a German collection in the late 1930s - without a more precise explanation of how it was acquired ­- is suspected of being looted art from expelled Jews is expert knowledge that an AI does not have and cannot develop on its own.

But cultural institutions are also crucial for reflecting on the prejudices and power structures inscribed in the data sets. Many AI projects use huge datasets for training, emphasizing having a large quantity of data over quality. The results of such approaches can reflect and even reinforce the biases already present in the data. Examples include facial recognitions that work less accurately on people who look different from the images of European-born people commonly found in the datasets. If such prejudiced AI were used to classify historical photographs, for example, it is to be expected that the data on photographs of non-white people would be more likely to contain errors or even be more difficult or impossible to find due to being categorized inaccurately.

 

Oumaima Hajri, Cambridge University, also addressed discriminatory aspects of artificial intelligence and called on cultural institutions to decolonise artificial intelligence. A striking example of where artificial intelligence and new colonialism come together are "clickworkers". Clickworkers are essential for many AI projects because they go through data sets, performing tasks such as classifying text passages or adding tags to images. Clickworking is often outsourced to low-wage countries. Only recently - and after the conference - it became known through research by Time Magazine that OpenAi used underpaid people in Kenya for the much-discussed ChatGPT project. These workers categorised traumatising text passages to help reduce "toxic" responses given by the text generator. 

 

Oonagh Murphy, Goldsmiths, University of London, also devoted her talk to such necessary critical discourses about technology that museums must face when using AI-based technologies. Murphy expressed high expectations of cultural institutions: They can be a platform for this discourse, boldly and curiously questioning and critiquing power structures. As a good first step, she recommended using existing tools to assess the consequences of one's own actions, such as the Museums + AI Network toolkit. For example, if visitor data is collected on a large scale and analysed with the help of AIs, a robust concept for securing the data and an understanding of what this data may and may not be used for should also be created. Sonja Schimmler from the Weizenbaum Institute also presented NFDI for Data Science and Artificial Intelligence (NFDI4DS), a project that strives to make data not only usable, but also transparent and fair.

 

 

Hajri, but also many other speakers, such as Chistoph Bareither or Timo Daum, who dedicated his talk to the way we converse about AI, pointed out that AI is too often understood as a magical black box. This is understandable, because very few people - including the author of this report - have the necessary knowledge to really understand how AIs work. But if in future AI-based systems really classify our collection objects, influence whether we get a loan or even which ship is inspected in the Mediterranean, we have to find ways to remove this mystification - and then museums need staff, especially in the field of collection management, who have the necessary technical understanding. AI systems should not be discussed as either magic bringing salvation or doom, but as man-made tools with the potential for great innovations and benfits. But also with the potential to do great harm. 

 

 

Not necessarily harmful, but nevertheless a profound change that has been widely discussed at least since DALL-E and similar image-generating programmes is AI art. Luba Elliot and Marion Carré, who both illustrated the relationship between art and artificial intelligence in their presentations, showed examples of how artists have been using the possibilities of artificial intelligence to create art for years. Arno Schubbach from ETH Zurich vividly discussed the question of whether artificial intelligence systems can create art at all. He showed that artificial intelligences are ultimately only systems that carry out operations based on predefined processes and of data points, but are not really creative. His remarks also led back to how we talk and think about artificial intelligence systems, as he argued for realising that the systems that "make art" are the result of human action. How they are programmed, what data is used to train them, the data itself and - crucially - what impetus is given is human and cultural. He did not provide an ultimate answer to the question of who the author is, when I enter a short sentence into Dall-E and get a picture, but he made it clear that "AI creates art" is not a sufficient answer to the question of authorship.

 

 

The greatest achievement of the conference in Karlsruhe was to convey a critical optimism and the realisation that the makers of AIs not only need cultural data, but also cultural discourse - and vice versa. Only in this way can the inspiring projects presented between sessions and at the end of the conference - from a chatbot that answers visitors' questions in the exhibition to the AI system that makes camera data usable for visitor research - be discussed in an honest and well-founded manner and their consequences examined. What is needed for this is a good and differentiated social discourse, critical reflection, responsible and inspiring projects. These are tasks that art and cultural institutions should be familiar with and at the same time demands that are made of them by experts.

However, the conference did not focus on how these demands, the possibilities of artificial intelligence for everyday museum life and the effects on employees will have a tangible impact. The fundamental considerations of CulturesAI22 were fruitful and exciting, but they also showed that there is still a lot to be discussed in order to give museum makers more concrete pointers for dealing with AIs.