Kick-off Webinar

Aims of the Cultural AI Lab

“What is the difference between Cultural AI and many other recently established cross-overs between computer science and the humanities, such as Human Centred AI/Ethical AI?”

— The Cultural AI lab is one of the new ICAI labs, and like the other ICAI labs we have multiple PhDs working on a common theme. What makes our lab unique is the key role for heritage partners and their collections in the projects we do, and our focus on addressing culturally sensitive problems in AI.

“What is the difference between 'Culture for AI' and 'AI for culture'?”

— 'Culture for AI' refers to what AI can learn from Cultural Heritage institutes and Humanities scholars. For example, how to make sure source data as well as data processing pipelines are trustworthy. 'AI for culture,' on the other hand, is the development of AI systems that work on the rich, subjective data of cultural heritage institutes. But the best answer to this question was given by one of the participants of the webinar on Twitter.

“Are there any similar multidisciplinary Digital Humanities projects of this scale?”

— The Netherlands has a pretty good track record when it comes to heritage/informatics/humanities collaborations, most notably through the Continuous Access to Cultural Heritage programme which ran for 10 years in which many of the Cultural AI staff were involved and take inspiration from. Abroad, the Living With Machines project is the first that comes to mind. I think one of the issues is that it is sometimes difficult to obtain funding as often research councils fund either humanities or computer science projects and it's rarer to see interdisciplinary funding, in particular for big initiatives, but there is a lot of great work happening in many places, check out for example
the University of Bologna and Finland.

Lab Structure and Operations

“Are you planning some 'introspection' phase where CH professionals (or others) would dive in the collections and produce some knowledge, possibly changing the scope of current cataloguing?“

— Absolutely! In all the projects we have formulated so far, we have planned workshops with the CH professionals to learn about their workflows and for them to provide input on our tools from the start. When we can go back to working on location, the idea is also for the researchers on the projects to spend a significant part of their time at the cultural heritage institutes to really immerse themselves in the day-to-day practie there and work closely with the domain experts.

“Is the Lab working with for profit partners, such as internationally operating technology providers?“

— We support a more general discussion, both about the potential harmful role of AI in society, and about the pros and cons of the current academic research funding ecosystem. We expect this discussions to continue and the Cultural AI lab is more than willing to engage in this discussion

“What data are you sourcing for the Cultural AI Lab? Is it mainly coming from the partners? And the heritage institutions share their data online together with the desired solutions?”

— Yes, much of the use cases and data is coming from the heritage partners. We also work with the Digital Heritage Network and Europeana. All results will be shared with the wider public, using open licenses by default.

“How are the cultural heritage organizations in this collaboration approaching adoption, modification, and sustainable use of AI?“

— The projects in the Cultural AI Lab have been designed in collaboration between researchers and cultural heritage partners. Evaluations will happen in the operational contexts of the heritage partners. We are fortunate that our founding partners KB National Library of the Netherlands, Netherlands Institute for Sound and Vision and Rijksmuseum employ expert staff that can help with
setting up evaluations.

Execution of the Projects

“One person's bias is another persons normal. How do you deal with that? Also: what about the bias you're inevitably putting in your algorithms that you are using in your project.”

— Actually, it is precisely our aim to raise awareness of biases that some might not recognise as such. Bias is of course a controversial issue and we cannot fully safeguard our algorithms against producing contentious outputs. Hence, we believe that transparent and explainable algorithms are core values for our project, so that users are enabled to retrace why a certain case of bias was detected, even if they disagree. These values can also help us deal with the problem of inevitably introducing bias through our algorithms, in the sense that users should be able to understand the intrinsic biases of our algorithms. We indeed need to take great care that we do not end up reinforcing existing social biases
or even being hurtful to users.

“How can these projects help to tackle data literacy inside the institutions and their staff? Because in order to discuss bias, we need their (internal) knowledge to work on the data/tools and to make it sustainable.“

— Research support services are vital for cultural heritage institutions when they want to engage with this kind of research. That also includes training to enhance digital and data literacy skills for internal and external users. It is also an area where co-operation between university libraries, knowledge institutes and cultural heritage institutions can be of real benefit.

“How are you dealing with/taking into account the quality of arguments and opinion“

— The "quality" of an argument or an opinion is in itself a separate multidisciplinary research topic. And there is no uniform solution to assess the quality of an opinion. In our research, we stress mainly the diversity of opinions or perspectives in heritage collections. We aim to retrieve different perspectives on objects in heritage data, despite their quality. Although, some of the perspectives might be biased, offensive, and discriminative. And yet they are a part of a collection, and they need to be studied to properly approach and represent them in a digital collection. For example, labelling discriminative words in object descriptions as inappropriate and providing explanations
(or counter perspectives) on them in data-stories.