Interview Felix Stalder

Generative Systems and the Politics of Knowledge

Interview with Felix Stalder

August 19, 2026

This interview was originally conducted on January 20, 2025.

Felix Stalder is a professor at the Zurich University of the Arts, teaching in the Department of Fine Arts „Art:ifical Studies“ major. His research focuses on the intersection of cultural, political and technological dynamics. In his work he deals with commoning, copyright, datafication, AI, and the transformation of subjectivity. Felix Stalder is also a member of the World Information Institute and the Technopolitics Working Group based in Vienna, while also moderating nettime, a fediverse node for the discussion of critical internet culture.

Felix Stalder examines AI as an infrastructural and political formation whose effects exceed the familiar question of whether machines are intelligent. He argues that the generative dimension is analytically decisive because these systems reorganize knowledge claims, redistribute institutional capacities, reshape labour, and intervene materially in the world. The conversation connects this perspective to his account of digital culture, especially the growing dominance of algorithmic processes within meaning-making. It also addresses the limits of AI ethics under extractive economic conditions and the epistemic authority granted to quantified knowledge. Models may enable interdisciplinary work and new forms of scientific creativity, yet their blind spots can narrow what remains perceptible or politically actionable. For Stalder, critical inquiry therefore begins with the infrastructures and dependencies through which AI acquires social force and redistributes power.

Andreas Sudmann: You began working on digital culture and network technologies well before the current AI boom. When did AI first become relevant to your research, and how did that interest develop?

Felix Stalder: I became interested in digital culture and internet networks in the early to mid-1990s. AI was always somewhere in the background, or in the more distant future. We were still in the second AI winter, yet AI remained one of those grandiose technological promises. In early internet culture, people often said that the internet was as important to civilization as the invention of fire. AI was part of that horizon, though it was not a focus of my research. I knew some of its history, from its beginnings to expert systems, which were decidedly unsexy. I was more interested in the political dynamics surrounding the use and expansion of network technologies. This changed around 2011 or 2012, when neural networks re-emerged and deep learning began to produce transformative advances in image recognition. Something that had long seemed rather dusty was moving closer to an immediate reality and becoming a technology capable of shaping society. My perspective is strongly informed by science and technology studies. I am interested in how technology shapes society and how society shapes technology, as mutually transformative processes that cannot be separated. It seemed clear that something new was happening. Social networks had matured and stabilized to some degree, although this was still before the Snowden revelations. My attention then shifted again. Snowden exposed a hidden layer of network infrastructure and substantially changed how we understood these technologies. At around the same time, blockchain also gained momentum, although I personally never liked it and always thought it was a bad idea. At the time of this interview, Donald Trump had launched his own cryptocurrency and was returning to the presidency, so it seemed unlikely that blockchain would disappear soon. With the emergence of Transformer-based language models around 2018, it quickly became clear that something consequential was happening, even if it remained, and still remains, difficult to say exactly what that was.

Andreas Sudmann: When you refer to developments around 2018, do you primarily mean Transformer models and earlier generative systems, rather than the public release of ChatGPT in 2022?

Felix Stalder: Yes. I mean earlier developments, including Google DeepDream and generative adversarial networks, as well as the Transformer-based language models that emerged around 2018. It was becoming clear that this was more than another increase in speed or efficiency. The development consolidated with ChatGPT two or three years later and then expanded into a much broader public sphere.

Andreas Sudmann: You completed your PhD in Toronto, a city now closely associated with both media theory and AI research. How did that academic environment shape your perspective on AI?

Felix Stalder: When I was in Toronto, Geoffrey Hinton’s research was taking place just up the road, so to speak, but it did not really percolate beyond his labs or the specialized domain in which he worked. That was simply not my field, and I did not hear about him despite being in the immediate vicinity of his institution. More important for me was my own work in media theory at the McLuhan Program in Culture & Technology, even though that tradition already seemed somewhat dusty at the time. I nevertheless saw an affinity between McLuhan’s ideas about the ways technology shapes society and actor-network theory. McLuhan used the relation between figure and ground. In a sense, an actor-network describes the ground from which an effect, the figure, emerges. My approach developed through a combination of these perspectives, bringing media theory into relation with science and technology studies, although I would not describe myself as an STS scholar anymore. Experimental art also became important because it brought questions of meaning, and of the constitution and transformation of meaning-making, into this framework. I began my PhD in 1996 and completed it in 2001, during the early dot-com boom in North America, which was very present in Toronto. I then held a postdoctoral position with the Surveillance Project in the Department of Sociology at Queen’s University in Kingston, Ontario. The project examined surveillance very early as a constitutive mechanism in the governance of society. This attuned me to questions that later became widely discussed under the heading of surveillance capitalism and that are now central to debates about AI.

Andreas Sudmann: Looking back at your work, especially The Digital Condition, which arguments remain particularly useful for understanding how AI is changing science and digital culture more broadly?

Felix Stalder: One central argument in The Digital Condition is that meaning-making in digital media environments proceeds through three processes. The first is recombination, or remix, rather than original creation. The second involves epistemic communities, while the third consists of algorithmic processes of pre-sorting. At the time, Google Search and social media recommendation systems were the main examples of systems that pre-sorted the world before our own cognitive processes could begin. These selections then passed through communities, where they were shaped and consolidated into smaller or larger forms of consensus about meaning, importance, and desirability. I still find this framework suitable for analysing AI. When I wrote most of the book in 2015, I regarded these processes as roughly equivalent in weight. Today, the algorithmic dimension has clearly become dominant. Yet the emergence of something new still depends on recombination, now increasingly at a statistical level. Processes in which communities negotiate positions and evaluate criteria can also be observed within algorithmic systems. Where no single position can decide, local criteria of assessment become decisive. Processes that were once more distributed have moved into the algorithmic domain, along with the actors involved in it. This dominance is visible in the rankings of the world’s largest companies and wealthiest individuals, whose fortunes derive from these processes. The framework therefore helps us understand how social relations reappear within seemingly technical processes, where they are extracted, reformatted, and embedded again.

Andreas Sudmann: What does it mean, in your view, to study AI critically, especially in relation to its use across academic fields?

Felix Stalder: I published an article yesterday titled „Das Generative der Künstlichen Intelligenz.“ Its main point is that neither the term „artificial“ nor the term „intelligence“ is particularly helpful analytically. We do not know what intelligence is, and the word has an evocative quality that helps explain why „artificial intelligence“ has retained such a long technological and utopian shelf life. Is intelligence defined as the one-to-one replacement of human capacities? Are only humans intelligent? Is machine intelligence comparable to human intelligence, or is it a different form? If it is different, what umbrella concept brings the two together? Analytically, this is a mess. The same problem applies to the word „artificial.“ The more closely we examine the material infrastructures that make AI possible, the clearer their deep natural roots and repercussions become. I therefore propose focusing on the generative dimension and asking what is being produced. This includes content and knowledge systems. How do we come to know something when an answer appears as an oracular pronouncement from a black box, and what kinds of knowledge claims result? It also includes political economy. AI can transform hierarchies among academic fields and create new dependencies on commercial infrastructures, which affect what researchers are able to do. Beyond that, it produces powerful ideologies about what these systems are and what they enable. AI has become a promise that climate change might be resolved without fundamentally transforming society. At present, the industry acknowledges that AI requires enormous amounts of energy and may delay the transition towards cleaner sources. It then promises that a future, vastly more capable AI will help clean up the resulting mess through geoengineering. This is a powerful way of responding to a problem that has become increasingly difficult to ignore. None of these questions depends on deciding whether the technology is truly intelligent or in what sense it is artificial. A critical analysis should examine what AI is generating in the present. That’s difficult enough.

Andreas Sudmann: Are we also witnessing a shift in authority from interpretation and critical inquiry towards forms of knowledge that present themselves as data-driven and purely analytical?

Felix Stalder: Ideologically, yes. Critical thinking and interpretation clearly have a difficult time, while the seemingly factual character of quantified knowledge grants it immense authority. Yet close analysis of how these technologies are developed reveals assumptions and theories, often operating at the level of folk theories about how society works and what its objectives should be. Interpretation never disappears from the technology itself. In terms of discursive authority, however, there has been a clear shift.

Andreas Sudmann: Much of the current debate is organized around AI ethics and the search for guidelines for responsible use. How do you assess this discourse, especially where it incorporates critical perspectives while still aiming to distinguish desirable from harmful forms of AI?

Felix Stalder: My lawyer friends in this field like to say that ethics is for people who do not want to be punished when they are caught cheating. That is how they distinguish ethics from law. I am honestly dubious about much of the current discussion. The main engines of AI development are highly extractive processes that redistribute wealth upwards into very few hands. Wealth brings power and decision-making capacity. We have seen this in technical infrastructure for some time, and it is now becoming visible in other forms of power through the current oligarchic movement, especially in the United States. Given the underlying structure of the industry, I doubt that sprinkling ethical dust on top will restrain these developments. This can be observed at many levels. The internet has long had more or less established conventions indicating which data a crawler may access. A robots.txt file can ask crawlers to exclude parts of a website or the entire site. Major search-engine crawlers generally appeared to respect such instructions, at least as far as we could tell. The most basic ethical principle would be to refrain from using data without permission. Yet the current generation of crawlers often overwhelms websites and retrieves the same material thousands of times even when explicitly instructed not to do so. In this gold-rush moment, there is little willingness to accept restrictions. The political backlash against supposedly „woke“ positions also expresses a broader refusal of democratic oversight, respect for minorities, or limits on appropriation. At an abstract theoretical level, ethical debates remain valuable because we will eventually need shared ideas about how to deal with these technologies. As instruments for curbing current excesses, however, I regard them as largely ineffective.

Andreas Sudmann: What forms of interdisciplinary collaboration are emerging through the use of AI in science? Can the humanities and natural sciences be connected more closely through these systems, and should that be an objective in itself?

Felix Stalder: Any problem beyond a certain degree of complexity requires collaboration among specialized disciplines that understand the internal logic of different actors. Research on climate change may require expertise in atmospheric models, water, soil, or other environmental processes. It also requires sociologists who understand the logic of human actors and economists who can analyse the behaviour of firms. Agency is distributed across a multi-actor world, and understanding the interactions therefore requires a complex set of perspectives. Even before AI, modelling was one way to study such complexity. Modelling languages provided common ground on which quite different disciplines could work together. If you want to model traffic in a city, you need distinct forms of knowledge. The fluid dynamics may contribute something, but it is insufficient. The technical structure of the model establishes a shared language and requires each discipline to reformat its knowledge so that it can enter the model. This has been part of complexity studies since cybernetics. To some degree, AI can also function as a common frame of reference and as a shared material ground into which specific forms of knowledge must be translated and transformed.

Andreas Sudmann: Turning to artistic practice, can AI contribute to artistic creation within scientific contexts? What forms of expression or inquiry become possible when artists and scientists work with these systems?

Felix Stalder: From the perspective of art, the output of generative models is often rather boring. The visuality produced by image generators is too generic. Fine art is oriented towards concerns quite different from easy reproduction, although the situation may be different for illustration. What I find more interesting is the capacity to move between modalities, from text to image or from text to video. This can provide different forms of understanding and different epistemological lenses on a shared data object. That matters because the current shift from critical and analytical understanding towards statistical understanding produces an enormous flattening of our worldviews. What cannot be expressed quantitatively risks becoming inexpressible and losing its aesthetic quality, in the sense of its perceivability rather than its beauty. A new hegemonic perspective is emerging that is narrow, powerful, and limited. Its power derives partly from that limitation, yet it also creates blind spots. If these blind spots remain unacknowledged, they can lead to significant forms of violence and unintended consequences. Things may be eradicated because they are not seen, understood, or rendered useful. It is therefore crucial to insist on other forms of knowing outside AI while also creating a multiplicity of epistemological approaches within these systems.

Andreas Sudmann: To what extent can AI promote creativity and innovation in science? Can it already contribute to new research questions or unconventional solutions?

Felix Stalder: AI offers a different way of understanding and handling data, including different forms of aggregation. I would be surprised if this did not also produce a changed understanding of the world that the data is supposed to represent. Certain degrees of complexity are very difficult to grasp outside complex models. At this point, I would not draw a sharp distinction between conventional models and AI models. The central question is how we gain access to complexity. Many processes cannot be predicted, although they can be modelled. This has already proved powerful in climate modelling and politically difficult at the same time. Politics often fails to understand this type of knowledge, or finds it convenient to ignore it. Models may produce claims with ninety or ninety-nine per cent certainty, but they will never provide complete certainty. This uncertainty is built into their epistemological basis, especially in Bayesian statistics. A one per cent probability does not mean that an event cannot occur. If political actors want to disregard a finding, the remaining uncertainty makes that easy. For model-based knowledge to feed back into political processes, we need a different understanding of these claims and a better acceptance of what they can and cannot do. The question of what we can know about the world through models is inseparable from the question of how that knowledge enters collective decision-making.

Andreas Sudmann: Felix Stalder, thank you very much for the interview.

Felix Stalder: Thank you.