Interview Clemens Apprich

„Let the Alien Machine Be Alien“ Media Theory, Errancy, and the Conjectural Logics of AI

Interview with Clemens Apprich

September 01, 2026

This interview was originally conducted on February 13, 2025.

Clemens Apprich is the head of the Department of Media Theory as well as the Peter Weibel Research Institute for Digital Cultures at the University of Applied Arts in Vienna, where he holds the Professorship for Media Theory and History. In 2023 he was appointed Vice-Rector for Research & Digitality at his University. He is guest researcher at the Centre for Digital Cultures at Leuphana University of Lüneburg, as well as an affiliated member of the Digital Democracies Institute at Simon Fraser University and of the Global Emergent Media Lab at Concordia University.

In this conversation, Clemens Apprich approaches AI through the competing and entangled logics of deduction, induction, and abduction. He connects machine learning to longer histories of computation, psychoanalysis, and media theory, and argues that media studies should neither reduce AI to a tool nor measure it exclusively against human cognition. The discussion turns repeatedly to errancy, conjecture, and the machine’s “alienness.” It concludes by contrasting two possible roles for AI in science: the automation of inherited patterns and the diagnostic exposure of the assumptions embedded in data and models.

Andreas Sudmann:  Clemens, your work has long engaged with digital cultures and, more recently, with artificial intelligence and machine learning. An example of this is your most recent book, Errant Intelligence, published by Routledge. Could you begin by outlining the research in which AI currently plays an important role for you?

Clemens Apprich:  In my research, I am treating AI not as a simple tool, but I wouldn’t go so far as to call it a medium either. Rather I am interested in the different logics and histories behind what we call AI today. From this perspective, AI turns out to be a misleading term, because all intelligence is artificial and has always been. From writing to radio to computation, intelligence comes to the fore through forms of artificial mediation. My media-historical approach therefore goes back at least to Turing, especially to his 1936 paper on computable numbers and the ‘Entscheidungsproblem’ and his 1950 paper, “Computing Machinery and Intelligence.” Turing does not approach the issue as an ontological question about whether intelligence is artificial or. He asks whether a machine can learn and what are the means to do so. That question really has sparked the whole project of machine learning, from which different branches emerged. While Symbolic AI typically works deductively – you have an input and explicit rules, and you derive an output – connectionism turns this process around: you have an input and an output, from which you induce the rules, which then can be applied to unseen data. Even though this inductive approach has become the dominant paradigm, fueled by an exponential growth of data, I would still insist that it is not synonymous with machine learning. A deductive system – nowadays called Good-old fashioned AI – is still a learning system, and Turing is very clear about this. What is more is that a purely inductive approach, as proposed by connectionism, which become more and more prominent since the 1980s, is about to hit a wall. This is also the reason why recently research projects, such as Bilateral AI in Linz, emerge that attempt to combine the two approaches: bringing symbolic and sub-symbolic AI together. At a commercial level, we can see that companies such as Google, OpenAI or Anthropic, have introduced what they call a ‘thinking mode’ or ‘reasoning mode’. Although these are not strictly deductive, they attempt to generate an explicit, visible sequence of intermediate reasoning steps before providing a final answer. However, from a media-theoretical and -historical perspective, I ask myself: why stop there? Alongside deductive and inductive logic, Charles Sanders Peirce identified a third type of reasoning: abductive reasoning. According to him, abduction is the only logical operation that introduces new ideas, and as such it precedes both deduction and induction. As a form of hypothesis building, it hints to the conjectural basis of all learning processes. Therefore, the interesting next step for me is not to ask whether a machine can learn, but rather how it learns.

Andreas Sudmann:  A few years ago, Matteo Pasquinelli argued that artificial neural networks cannot perform abduction in a strong sense. While they can, according to him, learn from established data and identify patterns in inherited data, they cannot produce something emphatically new. Do you agree?

Clemens Apprich:  I partly agree. Pasquinelli is arguing on the basis of inductive machine-learning systems. In a purely statistical sense, machine learning models cannot perform strong abduction. They can only perform what he calls a form of “weak abduction”, and, if I remember correctly, he links the inability to perform “strong abduction” to Umberto Eco’s account of metaphor. Accordingly, machines cannot create genuinely new metaphors because they cannot jump between categories. While I consider this to be a valid and important critique, but it is confined by a statistical, that is inductive understanding of AI and machine learning. And Pasquinelli is not the only one addressing this issue. Luciana Parisi, for example, has argued that an inductive system always produces something additional – a kind of surplus – when processing data. Drawing on Gregory Chaitin and related mathematical concepts, she describes the surplus as abductive. I agree with this point, but would like to take the idea further: abduction not just an additional form of inductive machine learning, but a different logic that could change the way we think about it. Perhaps abductive AI would constitute a different kind of AI system altogether. Whether this is possible remains an open question, though.

Andreas Sudmann:    For readers who are less familiar with Peirce, could you clarify what abduction means? It is a contested concept and is not universally accepted as a scientific form of reasoning.

Clemens Apprich:  True, but this already raises the question of how we define scientific reasoning. There is no clear-cut definition, of course, and even Peirce’s own account of it changed over time. His basic idea is a triad: deduction, induction, and abduction are not linear stages, but they are interrelated. In deduction, you have an input and a rule from which you derive an output. This resembles classical programming and the older form of syllogisms. Deduction can be a bit boring, because, in a sense, you already know the outcome. In contrast, in induction – a process associated with the rise of the experimental sciences – you know the inputs and outputs, and you try to infer the rule. However, the problem of induction, as David Hume already formulated it in the 18^th century, is that a rule can never be established as an absolute truth. It must remain open to revision or falsification – a fact that has become central to our modern understanding of science. Abduction is an entirely different matter. In Peirce’s earlier work, the distinction with induction is not always clear, but later it acquires its own domain. Abduction is usually associated with the formulation of a hypothesis or the famous “aha” moment. However, in a broader sense, it may also relate to the creative process in general, bringing it closer to the arts. It is something that emerges as a problem or something that intrigues you. From there, you can proceed inductively to test your hypothesis and deductively formalize it as a rule. All three operations remain connected.

Now this may be central to how we understand machine learning, because hypotheses are always part of the modelling process but rarely accounted for. When I model a system, I already have an idea of what it is supposed to do and of which data are worth collecting. Preconceptions enter at the beginning, but then disappear from view. Louise Amoore and Luke Stark, among others, have written about how such assumptions become hidden in technical processes. If a particular group of people defines the problems, their conscious and unconscious assumptions enter the model. This results in data and modeling bias. However, such conjectures should not be viewed negatively only. In fact, you could argue that they are constitutive to machine learning, as to any other form of learning, and as such they should be approached critically, but also creatively.

Andreas Sudmann:  Let us step back to a question that concerns both of us. What can media studies contribute to AI research today? What becomes visible from a media-studies perspective that other approaches tend to miss?

Clemens Apprich:  As you know, I work at an art university. Art has always been good in reflecting on its own medium. Media studies shares this capacity. Although they work with different approaches and means, both allow for a self-reflexivity on the medial processes they are working on. And I think this capacity is of utmost importance in the context of AI, as it enables us to make the processes involved in machine learning explicit. By looking at the histories of logic, mathematics, and computation, including the work of Turing and others, we can grasp how different logics operate within these systems and how they have evolved over time. This media genealogical understanding protects us from a short-sightedness that often prevails in the technical sciences: It is not just about the latest developments, such as neural networks, deep learning or LLMs, but about the underlying structures, such as language, symbolization or culture. So there is a historical reflexivity. Another characteristic of media studies is thinking not only about a medium, but also through it. This involves a hands-on engagement with the media technologies themselves, which is often lacking from a purely humanities’ perspective. Media studies can therefore mediate between applied engineering and computer science on the one hand, and cultural analysis or critique on the other. Bernhard Rieder, in his book Engines of Order, describes this as a middle layer. Bringing these two sides into contact is an interesting task, and one that has become very important in regard to AI and machine learning.

Andreas Sudmann:  It is also a difficult task. There is an ironic line, which I believe comes from Alexander Galloway, who said that the digital humanities are for computer scientists who do not understand literary studies and literary scholars who know nothing about computers. A media scholar who mediates between computer science and another scientific domain may need competence in three fields at once. Are we equipped for that? What genuinely media-theoretical contribution can we make to such translation, perhaps through an understanding of the media of science?

Clemens Apprich:  There are certainly many debates within media studies about which direction it should take. For me, again from the perspective of an art university, one contribution is speculation. I think we have a strong tradition of storytelling and narrative construction. While others see this as a scientific shortcoming, I actually see it as a strength. Sometimes, taking a narrative shortcut can reveal much more than the careful zooming in and out of the story. Kittler is a prime example of this. His books may not to be entirely accurate in terms of historical detail, but they certainly offer highly important theoretical insights. If we find ourselves caught in a particular paradigm – as I believe we are with AI and machine learning, which currently overemphasize statistical induction – media theory, particularly from a historical perspective, can remind us that things have been different and will be different again. There may be a completely different paradigm in ten, thirty or fifty years. Reaching it requires imagination. In this respect, media studies is similar to art. It can offer alternative narratives by speculating about the past and the future. This speculative part, which corresponds to the aforementioned conjectural basis of knowledge production, is absent from the so-called hard and techno-sciences, or at least not accounted for. However, this does not mean that media studies, or the humanities and social sciences in general, possess some form of innocent or pure knowledge that other sciences must follow. As Wendy Hui Kyong Chun has shown in Discriminating Data, knowledge from sociology and urban studies has been incorporated into today’s machine learning systems. In this sense, the humanities and social sciences are as complicit as computer and data sciences, and should therefore also be part of the discussion.

Andreas Sudmann:  We began by discussing induction, deduction, and abduction much as philosophers or historians of science might. What changes when we approach these concepts through the lens of media studies?

Clemens Apprich:  In the book I have just finished writing, I develop the concept of errancy, in order to shed light on the historical but also logical development of what we call AI today. It is a truly errant history, driven by a great deal of speculative energy that a conventional history of science might not fully grasp. The history of computation can be told as one of large-scale failure. From a Western perspective, it begins with Leibniz and his desire to construct a universal formal language. Later, Boole, Frege and others built on this project of formalization until it reaches its paradoxical limit, as expressed by Gödel’s incompleteness theorems. Turing, for his part, basically provided a proof of concept for Gödel’s problem, and, by doing so, invented the universal Turing machine. Thus, the failure of a centuries-old project became the conceptual basis for one of the biggest successes in modern history: computing. That errant history interests me, and I believe that media studies is best placed to work out the genealogical but also paradoxical details that remain important today. It allows us to sharpen our sensitivity to shifts, detours, and errant concepts.

Andreas Sudmann:  Media studies has historically concentrated on what Kittler might call the circuit intelligence of the computer, and thus on symbolic AI. Connectionism, as a form of adaptive learning with neural networks, was largely ignored, with a few exceptions. In Computer als Medium, for example, Norbert Bolz mentions sub-symbolic AI in the introduction and recognizes its distinctive features through work by Rumelhart, Hinton, and others, but does not develop a sustained media-theoretical account. Cybernetics remained central, including Claus Pias on McCulloch and Pitts, while the symbolic trajectory dominated histories of the computer, as in Bernhard Dotzler’s Diskurs und Medium. You seem unconvinced that artificial neural networks will determine AI’s long-term future. Looking back, however, media studies excluded an important mode of computational information processing. How would you explain that exclusion? Did it follow from computer science itself?

Clemens Apprich:  Computer science, which only emerged as a distinct academic discipline in the late 1950s and 1960s, had a problem with connectionism itself. While computer science is historically indebted to mathematics, logic, electrical engineering and information science, connectionism stems from data science and its proximity to experimental natural sciences. The former invokes deductive logic, while the latter follows an inductive paradigm. Despite these differences, there were, of course, overlaps. Cybernetics, for example, has always contained both approaches, and at the end of his famous 1950 paper, Turing mentions both of them and asks whether they might be combined. In recent decades, however, connectionism, or what you call sub-symbolic AI, has gained the upper hand, and I wonder what happened to the other approach. This is all the more important given that, precisely because of its history, connectionism is deeply rooted in a biological framework. Today, names such as ‘neural networks’, ‘Google Brain’, ‘DeepMind’, or ‘Cortex AI’ are a reminder of this history, implying that the intelligence lies somewhere in the brain rather than in learning, creativity, or social interaction. The tendency to biologize intelligence is, of course, an older phenomenon, already present in the work of McCulloch and Pitts, as well as in Geoffrey Hinton’s continuous attacks on symbolic AI. What bothers me is that this tendency has also started to shape some of the discourses in media studies, particularly around the concept of nonconscious cognition popularized by N. Katherine Hayles’s book Unthought. Neuroscience becomes a new reference system here. This is certainly remarkable when you consider that psychoanalysis, which would later have a decisive influence on media theory, at least in its German-speaking form, had already criticized this aspect of brain physiology over a hundred years ago. By ignoring this critique, we risk losing some of the keys that could help us to better understand today’s AI systems. As mentioned before, conjectures matter because unconscious assumptions are central to every learning process, artificial or otherwise. A critique of the current connectionist paradigm should therefore not simply return to a deductive approach or understanding of intelligence, but rather move towards new conjectures. What role does the symbolic play in these systems? Do LLMs perhaps alter the structure of language itself? How does our notion of the digital change with neural networks?

Andreas Sudmann:  This last point relates to my argument that artificial neural networks instantiate a post-digital, or quasi-analog, form of information processing. They still depend on digital substrates, but pragmatically their quasi-analog structure makes a decisive difference. This raises a broader disciplinary question. AI-based computational media studies could become important across many different fields. Should we primarily critique this development, or can media studies also embrace it?

Clemens Apprich:  The relationship between the analogue and the digital is a complex one. In fact, one could argue that this relationship constitutes human culture as a whole. So yes, of course it is also relevant to our discussions on AI and machine learning. I would, however, still insist that these technologies are fundamentally digital, whether in the data being processed or the logical structures of the processes themselves. In this sense, these systems are not so different from other cultural techniques where analogue signals are broken down in order to produce meaning, such as in music or writing systems. Yet 99 percent of the world might not even be possible to be discretized.

Neural networks can exploit that analogue aspect of our world to become more flexible and productive. Quantum computing in a sense also operationalizes this analogue counterpart. But also here, in order to control the computational processes, discrete instructions, in the form of high-level programming languages, are still needed. The age-old dream of wresting a discrete, and, therefore, controllable part of the world from it remains. In our discussion of AI, the term computational is probably more precise because it considers what happens when such discrete processes are no longer tied to our analogue senses. As M. Beatrice Fazi has shown, the Turing machine transcends our notion of aesthesis, that is the immediate, sensory and physical perception of the world. Cantor’s infinity, Gödel’s incompleteness, and Turing’s halting problem are all relevant here. Call it indeterminacy, incompleteness, the analogue, or something else. Computation reflects back on us, thereby transforming the way we think about the world. Computational processes – and, for me, AI is just the latest development in this regard – demonstrate that we require a discrete world in order to navigate it; however, in by doing so, we always fail to fully grasp it. Media studies has not always taken the profound impact of the computer on our culture seriously enough. However, new work is now emerging that does, for example at the Digital Theory Lab in New York.

Andreas Sudmann:  Another task for media studies may be to connect AI’s imaginaries with its material and technical operations. AI has always had a speculative dimension. The current discussion of AGI remains future-oriented and has been shaped by cultural industries, literature, and film. On the other side are infrastructures, technical systems, and empirical development. How can media studies relate these domains without reducing one to the other, especially in light of your work at the intersection of art and media theory?

Clemens Apprich:  When people talk about an AI revolution, or the emergence of a singularity, it’s useful to consider how the actual situation looks like in our daily lives. For example, a virtual assistant built on new AI technologies, can make a restaurant reservation for me. This is indeed impressive and also very handy, but hardly revolutionary. For me, there is still a big difference between imagining how technologies can enhance what we already have and actually speculating about what these technologies might bring to the fore. Media studies, in my view, should focus on the latter, by taking the perspective of the machine. Not in the sense of a post-humanism, as this is, as Leif Weatherby has recently argued in Language Machines, just the inverse of a still pertinent ‘remainder humanism’, but rather a true change of perspective, that allows us to learn something new. So it’s not about bringing human values back into the machine, but to advocate for a re-evaluation of those values. This is where the arts come in, as they do allow for a different epistemological approach, one not already constrained by the so-called exact or natural sciences. As an alternative form of knowledge-making, art can foster true speculation about the future. The task is to navigate between humans and machines, finding connections without erasing their differences. Art, in collaboration with media studies, can highlight technology’s alienness and otherness, rather than merely imagining the machine as a servant or assistant. It is this otherness, the difference the machine represents, that may tell us something new about ourselves.

Andreas Sudmann:  You still emphasize machines in relation to humans, although one could argue that statistical computation and human cognition belong to different domains. Neural networks were loosely modeled on neuroinformatic accounts of animal and human brains, but their genesis differs radically from ours. The Turing test concerns similarity at the level of observable performances. While machines can appear human, and humans can sometimes appear machine-like, the processes that produce these performances remain fundamentally different. One influential vision of AGI nevertheless defines it through superhuman or human-related capacities. Yet the digital computer, particularly the general-purpose computer in the tradition of Turing and von von Neumann, can be understood as a form of general intelligence in its own right. The dominant AGI imaginary instead projects a human-like agent with broadly distributed learning capacities. Even this does not resemble actual humans, whose learning abilities vary dramatically across domains. Why does AI discourse remain so obsessed with defining machines in relation to humans?

Clemens Apprich:  The problem arises when we refuse to accept the alienness of the alien machine. A true relationship can only come from difference. Gilbert Simondon understood this, and famously criticized our tendency to anthropomorphize machines by projecting ourselves onto them. This tendency can be seen in myths, ranging from the Golem to Frankenstein, but also in today’ virtual assistants with their often female voices. The alienness of the machine is difficult to grasp, and perhaps even more difficult to endure. Although human and machine cognition differ, they nonetheless operate within the same symbolic realm, which is also the reason why I do not believe in a singularity that would all of a sudden produce an entirely separate intelligence. Both humans and machines operate through language, and with LLMs, this common ground has become even more obvious. However, this common ground itself is being transformed, in particular due to the automation of language processing. This is where we should focus on, not least because it is also relevant to our understanding of learning. We should move away from narrowly Western, individualistic models of learning, which are now being reinvigorated within AI systems. In contrast, Soviet learning theory, especially Lev Vygotsky, has long emphasized ‘concept learning’ as a collective process. Humans do not individually learn pre-existing knowledge. We develop concepts together in places such as kindergartens and schools, and during conversations like this one. Those concepts remain in flux, a fact that is also of interest to machine learning. Google’s Gemma Scope, for example – at least as I understand it – attempts to identify shared concepts and their relational status within a model. Similarly, Adrian Mackenzie asks how humans and machines might learn together. In his view, we are all machine learners in this process. And psychoanalysis has long insisted on the centrality of language, reminding us that learning is not just about accumulating information, but rather a process of entering shared symbolic worlds. These worlds are increasingly becoming entangled with computational systems and machine learning models today.

Andreas Sudmann:  I would like to continue that line of thought, but let me close with the question we ask all our interview partners. How is AI changing science? Or does the question itself need to be reformulated?

Clemens Apprich:  Based on what we have discussed so far, I would first like to ask: what kind of science? I think this is necessary in order to understand how AI might have changed it. Following Jacques Lacan, we could say that there are, at least, two conceptions of science: the exact sciences and the conjectural sciences. While the former, often also called hard or natural sciences, emphasize the accumulation of knowledge, the latter, which I associate with the humanities but also the arts, are more concerned with truth. It is important to note here that knowledge does not automatically lead to truth, because truth is always the result of collective negotiation. And this is precisely the problem with today’s AI and data models, which too heavily rely on the exact sciences. Instead of negotiating what we want from these technologies, we establish a ‘ground truth’ based on conventions, and if the model does not reproduce this correctly, then we simply give it more data, adjust, and optimize it further. There is usually no discussion about the ground truth or conventions themselves, which are always based on conjecture. Even the hardest of all hard sciences is contingent on it. Therefore, rather than using AI systems to simply optimize already existing knowledge, we could use them to test and build new hypotheses. AI can help us reflect on the conjectures and assumptions that underpin not only scientific but also social processes. Rather than extrapolating patterns from past data into the future – which, by default, only foreclose that future – we can use AI models to diagnose the patterns in our data, the conjectures and assumptions we have built into it. We could then act on these patterns instead of merely automating them. AI may therefore push science, and society with it, in two opposing directions: It can either reinforce existing biases in our data and knowledge base, keeping us in an endless loop of the same, or it can help us to better understand those assumptions, enabling us to develop better ones. The choice is ours to make. In this sense, AI we might enter into a new truth game, one concerned less with discovering facts than with collectively negotiating the conditions under which facts become meaningful, relevant, and actionable.

Andreas Sudmann:  Hopefully. Thank you very much for the conversation, Clemens.

Clemens Apprich:  Thank you. I appreciate our dialogue.