By Mustapha El Moussaoui (Free University of Bozen)

The Western tradition has long privileged vision as the royal road to truth, an ocularcentric inheritance that phenomenology — from Husserl's zu den Sachen selbst through Merleau-Ponty's embodied perception to Pallasmaa's critique of the disembodied eye — has consistently sought to displace. This paper argues that the proliferation of AI-generated imagery produces an unexpected phenomenological consequence that completes, by technological means, what philosophical critique alone could not: the collapse of visual evidentiality and the opening of a post-ocular condition.

The standard critique of generative AI, indebted to Debord and Baudrillard, reads it as an intensification of the spectacle: more images, more hyperreality, deeper alienation. I propose a dialectical reversal. When every image can be hallucinated, when the deepfake renders any visual document suspect, the image loses its centuries-old claim to indexical authority. The mode of givenness (Gegebenheitsweise) of the photographic image shifts from evidential to suspicious. This is not an epistemological problem in the analytic sense but a transformation in the style of appearing itself: the image now solicits belief differently, and in doing so throws the subject back onto other registers of verification — bodily presence, lived encounter, multisensorial testimony, the situated Leib.

The deepfake does not betray the image's truth-function; it exposes the fact that the image was never the ground of truth to begin with. Photographic indexicality (Barthes's ça-a-été) emerges, retrospectively, as a modern myth that ocularcentrism naturalized. AI dispels this myth by negation, enforcing what phenomenology has argued for a century.

Drawing on prior fieldwork in Bolzano — where 80% of participants in a public urban-design experiment selected aesthetically spectacular AI-generated scenarios over a phenomenologically grounded one addressing real lifeworld concerns — the paper reads such preference as the residue of an ocular regime now entering crisis. The participants chose under the old style of givenness, in which the beautiful image still carried authority. That style is collapsing in real time.

The post-ocular condition does not, however, guarantee emancipation. It opens onto two divergent paths: epistemic nihilism (the post-truth retreat into algorithmic curation) or a renewed phenomenological aesthetics grounded in embodied, necessitous, situated appearing — what I have elsewhere called Phenomenacessity. Which path obtains depends on whether we cultivate the aesthesiological capacities required to receive appearance otherwise than through the eye alone.

The paper thus proposes that AI, far from being only an accelerant of the spectacle, constitutes a paradoxical phenomenological event: the technological enforcement of the very critique of ocularcentrism that phenomenology and aesthetics have long pursued. The question is no longer whether to move beyond the visual, but what style of appearing will replace it.

By Gina Negelmann (Friedrich-Schiller-Universität Jena)

AI alignment is often understood as the task of playing, and ultimately winning, the immensely difficult specification game (Krakovna et al., 2020). To prevent misalignment, such as reward hacking, we must specify our goals, values, and preferences to an AI system with sufficient precision. In the context of reinforcement learning, this means finding a reward function that closely corresponds to what we want.

However, the AI alignment specification game is not the first of its kind in the history of artificial intelligence. For decades, the phenomenologist Hubert Dreyfus wrote about the specification game that symbolic AI struggled with in his time. One of Dreyfus's central criticisms of symbolic AI was that something essential would inevitably be left out when attempting to explicitly specify human knowledge and abilities for AI systems. Can we really find rules that determine how to drive a car in every possible situation, or when an object in an image counts as a cat? Dreyfus's answer was no. For Dreyfus, symbolic AI was engaged in a game of specification, one that it was bound to lose.

While Dreyfus’s reflections are widely regarded as relevant for contemporary AI development (Schuering & Schmid, 2024; Reynolds, 2024), they have not been applied to the AI alignment problem. My central thesis is that in the age of deep learning we are witnessing a transition of the specification problem described by Dreyfus: it has shifted from the level of actions to the level of goals. It is this shift, and the emergence of a new specification game at the level of goals, that contemporary approaches to AI alignment are struggling with.

For Dreyfus, the specification game of symbolic AI was grounded in a fundamentally mistaken conception of intelligence, one that relied on the rule-governed manipulation of representations (Newell & Simon, 1976). By analogy, I therefore ask how human goals, values, and preferences are conceived in today’s specification game. I take this question to be of crucial importance, yet it receives comparatively little attention. Much of the debate focuses on which goals should be implemented in AI systems or on how such goals can be implemented, ethical and technical questions that I do not primarily address here (Gabriel, 2020). Rather, I ask the ontological question of what goals are in the first place. What conception of human goals, preferences, and values underlies current approaches? Are they understood as representations located somewhere in the mind, or should they instead be understood, in Dreyfus’s sense, as embodied?

My thesis is that human goals should be conceived as embodied, context-dependent, and situated. It then follows that an important part of the alignment problem concerns the question of how such embodied goals could be incorporated into disembodied machines. In Section 2, I present several of Dreyfus’s criticisms of symbolic AI and show how they reappear in the context of AI alignment. On this basis, I argue in Section 3 for extending Dreyfus’s conception of human intelligence as embodied, context-de-pendent, and situated to human goals, values, and preferences.

By Thomas Schimmer (Mainz University of Applied Sciences)

Since its founding by Edmund Husserl, the tradition of phenomenology has engaged with technology and science. While Husserl focused on a critical reflection on modern sciences and the ideological positivism arising from them, other thinkers such as Heidegger and Henry adopted a skeptical stance toward modern science and technology. Against this backdrop, phenomenology can also provide important, critical insights into digitization and artificial intelligence (AI), a fact that has already been taken into account to some extent.

The proposed paper focuses on the significance of AI as a tool in science. Science has always generated and benefited from technological innovations. Digital tools – e.g. for the collection, processing, and analysis of research data – as well as computer technology (word processing, calculation, modeling, literature management, etc.) and application areas have enhanced the performance and quality of research. Many of these technologies have their origins in scientific problems and collaborative research – among these, the internet, as a solution to exchange information between research institutions, represents the elephant in the room. In 2020, the Deutsche Forschungsgemeinschaft/German Research Foundation (DFG) advocates for the use of digital technologies, but also acknowledges the need to reflect on new “research practices”. Two years later, OpenAI will introduce ChatGPT. AI models are also rapidly gaining traction in research in nearly all disciplines and DFG opened up to AI since 2023.

A phenomenological analysis of AI as scientific tool encounters a point that is critical to our understanding of scientific rigor. In his late work Crisis, Husserl reconstructs the emergence of the modern sciences from the interplay of practical lifeworld knowledge, experience, measurement, theory-formation and abstraction. Philosophy as well as the increasing perfection of measurement and theory-formation of sciences constitutes the episteme as certain scientifical knowledge in contrast to doxa as the practical lifeworld knowledge based on subjective experience (e.g. regarding the physical ‘habits’ of things). Although the elevation of the abstract world of scientific thought obscures the subjective world as ideologically positivism, with negative consequences to subject and society, the distinction between episteme and doxa nevertheless marks the difference between certain scientifical and practical lifeworld knowledge. The advent of AI into scientific methodology reveals a shift in meaning that is dangerous in the Husserlian sense. For AI models are not based on theory-building, but on algorithms that evaluate a learned dataset. In other words, AI calculates results based on its data knowledge. And this ultimately means that AI carries certain elements of doxa into episteme.

The paper reflects on this hypothesis and discusses advantages and dangers of AI for scientific self-claims, also in respect of the claim of the „end of theory“, as well as the consequences for everyday life when powerful doxical elements enter sciences via AI and become taken for granted.