By Hannah Peuker (Ruhr-University Bochum)

In current AI development, generative world models (e.g., Google Genie 3 or Meta V-JEPA 2) represent a highly dynamic paradigm that goes beyond language-based models (LLMs and chatbots). World models are developed with the goal not only of describing complex environments but also of making them usable for interactive simulation, the entertainment industry, and robotics. World models operate on large-scale video datasets and claim to reconstruct physical reality – with its entirety of rules and implicit laws – in a way that is consistent and all-encompassing. In this context, the idea that the shift from linguistic to visual representation relies on the assumption that the complexity of the perceptible world is, in principle, fully modelable and therefore algorithmically operationalizable. Representation promises a structurally identical reproduction of physical reality.

In my talk, I want to explore the extension of language-based AI systems to visually operating world models from three perspectives: First, I will provide a brief overview of the cognitive theory of internal representation that underpins the idea of technological world modeling. Second, I will introduce some of the philosophical frameworks used by AI researchers at Google DeepMind, Microsoft AI, Meta Platforms, and other major tech companies to conceptualize their shift to the world model approach. These references range from the idea of creating ‘a digital Plato’s cave’ to phenomenological views of perceptive intelligence inspired by Merleau-Ponty. Shedding light on these philosophical self-descriptions illuminates how Silicon Valley discourses incorporate philosophy into their understandings of technologically implemented intelligence and rationality. This line of thought, hence, presents a first thread of current techno-philosophical thinking. In the third and final, yet crucial, step, I will present a critical perspective on the ontological promises underlying the debate about world models as the next frontier for reaching Artificial General Intelligence (AGI). I will explore this perspective by connecting two techno-philosophical viewpoints: German media ontology after Kittler and French simulation theory after Baudrillard. By combining these two lines of thought, I aim to argue that the transition from language-based AI to visual world models rests on a paradoxical promise: that AI can attain an ontological status, with presumably immediate access to and intensive consequences for physical reality, precisely through the complete and self-contained visual simulation of that reality. In my talk, I want to ask how this ontological promise relates to earlier theories of visual representation in media and techno-philosophy. In this context, bringing Kittler’s media ontology into dialogue with Baudrillard’s simulation theory promises a multilayered analysis of the relation between our understanding of physical reality, technological materiality, and codified virtual worlds.

By Ivanna Gladysh (SacreCode, Prague)

Generative AI systems now produce artefacts semantically indistinguishable from human-authored content, raising a concrete problem at the intersection of philosophy of technology and AI governance: how to measure semantic authenticity in a principled and computationally reproducible way. Existing approaches — binary classifiers, provenance standards such as C2PA, and watermarking — operate on the level of signal or metadata and are systematically defeated by simple post-processing (re-encoding, screenshotting). None addresses authenticity at the level of meaning itself.

This paper introduces Δsim (Simulacral Distance), a computational metric grounded in Baudrillard’s theory of simulacra and operationalised through dense embedding geometry. Where Baudrillard diagnosed the proliferation of self-referential signs without originals as the cultural condition of the late-modern image, Δsim translates this diagnosis into a measurable geometric property: the angular distance, in a shared semantic embedding space, between a content artefact and the nearest point of an adversarially constructed cloud of AI-imitation descriptors. In its simplest formulation:

Δsim(W) = 1 − maxₖ { cos( E(W), E(Pₖ) ) }

where E denotes a pre-trained vision-language or sentence embedding function and { Pₖ } is the simulacral cloud constructed for W. The construction of this cloud — the protocol ensuring the cloud is genuinely adversarial rather than merely descriptive, the diversity controls preventing collapse onto a single stylistic mode, and the calibration procedure making Δsim values comparable across domains — constitutes the operative substance of the method and is the subject of the underlying patent application. This paper presents the philosophical architecture and the empirical behaviour of the metric; implementation specifics are reserved for the patent disclosure.

The contribution is threefold. First, Δsim is, to our knowledge, the first formally defined metric to operationalise a Baudrillardian concept as a computable geometric quantity, opening continental semiotic theory to empirical inquiry without dissolving its conceptual content. Second, the metric is model-agnostic and signal-invariant — independent of which generative system produced the content, and robust to the post-processing operations that defeat provenance and watermarking, addressing authenticity at an orthogonal layer. Third, Δsim is positioned within the regulatory context of EU AI Act Article 50, which from August 2026 mandates disclosure of AI-generated synthetic content — a framework presupposing the detectability of simulacral saturation without supplying the infrastructure to achieve it. Δsim is proposed as one such infrastructure: a bridge between philosophy of technology’s critical diagnosis of simulation and the computational demands of AI governance.

The framework has been developed as a patent-pending system (filed with the Czech Industrial Property Office, ÚPV; EPO priority window active) and a working prototype, enabling empirical demonstration of the theoretical claims alongside their philosophical analysis.

By Jakub Mácha (Masaryk University, Czech Republic)

Large language models (LLMs) produce remarkably fluent and coherent text, yet they remain prone to hallucination and other forms of error. In this paper, I argue that their characteristic strengths and some of their failures arise from the same source. LLMs instantiate a sign-first rather than a proposition-first account of language: their training objective and their architectural primitives operate on tokens and their relations, not on propositions. I distinguish four grades of unity: attentional, syntactic, semantic, and propositional. LLMs attain the first three, but nothing in their objective or primitives secures the fourth: the unity of a content that is true or false and remains the same across its different expressions. This explains hallucination, semantic drift, and quotation failure insofar as they consist in a content being lost when its wording changes. It also gives concrete form to the long-standing objection that structuralism and deconstruction leave the truth-relevance of language unaccounted for. Deconstruction correctly describes the instability that LLMs display. However, its inference from the iterability of marks to the instability of meaning holds only on sign-first premises and does not reach propositional content; the same conditional limits structuralism. Against the objection that propositional content emerges from token-level processing, I argue that propositional unity is not a statistical feature of text and is therefore not inherited from training data. The paper closes by asking whether a proposition-first language model is possible and where it would have to secure propositional unity.