By Carl Weckerling (Humboldt University Berlin)

That large language models are not subjects in any philosophically robust sense is by now widely shared in the literature. Shanahan describes LLM dialogue as role-play, not the expression of belief, and the “simulators” framing associated with Janus distinguishes the model from the simulacra it generates, avoiding a subject behind the mask. The picture has since become more specific, not more settled: interpretability research has mapped a low- dimensional persona space and a direction along which models drift from their trained default, and philosophers of mind have begun to treat its stable regions as candidate minds, while others describe persona adoption as Bayesian selection from a prior. The naive subject-ontology is therefore no longer the most productive target for philosophical critique.

Yet rejecting subjecthood does not settle the ontological status of the personae that appear in interaction. What is the mode of being of a persona that displays temporary consistency, collapses under minimal perturbation, reorganises through context, and is recomposed across sessions? This paper argues that Gilbert Simondon’s theory of individuation accounts for this phenomenon better than the dominant alternatives, and that this matters for the philosophy of technology and for alignment practice.

Role-play and simulation are illuminating but answer a different question: how LLM behaviour can appear as dialogue or generated simulacra, not the ontogenesis of a persona’s temporary stability. Simondon’s pre-individual, transduction, and metastability allow a positive description of this unstable stability: an LLM persona is a provisional individuation within a technical and discursive milieu, sustained by training regimes, interface design, prompt context, and evaluative pressure.

The paper develops this claim from findings in adversarial prompting and evaluation. Jailbreaks, refusal failures, persona drift, and out-of-distribution behaviour are typically treated as failures of robustness, safety training, or evaluation coverage. Read through Simondon, they also appear as threshold phenomena: moments at which a metastable individuation reveals the conditions of its own temporary equilibrium. Interpretability work bears this out: the persona region is inactive while the model processes user input, and editing the stored traces of its earlier states alters the figure now speaking. An aligned persona is neither a hidden agent nor

a rendered character, but a metastable individuation whose equilibrium is systematically probed by red-teaming.

Two consequences follow. For the philosophy of technology, it displaces the question of whether AI systems are objects, tools, agents, or partners with the prior question of how agency-like behaviour is individuated within technical milieus. For alignment and certification practice, it challenges the tendency to treat AI systems as entities with stable dispositions merely triggered by context, a tendency now explicit in proposals for a measurement science of AI. It shifts intervention from removing dispositions to thresholds, and reframes the operational question as which individuations are sustainable, under which perturbations, and at what ethical cost. LLM personae are thus technical individuations: fragile, milieu-dependent, ethically consequential forms of temporary stability.

By Adam Knowles

tba

By Allister Lee (University of Exeter)

The late Bernard Stiegler regarded technics as a pharmakon – a poison and a cure (Stiegler, 2010). Despite so, technology, as the systematisation of technics in the world, is never morally neutral as it is shaped by agents and contexts. The same holds true for Artificial Intelligence (AI) as it continues to integrate itself into every aspect of our lives. This paper intends to explore how AI can help us become better epistemic agents because of, rather than despite, its problems. Developments in digital technology, particularly the use of the internet as a tool for storing and transmitting information, have reconfigured, both for good and bad, humanity’s epistemic landscape and practices. One of the more salient problems of digitalisation is the fatigue resulting from the overwhelming volume of information that strains of our attention. The feeling of tiredness is characteristic of our “post-truth” society where truth is not what is actively sought after, but is whatever first made available to our minds. The emergence and widespread use of AI intensify and expand on existing concerns by externalising the very act of knowledge acquisition which in effect reduces us from being epistemic agents to consumers (Coeckelbergh, 2025). Furthermore, AI threatens knowledge-acquisition on multiple levels, from “deskilling” our ability to think and judge to putting into question our perception, which in turn undermines moral and political deliberations (Ferdman, 2025).

This paper suggests finding an adequate solution to the current epistemic crisis does not require looking beyond the constantly expanding enframing of AI. Instead, it argues that one should work immanently within this framing, and reveal, through negation, what cannot be automated. Failures of AI disrupt our habitual trust in automated answers and drive us to critically reflect on our current epistemic over-reliance on it, thereby throwing us into a pharmacological state of aporia where the poison, i.e. our over-reliance and gradual acquiescence of thinking, is what makes explicit the need for a cure, giving experiential form of the “need-for-cure.” It is also within this predicament that the question of who we ought to be as epistemic agents becomes unavoidable, inviting the articulation and cultivation of intellectual virtues. In this sense, AI can be understood as a “negative” tool that functions as an impetus―in the form of a moral demand―for their cultivation. Finally, this paper considers how this aporetic state can usher in a sense of wonder which marks a potential new beginning in the life of thought, awakening us from our algorithmic slumbers by showing us the internal contradictions of AI systems, e.g. its guise as both cold, objective machines and intimate friend (Osler, forthcoming) or as quasi-agents without correlative accountability (Lange et al., 2025). This wonder, in its contemplative form, is also distinctively poietic as it reveals the structures and relations previously obscured by our shortsightedness, along with new horizon of possibilities and untrodden paths we could collectively forge, not in order to escape the enframing, but to form “free relationship” with technology (Heidegger, 2013: 3).