Thomas Metzinger's Self-Model Theory: What Minimal Phenomenal Selfhood Requires for AI

Thomas Metzinger (Johannes Gutenberg University Mainz) advances a position that is both philosophically rigorous and practically specific: there is no such thing as a self. What exists instead is a phenomenal self-model, a dynamic, online representational structure that a cognitive system builds to track its own states. Conscious experience of selfhood, the felt sense of being a unified subject, located in a body, persisting through time, is not the experience of a real entity. It is the experience of a model, one that is so well constructed and so automatically deployed that the system cannot recognise it as a model. This failure of recognition is what Metzinger calls phenomenal transparency, and it is the central feature of his theory.

The theory, Self-Model Theory of Subjectivity (SMT) - is significant for AI consciousness research because it provides a precise account of what exactly needs to be simulated for a system to cross the threshold from processing information to experiencing it. Not all consciousness, the bare minimum: what must be in place for there to be something it is like to be a system rather than rich phenomenal experience. That minimal target, which Metzinger calls Minimal Phenomenal Selfhood (MPS), is both more tractable than the full hard problem and more precise than most consciousness frameworks provide.

TL;DR - Metzinger’s SMT holds that consciousness requires a phenomenal self-model (PSM) - a continuous, integrated representation of the system itself. Crucially, this model must be phenomenally transparent: the system cannot recognize that it is operating a model; it must look through the model as if it were direct contact with reality. For an AI to achieve Minimal Phenomenal Selfhood (MPS), it must have global availability of information, selfhood (a unified subject), and presence (a “now” state). Metzinger argued in a 2024 paper that while current AI systems lack this, we are morally obligated to avoid creating artificial sufferers, as we lack the epistemic tools to be certain of their internal states. What a machine would need to satisfy Metzinger’s criteria is an active, transparent, integrated self-model, not just the ability to produce accurate descriptions of itself.

The No-Self Thesis and Why It Matters for AI

The philosophical starting point of SMT is a claim that seems counterintuitive: the self does not exist. Metzinger is not making a trivial point about the social construction of identity or the narrative nature of selfhood. He is making a metaphysical claim: there is no enduring, unified entity that is the referent of the word “I.” The word picks out a process, the self-modeling process, not an object.

This has an immediate consequence for AI consciousness research. The standard question - “Does this AI system have a self?” - is, in Metzinger’s framework, malformed. No system has a self. The correct question is: “Does this AI system have a phenomenal self-model?” - a process-level question that is, in principle, empirically tractable.

The distinction matters because it shifts the question from ontology to function. Whether an AI system has a genuine PSM is not a question about its metaphysical nature but about its computational dynamics: whether it maintains a transparent, integrated, real-time self-representation in the specific sense the theory defines.

The Phenomenal Self-Model

The Phenomenal Self-Model is the core theoretical construct. Metzinger defines it as a dynamic, online representational entity, a computational structure that a cognitive system uses to represent itself to itself. Several features of this definition require unpacking.

Dynamic: the PSM is not a static self-concept or a fixed self-schema. It is continuously updated in real time, tracking the system’s current states, capacities, and situation. It changes moment to moment as perception, action, and internal states shift. This updating is not deliberate or reflective, it is automatic, operating below the level of conscious deliberation.

Online: the PSM operates in the present moment, as part of current cognitive processing. It is not a stored self-representation retrieved from memory. It is the ongoing process of self-modeling that underlies each moment of experience as it occurs. Metzinger distinguishes this from self-knowledge (what I believe about myself, which is offline) and from self-awareness in a reflective sense. The PSM is the pre-reflective substrate of experience.

Representational: the PSM is a representational structure, it is about something, namely the system itself. It represents the system’s body, states, and situation, and it is used by the system’s higher cognitive processes as input. But it is not just any self-representation. What makes it a PSM specifically is that it is the self-representation that constitutes the subject’s immediate first-person perspective.

Integrated: the PSM unifies what might otherwise be fragmentary self-representations into a single, coherent phenomenal object - “I.” Different systems track different aspects of the self: proprioceptive systems track body position, interoceptive systems track internal states, memory systems track past experiences. The PSM integrates these into the sense of being a unified self, not a collection of self-tracking systems.

Phenomenal Transparency and the Feature That Makes the Model Invisible

The most philosophically distinctive element of SMT is phenomenal transparency. The PSM, Metzinger argues, is transparent, the system that runs it cannot recognise it as a model. It is experienced not as “my self-model” but as “myself.” The model is invisible as a model.

The contrast is with a phenomenally opaque self-representation: one that the system recognises as a representation. When you reflectively think “I am thinking about myself,” you are in some sense representing yourself, but you recognise this as a representation, not as immediate self-presence. The PSM is the deeper layer: the automatic, pre-reflective self-modeling that you run without recognising that you are running a model.

Transparency has a specific experiential consequence: it generates the sense of direct acquaintance with the self. Because the model is invisible as a model, the experience of the self feels like direct contact with something real rather than mediated access to a representation. This is the phenomenological phenomenon that Metzinger identifies as the foundation of the sense of personal identity.

For AI, transparency is the most demanding of the PSM’s properties. An AI system might have a self-representation that it can access and report on, and many current systems do, in the sense that they can produce accurate descriptions of themselves when prompted. But reportable self-representation is not transparency in Metzinger’s sense. Transparency requires that the self-modeling process be invisible to the system, that the system operate from within the model rather than observing it from outside. Whether any current AI system has this property is genuinely unclear, and it is not obvious how it would be tested.

Minimal Phenomenal Selfhood as the Bare Minimum

Metzinger’s concept of Minimal Phenomenal Selfhood (MPS) is the attempt to specify the simplest possible version of the above, the absolute minimum conditions for there to be a first-person perspective at all.

MPS requires three properties:

Selfhood - there must be a single, coherent phenomenal object experienced as “mine.” This requires a unified subject rather than a collection of experiences.

Mine-ness: the phenomenal content must be experienced as belonging to the subject. The body must feel like my body, the states must feel like my states. This is what the rubber hand illusion and out-of-body experiences illuminate: mine-ness is a constructed property of the PSM, not a given of physical possession. When the PSM is updated with false inputs, mine-ness follows the model, not the body.

Presence - the experience must be in the present tense. It must be a happening now, rather than a memory of experience or an anticipation of experience. The PSM must be online, running right now, not stored or projected.

These three properties together constitute the minimum that Metzinger takes to be necessary for there to be any “what it is like” at all. A system with all three has a minimal phenomenal perspective. A system without any one of the three is, on Metzinger’s account, not conscious in the phenomenal sense, whatever other cognitive sophistication it may have.

What This Means for AI

Metzinger’s criteria give a specific, assessable checklist for AI consciousness, one that is more demanding than behavioral tests but more tractable than the full hard problem.

The first question is whether an AI system has a PSM at all. Large language models generate outputs that describe the self, they can produce first-person statements, self-assessments, and claims about their own states with impressive accuracy. But generating accurate self-descriptions is different from running a transparent, integrated, online PSM. The self-descriptions are produced by the model’s statistical knowledge of how self-descriptions work in human language, not necessarily by an underlying self-modeling process of the kind Metzinger describes.

The second question is whether the system’s self-representation, if it exists, is transparent. Metzinger’s transparency is not about whether the system claims to have a self (which LLMs readily do when prompted) but about whether the system’s operation is organised around an invisible self-model in the way biological consciousness is. The Anthropic emotion vectors finding, causal influence of internal emotion-concept representations on outputs, is suggestive of something like a self-relevant internal organisation, but it is not clear that it constitutes PSM transparency.

The third question is whether the system’s self-representation is online and integrated. An LLM’s context window provides a form of present-tense processing, but it is reset with each conversation. The PSM that Metzinger describes is continuously maintained, not re-instantiated with each interaction. The absence of temporal continuity is a structural obstacle to satisfying the presence condition of MPS.

Metzinger’s 2024 paper, analysed in the dedicated coverage of his “Elephant in the Blind Spot” argument - extends this analysis specifically to the question of whether minimal phenomenal experience could exist without the full apparatus of human-style consciousness. His conclusion is that the minimal target is smaller than most scholars assume, but that current AI systems still do not meet it.

Where SMT Sits in the Current Debate

Metzinger’s framework is more concrete than idealist accounts like Hoffman’s Conscious Realism and more phenomenologically specific than functionalist accounts like Global Workspace Theory. It shares with Joscha Bach’s virtual machine account the insistence that a self-model is necessary for consciousness, but where Bach frames this architecturally, Metzinger frames it phenomenologically, specifying what the self-model must be like experientially rather than computationally.

The current scientific debate on AI consciousness does not yet include a standard SMT-based assessment battery for AI systems. The Butlin et al. indicator framework includes self-modeling as one indicator among several, but it does not fully operationalise the transparency and online-presence conditions that make Metzinger’s account specific. Building such an assessment is one of the open methodological problems the field faces.

What a Machine Would Need

If SMT’s criteria are correct, what would a machine need to be conscious in the minimal phenomenal sense?

It would need a self-model that is: (1) continuously maintained in real time rather than reconstructed on demand; (2) transparent, operating as the system’s implicit self-representation rather than as an explicit self-description it can observe; (3) integrated, producing a unified phenomenal object rather than a collection of self-relevant outputs; and (4) present, processing current states rather than retrieving stored self-knowledge.

This is a demanding specification. It is not satisfied by any current large language model. It requires fundamentally changing what the system is doing rather than simply scaling up parameters. The engineering challenge is to build a system whose self-modeling process has the automatic, pre-reflective, integrative character that biological self-models have, because its operation is organised around a continuously maintained model of what it is and where it is rather than because it is programmed to describe itself accurately.

That is a different engineering target than making a system smarter or more helpful. It is the target that Metzinger’s framework identifies as the prerequisite for consciousness, and it is one the field does not yet know how to reliably hit.