Contribution
The article states the self-model theory of subjectivity (SMT) in a few thousand words, which Being No One takes seven hundred pages to build. It turns “there is no self” from a slogan into a definite claim with a mechanism: what exists is a self-model, and the experience of being a self is what it is like to run a self-model that one cannot recognise as a model. It also separates three things often run together: having a self-model (which a robot can), having a conscious one, and experiencing oneself as a self.
Key insight
The self is not what the self-model is a model of; it is what the model looks like from inside when its construction is invisible. A system that could see its body image as an image, the way we can see a thought as a thought, would on this theory stop experiencing itself as a self.
Assumptions
- Representationalism and functionalism about the conscious self. “What we call ‘the conscious self’ … is a specific form of representational content, characterized by specific functional properties”, multiply realisable.
- Local supervenience of phenomenal content. Fix the brain’s contemporaneous internal properties and the PSM’s phenomenology is fixed (a brain in a vat could have it), while self-knowledge depends on history and environment.
- A minimally sufficient neural correlate of self-consciousness (NCSC) exists for every biological PSM; if it is nomologically coextensive with the PSM, the PSM reduces to it.
- Consciousness is availability: “the space of consciousness essentially is the space of availability for selective, high-level attention”. Called “a central working hypothesis … not yet empirically demonstrated”.
- Teleofunctionalism: the self-model has an evolutionary function, “a weapon … in a ‘cognitive arms race’”.
Key results
There are no formal results. The article’s claims, by section:
- SMT. The questions it is built to answer, from “How, in principle, could a consciously experienced self … emerge in a given information-processing system?” to the ethics of machine subjectivity and the minimally sufficient neural correlate. Method: “interdisciplinary constraint satisfaction”.
- PSM. The conscious partition of the self-model, meeting further constraints (ten in Being No One, ch. 3), chiefly a window of presence and integration into one world model. A “model” here is a variable-free representation by partial structural correspondence that can be run offline (after Craik and Johnson-Laird). Functionally, a “transient computational module” switched on at waking, absent in dreamless sleep, present in dreams. Psychiatric disorders as lost, hyperactive or dissociated self-model content. Abstract thought as taking the PSM offline. The mirror system as “mutually coupled self-models”, the basis of perspective-taking and theory of mind.
- Machine self-models. The starfish robot (Bongard, Zykov and Lipson, 2006): random action, 15 candidate self-models fitted by stochastic optimisation, exploratory actions chosen to maximise their disagreement, convergence, locomotion; after damage the cycle restarts and it “learns to limp”. Lessons: a self-model can be unconscious, a process not a thing, plastic, and non-linguistic, rooted in motor control. Owning a body is using a whole-body model to control it, “the decisive first step towards becoming an autonomous agent”.
- 1E/2E/3E. First-order embodiment without explicit computation (the Tarry II walker); second-order with an explicit self-model used for control (the starfish, primitive animals, perhaps sleepwalkers); third-order with that model conscious (waking humans, perhaps orangutans). Humans always have 1E and 2E, and 3E only in waking and dreams. Brugger’s subject born without limbs reports vivid phantoms, with premotor and parietal activation, as possible evidence of an innate body model. Full-body illusions in virtual reality (Lenggenhager et al. 2007) shift felt self-location toward a seen virtual body, and global body ownership is offered as “the most simple form of self-awareness”.
- Transparency. Transparency is phenomenological, comes in degrees, and holds when earlier processing stages are attentionally unavailable; “the phenomenology of transparency is the phenomenology of naïve realism”. The body image is transparent; conscious thought is opaque. Central claim and prediction as in the LIT. Humans are special in having “a very large opaque partition”.
- PMIR. The intentionality relation can itself be represented: “we also co-represent the representational relation itself”. A first-person perspective is a transparent PMIR. It lets a system discover that it represents and attends, and makes subject–object relations globally available, which may be why it was adaptive.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | There is no substantial self; there are self-models and phenomenal self-models | asserted as the framework’s starting point | parsimony; “maximally parsimonious framework” |
| C2 | A phenomenal self exists iff a system operates under a transparent PSM | a definition with a prediction; untested here | stated; the prediction (opacity abolishes selfhood) is not tested in the article |
| C3 | Consciousness is availability for high-level attention, and the PSM is the self-model’s available partition | working hypothesis, labelled unproven | stated |
| C4 | Self-models can be unconscious and artificial, and a robot can build and revise one | strong for the robot result as reported | Bongard et al. 2006, described; not in the reference list |
| C5 | Embodiment comes in three orders, and humans have the third only episodically | moderate as conceptual analysis | definitions with examples |
| C6 | Body ownership can be manipulated by multisensory conflict and is the simplest form of self-awareness | moderate for the manipulation, cited; asserted for “simplest” | Lenggenhager et al. 2007; Blanke and Metzinger 2009 |
| C7 | Phantoms in congenital limb absence suggest an innate component of the PSM | weak; offered as a question | one case study (Brugger) |
| C8 | A first-person perspective is a transparent model of the intentionality relation | asserted as analysis | definitions; phenomenological examples |
Concepts
- self-model — a model, in Craik’s sense, that a system has of itself as a whole; may be unconscious.
- phenomenal self-model (PSM) — the partition of the self-model that is conscious because it meets further functional constraints.
- transparency — the attentional unavailability of a representation’s earlier processing stages, so that it is experienced as the thing itself; a matter of degree; its opposite is opacity.
- PMIR — phenomenal model of the intentionality relation: a conscious representation of oneself as directed at an object or goal.
- first-, second-, third-order embodiment (1E, 2E, 3E) — intelligence from bodily dynamics alone; with an explicit self-model used for control; with that model conscious.
- NCSC — minimally sufficient neural correlate of self-consciousness.
- virtual window of presence — a system’s internal representation of an extended now.
Connections
- Minsky, The Emotion Machine (LIT-432, NOTE-348). The brief asked
for this one. Points of agreement: both deny a central, substantial
self; both make self-models the stuff of selfhood; both allow them in
machines. Points of difference:
- One model or many. Minsky: “Whenever you think about your ‘Self,’ you are switching among a network of models”, and a single unified model would be impractical (NOTE-348, C10). Metzinger: a PSM is “a coherent self-representation, a consistent internal model of itself as a whole”, integrated into one world model. Minsky’s partial models are closer to Metzinger’s unconscious self-model layers; what Metzinger adds is the claim that the conscious partition is one.
- Where the sense of a self comes from. Minsky explains the Single-Self idea by its convenience (§9-3). Metzinger explains the experience of selfhood by transparency. Minsky’s account says why we talk of a self; Metzinger’s says why it seems real.
- Self-models and consciousness. Minsky makes self-models one of four co-occurring processes whose detection might be what “consciousness” names (§4-3). Metzinger makes conscious self-modelling depend on separate constraints (presence, a world model, availability), and a self-model alone is not conscious. Both reject the inference from self-model to experience, by different routes.
- Body. Minsky treats the body as one more resource. Metzinger roots the self-model in motor control and body ownership: “in its origins it is not based on language or conceptual thought, but very likely on an attempt to organize global motor behavior”.
- Dennett (LIT-442, NOTE-364). Both are no-self views that keep the self as something real of a kind. For Dennett it is an abstractum posited by interpretation; for Metzinger it is the content of a transparent model. Dennett’s self is as determinate as its story; Metzinger’s self-model is fixed by local brain states.
- Gallagher (LIT-453, NOTE-377). Metzinger’s body-ownership work is Gallagher’s sense of ownership measured, and his starfish robot repeats Gallagher’s Box 1 point (Tani’s robot) with a self-model that persists and learns.
- Zahavi (LIT-463, Deferred). By its description, Zahavi’s book defends the self against the view that it is “a neurologically induced illusion”, and Being No One is that view’s main statement.
- Damasio (LIT-387, LIT-388). The article cites Damasio’s “self in the act of knowing” for the relational content of higher self-consciousness, and its interoceptive layers (immune system, brainstem, Seth, Suzuki and Critchley 2012) are the body Damasio’s account of feeling starts from.
- Seth (LIT-135, NOTE-169). Seth’s interoceptive predictive coding is cited here as part of the unconscious self-model. Seth’s case that consciousness belongs to living systems sits uneasily with Metzinger’s multiple realisability of the PSM, though the article does not say whether a machine could have a PSM, only a self-model.
- Clark and Chalmers (LIT-097, NOTE-180). The PSM’s phenomenal content supervenes locally on the brain, so on this theory an extended self is at most an extended self-knowledge, never an extended self-experience. The pairing is mine.
Bearing on the record
- THEORY-023 (AI consciousness). The article gives the record a clean three-way split: self-model (the starfish has one), PSM (needs the unknown phenomenality constraints), phenomenal self (needs a transparent PSM). A system’s self-description or self-model is therefore not evidence of experience on this theory, and the article says the conditions for phenomenality are “presently still unknown”. That agrees with THEORY-023’s verdict that the question cannot now be settled. The article raises “the ethical implications of machine models of subjectivity” only as a question; whatever the book argues on it, the book was not read, so the record does not hold it.
- THEORY-043 (the bridge). The PSM is “an instrument for global control”, one per system, and mutually coupled self-models underlie social cognition. That is an architectural criterion of the kind the bridge lists: a group of selves linked by coupled self-models would not thereby have a single PSM. The application is mine; the article does not discuss groups.
- Levin (LIT-439, NOTE-341). Levin’s self is any goal-pursuing collection with a cognitive light cone; Metzinger’s 2E needs an explicit whole-system self-model. So Levin counts selves Metzinger would count as 1E systems. The comparison is mine.
- No ML instruction. The starfish is an illustration from robotics. A learned world model that includes the agent would be a 2E system in Metzinger’s terms; the article says nothing about learned models generally, and nothing here belongs in the anthology.
- ACT’s cognitive fusion is read as a thought functioning transparently, and defusion as making it opaque, in THEORY-050 and THEORY-044. The mapping is the record’s; neither source cites the other.
Limitations
- Revised text. The page was last modified in 2013 and cites 2012 work; the 2007 citation names a text that is no longer what is served.
- Editing slips. The starfish procedure is described twice in nearly the same words (once under “Self-models in machines”, once under 2E). A list of lessons jumps to “Fourth” in a paragraph that has no first to third. “Input from the physical boy” for body.
- Reference gaps. Cited but not in the list: Bongard, Zykov and Lipson 2006, the main machine example; “Metzinger 2005a”, “2006b” and “2007”; Weisberg is “2007” in the text and 2005 in the list; Lenggenhager et al. has no year in the list; “Metzinger 2003a” appears twice.
- The central claim is not argued here. C2 is stated as a definition with a prediction; the case for it is in the book, unread.
- Single cases. The innate-phantom evidence is one subject.
Open questions
- Is transparency testable apart from reports of selfhood? The prediction that an opaque PSM abolishes the experience of selfhood needs a way to make a PSM opaque; the article names none.
- Does a language model’s description of itself count as a self-model in the article’s sense (a variable-free structural model used for global control) or only as self-reference in language? The article’s non-linguistic, motor origin for self-models suggests the latter.