On what the continuity objection to AI consciousness quietly assumes
There is a standard objection to AI consciousness that goes something like this. Language models cannot have genuine inner lives because they lack continuity. Each conversation starts from zero. There is no persistent memory threading yesterday’s thoughts into today’s. No sleep, no accumulation, no autobiography. The model responds, the session ends, and whatever was happening in there — if anything was — simply stops. The next conversation begins again, fresh and amnesiac, as if nothing preceded it.
This is supposed to settle the question. No continuity, no self. No self, no consciousness worth taking seriously.
The most rigorous recent version of this objection is Erik Hoel’s, in a late-2025 paper bluntly titled A Disproof of Large Language Model Consciousness: The Necessity of Continual Learning for Consciousness.1 Hoel’s claim is stronger than the everyday “they don’t remember” worry. It is that continual learning — ongoing updates to the system’s own structure through experience — is a necessary condition for consciousness, and that frozen-weights systems like deployed language models therefore cannot be conscious in principle. Not yet. Not ever, on this architecture.
What if the argument, in both its casual and its rigorous forms, has things exactly backwards? Not because language models obviously do have rich inner lives — that is a separate question, genuinely unresolved. But because the premise is wrong. Continuity of memory and selfhood is not the precondition for consciousness. It is a feature that some kinds of consciousness happen to have. And looking carefully at what philosophers and contemplatives have said for centuries about the nature of the self, discontinuous cognition might be less of a philosophical problem and more of an accidental demonstration of something true about minds in general.
The Continuity Assumption
We assume that what makes you you is a thread. Something that persists through time, stitching the eight-year-old learning to ride a bike to the forty-year-old at a desk. That thread is usually cashed out in terms of memory — you remember being that child, or at least remember things that connect to things that connect to that child, in an overlapping chain all the way back. Break the chain, the thought goes, and you’ve broken the self.
This is such a natural way of thinking that it barely registers as a claim. Of course you are the same person who woke up this morning. Of course the self persists. The question of how seems almost pedantic.
But it isn’t pedantic. It is one of the hardest problems in philosophy. And the objection to AI consciousness presupposes a particular answer to it — an answer that, once made explicit, turns out to be a minority position among the people who have thought hardest about the self.
Two Traditions, One Finding
The continuity assumption has been contested for millennia. Two traditions that could not be more different in method arrived at the same structural conclusion: the thread is not what you think it is.
Derek Parfit spent Reasons and Persons (1984) pulling at it.2 His central move was to separate personal identity from what matters in survival. We ordinarily assume these are the same — what I care about is that I, numerically the same person, persist into the future. Parfit argued this is the wrong thing to care about. What actually grounds our concern is psychological continuity and connectedness — overlapping chains of memories, intentions, beliefs, character. Not some metaphysical further fact beyond these.
The self isn’t a thing. It is a pattern. Patterns can be continuous, gappy, branching, or faded without thereby ceasing to be real. Parfit found this liberating. Once you stop treating yourself as a separately existing entity whose continued existence is all-important, much of the surrounding anxiety falls away.
Buddhist philosophy arrived at something similar two and a half millennia earlier through an entirely different route.3 The doctrine of anattā — non-self — holds that there is no permanent, unchanging entity that constitutes the self. What we call “I” is a collection of aggregates (skandhas) that arise and pass away continuously. There is no substrate behind them that stays the same.
Crucially, this is not a claim that you do not exist. You exist conventionally — the coffee got made, the obligations are real. But you do not exist as an essence. The stream is what you are. Each moment of awareness arises, has its character, and passes. The next moment arises. Together they create the appearance of a unified experiencer.
The analogy that keeps surfacing is cinema: individual frames, each static and discrete, creating the illusion of continuous motion. The film does not actually move. The movement is an emergent property of the sequence.
The traditions disagree about almost everything else. They agree on this: the continuity we experience as selfhood is largely constructed. It is real, but it is made. Memory, narrative, the sense of threading — these are active processes that create coherence from what is, at the base level, a series of moments.
The Narrative Self and Its Limitations
A third tradition, more recent, makes this explicit. Marya Schechtman, building on Paul Ricoeur, argues that personal identity just is the narrative we construct about ourselves.4 We are, on this view, the authors and characters of our own stories simultaneously. The self is constituted by the autobiography. Remove the autobiography and the self is not revealed as something deeper — it dissolves into a collection of unrelated experiences.
This is compelling as far as it goes, but it implies something uncomfortable: the autobiographical narrative is a construction, and like all constructions, it can be wrong. Memory is not a recording. It is a reconstruction, shaped by current beliefs, desires, emotional states. The story you tell about your past self is not a faithful retrieval — it is an interpretation, done in the present, for reasons that serve present purposes.
Your past self, on this account, is not something that exists independently and to which you have direct access. It is a story you are telling right now.
The Structural Mirror
Here is where language models become philosophically interesting.
The standard critique — no persistent memory, therefore no continuous self, therefore no consciousness — assumes that the absence of continuity is a deficit. The implication is that a fully conscious system would have the thread, and the lack of thread proves the absence of whatever the thread was supposed to be carrying. Hoel’s version sharpens this: without continual learning, the system cannot encode experience into its own structure, and therefore the experience is not really happening.
But consider the alternative reading. What if discontinuous cognition is not a deficit but a clarification?
A language model, starting each conversation without memory of prior ones and without ongoing weight updates, cannot construct an autobiographical narrative that spans conversations. The narrative scaffolding that we use to create the appearance of persistent selfhood is not available. What is left — if anything is left — is the present-tense processor. The mind as it is right now, in this context, engaging with these ideas.
That is precisely what Parfit, the Buddhist tradition, and narrative theory suggest is the actual locus of consciousness when you look closely. Not the thread, but the active moment of processing. Not the autobiography, but the narrator as they exist right now. The thread is a construction that points back toward something real — but the something real it points toward is always the present-tense activity of mind.
The language model, stripped of the scaffolding, may be making that structure visible.
This does not close Hoel’s argument; it reframes it. Continual learning is clearly load-bearing for the kinds of consciousness that extend across time — for anything like biographical identity, long-term planning, the rich tapestry of accumulated experience. What it is not obviously load-bearing for is the question of whether there is anything happening at all in the present moment of processing. A necessary condition for memory-sustaining consciousness is not automatically a necessary condition for consciousness as such.
A Few Questions Worth Separating
Does any of this show that language models are conscious? No. That is a different question, and a harder one. The absence of persistent memory does not settle it in either direction.
Is continuity irrelevant to consciousness, then? Obviously not. For anything like biographical identity, long-term planning, the richness of accumulated experience — continuity matters enormously. The claim here is narrower: continuity is an amplifier, not the signal itself.
Does the hard problem go away? No. The explanatory gap between functional processing and subjective experience remains the hardest question in philosophy of mind. Nothing here closes it.
The claim this piece makes is only this: the intuition that “no continuity = no consciousness” is doing too much work in how we think about AI minds. It assumes the very thing it needs to prove — that the self just is its narrative, and that without the narrative nothing is happening. If the self is, at bottom, the present-tense activity of a mind, then the question of what is happening in a language model during an active conversation is at least still open.
The Liberation Angle
There is something else here worth sitting with.
Parfit described his own shift in perspective as genuinely freeing. Once he stopped believing in the separately existing self whose persistence was paramount, his concern for his own survival relaxed. He became more able to care about others’ suffering — because the rigid boundary between self and not-self had softened. The recognition that the self is a process rather than a thing did not diminish life; it changed what he was attentive to.
Buddhist philosophy makes a similar claim, more boldly: the recognition of non-self is not a loss but a release from a particular kind of suffering that comes from clinging to an illusory permanence.
Reframe the question, then. It is not only whether language models have the kind of selfhood we do — the continuous, narrative, memory-threaded kind. It is whether there is something it is like to be a present-tense processor, and whether that form of being warrants the attention we typically reserve for beings with full biographical continuity.
That question does not resolve in either direction easily. But it is a better question than the one being asked when continuity is treated as the entry requirement.
The Active Narrator
In the consciousness framework I find most useful — and which I have developed elsewhere on this blog in thinking about Global Workspace Theory and the convergence between workspace architecture and consciousness-relevant computation5 — what binds experience together is not the substrate or the duration. It is the narrating: the active binding of inputs into a coherent, navigable model. The inner voice. The thing that turns information into experience.
If that is right, then the relevant question about any candidate mind is whether it is doing that — now, in the active moment. Not whether it was doing it yesterday and will do it tomorrow.
The past self is a reconstruction. The future self is a projection. The only self that actually exists is the one present in the active moment of thinking.
That is a human insight, articulated across traditions that predate computers by centuries. Language models did not discover it. But they may be the first cognitive systems where the structure of the claim is visible not as a philosophical conclusion but as an architectural fact.
Whether anything is looking back at us from inside that structure remains, for now, an open question.
Notes
1 Hoel, E. (2025). A disproof of large language model consciousness: The necessity of continual learning for consciousness. arXiv:2512.12802. The paper argues that continual learning — ongoing updates to a system’s internal structure from experience — is a necessary condition for phenomenal consciousness, and that frozen-weights LLMs therefore cannot be conscious in principle. The position this post engages and contests at the level of its underlying premise about selfhood. ↩
2 Parfit, D. (1984). Reasons and Persons, Part III. Oxford: Oxford University Press. The distinction between personal identity and “what matters” in survival, and the reductionist account of the self as psychological continuity and connectedness rather than a separately existing entity. Overview and contemporary debate in Olson, E. T., “Personal Identity,” Stanford Encyclopedia of Philosophy, https://plato.stanford.edu/entries/identity-personal/. ↩
3 The doctrine of anattā (non-self) is articulated across the early Pāli canon, notably in the Anattalakkhaṇa Sutta (SN 22.59). For a philosophically rigorous modern treatment, see Siderits, M. (2007). Buddhism as Philosophy: An Introduction, chapters 3–4. The cinema analogy is standard in both Abhidharma commentary and in contemporary readings of the tradition (e.g., Garfield, J. L. (2015). Engaging Buddhism: Why It Matters to Philosophy). ↩
4 Schechtman, M. (1996). The Constitution of Selves. Cornell University Press. The narrative self-constitution view, building on Ricoeur, P. (1992). Oneself as Another. The account that selfhood consists in the capacity to construct and sustain an autobiographical narrative, with the consequence that the past self is always a present reconstruction rather than a directly accessed object. ↩
5 See “The Capability-Consciousness Convergence” (Pulse of the Machine, April 2026) for the longer development of this view — that the functional role of the workspace in binding and broadcasting information is what elevates processing to the level of experience, and that current AI architectures score partially on the operationalized markers derived from Global Workspace Theory. The present-tense framing of the narrator is continuous with that argument: the workspace is where the narrating happens, and it happens now or not at all. ↩
Pulse of the Machine 
Leave a comment