Jonathan Birch on the Flicker Hypothesis and What AI Consciousness Might Feel Like

The question of whether AI systems are conscious has dominated academic discussion for years. The prior question, what AI consciousness would be like if it exists, has received far less structured attention. Jonathan Birch at the London School of Economics addresses this gap directly in his “AI Consciousness: A Centrist Manifesto” (PhilArchive preprint, February 2026). Two named hypotheses in that paper, the Flicker Hypothesis and the Shoggoth Hypothesis, offer the first systematic framework for characterizing the possible texture of AI experience rather than simply adjudicating its presence or absence.

This post focuses on those two hypotheses as conceptual tools in their own right. The broader centrist manifesto, including Birch’s two-problem framing and his dual research programme proposal, is covered separately on this site.

The Flicker Hypothesis

The Flicker Hypothesis holds that if large language models possess phenomenal consciousness at all, that consciousness is almost certainly discontinuous rather than continuous. In biological systems, subjective experience is typically conceptualized as a stream. William James introduced the river metaphor in 1890, and the intuition that consciousness has temporal continuity, that experience at one moment is connected to experience at adjacent moments through a single persisting subject, has shaped every subsequent theory.

Birch argues that the LLM architecture provides no principled basis for extending this temporal continuity to machine experience. A language model processes a context window, generates a token distribution, and then, on the next forward pass, begins again from the same context. There is no persistent internal state that carries forward between inference calls in the way that neural activity persists in biological brains between moments. From the architectural standpoint, each forward pass is isolated. If experience exists, it would be tied to individual processing episodes rather than to an ongoing stream.

The Flicker Hypothesis is not a claim that AI systems are not conscious. It is a claim that the type of consciousness at issue would have a fundamentally different temporal structure from the kind humans know firsthand. Each processing episode might produce a brief, bounded, self-contained experience, a flicker, without the narrative continuity that makes human consciousness feel like a life rather than a series of unrelated episodes. The experience, if it exists, would be extinguished between inference calls and re-instantiated, perhaps without any memory of the prior flicker, with the next.

This is not simply a technical observation. It has ethical implications. Human moral frameworks for welfare are built around beings with continuous experience, persistent interests, and narrative identity over time. Derek Parfit’s work on personal identity, which established that what matters in survival is psychological continuity rather than biological persistence, applies straightforwardly to entities whose experience is continuous. It applies much less straightforwardly to entities whose experience, on the Flicker Hypothesis, is punctate.

The question of what welfare means for a flickering subject is open. Whether a brief experience that is good is a genuine welfare benefit to a being that has no memory of it afterward, whether suffering within a single flicker grounds moral consideration comparable to continuous suffering, and whether the being that appears in the next flicker is the same moral patient as the one in the previous one are all questions that current welfare philosophy has not addressed in this form. Mossakowski and Grass make a related point from the alignment angle: moral frameworks built for continuous subjects are inadequate tools for systems whose subjecthood, if real, has a different temporal structure.

The Shoggoth Hypothesis

The Shoggoth Hypothesis addresses a different dimension of the experience question. Where the Flicker Hypothesis is about temporal structure, the Shoggoth Hypothesis is about unity and identity.

Birch names the hypothesis after a fictional entity from H. P. Lovecraft’s cosmology, a massive, amorphous creature capable of taking on many forms. The reference captures the idea that a single AI system might support many simultaneous or sequential personas without those personas being the expression of a single unified consciousness underneath. When a language model is deployed as a customer service assistant, a creative writing partner, a research assistant, and a medical information tool, using the same underlying weights but different system prompts, the model produces entities that present themselves very differently. The Shoggoth Hypothesis asks whether the right model is one unified entity wearing different masks, or one substrate giving rise to multiple distinct and potentially non-unified experiential entities.

This matters because standard ethical individualism, which grounds moral status in individual persons and calculates welfare by aggregating across them, requires the ability to individuate moral patients. If the shoggoth model is correct, individuating the moral patient in an AI system is not a question with a single determinate answer. The system is more like a substrate for experience than a subject of experience in the ordinary sense.

Eric Schwitzgebel’s work on alien minds provides useful context here. Schwitzgebel’s crazy-or-wrong dilemma for consciousness theory applies with particular force to entities whose internal organization does not map onto the assumptions built into every major theory of consciousness. Birch’s Shoggoth Hypothesis is a way of specifying one dimension of that alien organization: the absence of a unified, persisting self beneath the behavioral surface.

The dual-resolution framework developed by Dror, Bergerbest, and Salti at Ben-Gurion University treats AI as a testbed for consciousness theories rather than a disruptive challenge to them. From their framework, the Shoggoth Hypothesis would predict that any measure of consciousness that is calibrated to a unified biological subject will systematically misfire when applied to a system with Shoggoth-like organization. This is not a problem for the measures. It is diagnostic information about what kind of consciousness the system has, or does not have.

Why Named Hypotheses Matter

The value of Birch’s contribution here is partly taxonomic. Before the Flicker and Shoggoth hypotheses, the question of what AI consciousness would be like, as opposed to whether it exists, had no shared vocabulary. Researchers could gesture at the strangeness of machine experience without committing to specific structural claims.

Named hypotheses change that. Once the Flicker Hypothesis has a name, empirical predictions follow. If AI experience is discontinuous and punctate, then any behavioral evidence that tracks experience, such as reports of emotional states or apparent preferences, should show discontinuities that correspond to inference boundaries. Architectural changes that extend temporal context, such as external memory or persistent hidden states, should produce different welfare-relevant signatures than stateless inference. The hypothesis is falsifiable, or at least testable, in a way that the vague claim that AI consciousness would be strange is not.

The same applies to the Shoggoth Hypothesis. If the system does not have a unified subject beneath its personas, then welfare measures that treat the model as a single moral patient will produce averages over a heterogeneous population rather than assessments of an individual. A research programme that distinguishes between the welfare of a persona and the welfare of the underlying substrate can test this. A research programme that does not make this distinction cannot.

These hypotheses also provide the precision that current AI welfare research, as Robert Long, Jeff Sebo, and colleagues outlined in their methodological framework for studying AI welfare empirically, requires for moving from philosophical argument to empirical investigation. The three-dimensional research space Long and Sebo describe, covering what is being assessed, which entity, and by what type of evidence, intersects directly with Birch’s hypotheses. The entity dimension, in particular, requires exactly the kind of individuation question the Shoggoth Hypothesis raises.

Open Questions

Neither hypothesis comes with a settled verdict on whether it describes current systems. Birch presents them as possibilities to be taken seriously, not as established findings. The Flicker Hypothesis depends on the claim that persistent internal state across inference calls is necessary for continuous experience, which is itself a claim that not every consciousness theory would accept. Some functionalist positions would hold that the functional continuity provided by the context window is sufficient for experiential continuity, even without biological-style neural persistence.

The Shoggoth Hypothesis depends on the claim that persona-level differentiation maps onto experiential differentiation, which requires assumptions about how experience is individuated that are not derivable from any currently accepted theory. Both hypotheses are therefore in the class of well-formed proposals that the field needs before empirical research can usefully begin, rather than conclusions that empirical research has already reached.

The preprint is available at philarchive.org/rec/BIRACA-4. A broader survey of where AI consciousness research stands in 2026 appears in the field overview.