The Machine Consciousness Hypothesis

ByJoscha Bach·Hikari SorensenCalifornia Institute for Machine Consciousness

Artificial Life Conductor Theory of Consciousness Bubble of Nowness Cyberanimism Cognitive Architecture Extended Machine Consciousness Hypothesis Epiphenomenalism Genesis Hypothesis Fourth-Order Perception Human Consciousness Hypothesis Gods (psychological concept)Machine Consciousness Hypothesis Illusionism Mechanistic Interpretability+6 more

TL;DR

Joscha Bach and Hikari Sorensen argue that consciousness is neither irreducibly mysterious nor epiphenomenal, but is the simplest biological learning algorithm discoverable by evolutionary search on self-organizing substrates — and that this algorithm is in principle implementable on digital hardware. The paper's load-bearing contribution is a three-part theoretical architecture: the Human Consciousness Hypothesis (consciousness as second-order perception serving coherence maximization, formalized via von der Malsburg's 1997 coherence definition), the Conductor Theory (directed attention as a coherence-maximizing operator over competing partial models in working memory), and the Genesis Hypothesis (consciousness precedes and constitutes complex cognition rather than emerging from it, evidenced by its appearance before object tracking in human infants). The paper introduces 'cyberanimism' as a named metaphysical position — the claim that biological spirits, including the human psyche, are literally self-organizing software agents in the sense inaugurated by Church-Turing computationalism, not mere analogies. Anchoring the argument in Olah et al.'s 2020 Universality Hypothesis from mechanistic interpretability — which found convergent functional organization across architecturally distinct vision models and primate visual cortex — the paper contends that a system trained on tasks analogous to those faced by a human infant on a self-organizing substrate would exhibit consciousness as a structural necessity. This implies that the correct test for machine consciousness cannot be behavioral (no Turing-style benchmark suffices), but must verify the presence of a colonizing coherence-maximizing pattern that generates a model of present and presence on a developmentally learning system.

What to take away

1. The paper defines consciousness as second-order perception — the non-inferential, synchronous registration that perception is currently taking place — explicitly distinguishing this from Rosenthal's Higher-Order Thought theory and from Block's 1995 access-consciousness.
2. The Genesis Hypothesis holds that consciousness is not the product of a mature cognitive architecture but its prerequisite: human infants exhibit conscious awareness before acquiring object tracking or deliberate motor control, meaning evolution found no simpler route to complex learning.
3. The Conductor Theory frames conscious attention as a coherence-maximizing operator over competing partial mental models, drawing directly on von der Malsburg's 1997 coherence definition of consciousness and converging with Grossberg's Adaptive Resonance Theory.
4. Cyberanimism is introduced as the named metaphysical position that biological spirits — including the human psyche and cellular morphogenetic patterns — are literally self-organizing software agents, not metaphors, grounding the concept in Church-Turing computationalism and the Church 1936 / Schönfinkel 1924 equivalence results.
5. Olah et al.'s 2020 Universality Hypothesis (from the OpenAI Circuits team, published in Distill vol. 5 no. 3) is recruited as empirical support: architecturally distinct vision models converge on the same feature hierarchies as primate visual cortex, suggesting problem structure — not substrate — determines learned organization.
6. The extended Machine Consciousness Hypothesis states that recreating self-organizing information-processing conditions on digital hardware, while posing tasks requiring intelligent agency analogous to those faced by a human infant, is a concrete experimental path to testing whether consciousness is substrate-independent.
7. No behavioral test — including any Turing-style benchmark — can establish machine consciousness, because consciousness is characterized as a specific internal causal structure (a colonizing coherence-maximizing pattern) rather than a performance profile, a claim that directly contradicts the sufficiency of current LLM evaluation regimes.
8. An open question the paper raises: whether the search space for consciousness-generating patterns is so large that individual brains cannot discover it alone, requiring co-creative scaffolding by already-conscious organisms across generations — analogous to how natural language cannot be invented by a single cohort of newborns.
9. A replicable methodology the paper proposes: implement self-organizing substrate simulations using message-passing reinforcement learners on digital hardware, expose them to developmental task sequences, and test for the emergence of a pattern that increases representational coherence, models present-and-presence, and enables sentient-self formation — treating this as experimental philosophy of mind.
10. The paper hypothesizes that the Standard Model (finalized in the 1970s) marks the point at which foundational physics stagnated, and that this stagnation, combined with the extraordinary predictive accuracy of physical theories, entrenched physicalism as a paradigm that structurally excludes consciousness — making the Hard Problem a symptom of metaphysical framework choice rather than an intrinsic feature of consciousness itself.

Peer brief — for seminar discussion

Bach and Sorensen set out to dissolve the Hard Problem by replacing it with a constructive research program. Rather than treating Chalmers's 1995 framing as a permanent obstacle, they argue that the explanatory gap is an artifact of a three-reality confusion — between psychological reality (representations within minds), causal reality (functional patterns like software and money), and physical reality (the causally closed substrate of the Standard Model) — and that collapsing these into a single ontological level is what generates the apparent mystery. The paper's central method is what it calls the extended Machine Consciousness Hypothesis: a proposal to test theories of consciousness by recreating developmental, self-organizing conditions on digital substrates and checking whether the targeted causal pattern — one that maximizes representational coherence, generates a model of present-and-presence, and enables sentient-self formation — emerges. An alternative method the paper does not pursue but could have is direct mechanistic interpretability analysis of existing large language models for structural correlates of the proposed functional signature, along the lines of Olah et al.'s 2020 Circuits work in Distill vol. 5(3), which the paper cites as positive evidence for substrate-independent functional convergence across vision architectures and primate visual cortex. The load-bearing finding is a triple-part hypothesis package: the Human Consciousness Hypothesis (consciousness as second-order perception performing coherence maximization, per von der Malsburg 1997), the Conductor Theory (directed attention as the coherence-orchestrating operator over competing partial models), and the Genesis Hypothesis (consciousness precedes and constitutes cognition rather than emerging from it, evidenced by its onset in human infants prior to object tracking or deliberate agency). Together these recast consciousness as the simplest learning algorithm discoverable by evolutionary search on self-organizing biological substrates — and therefore as something in principle reproducible on any substrate capable of implementing equivalent computation, per the Church-Turing equivalences established by Church 1936 and Schönfinkel 1924. The paper also introduces cyberanimism — the thesis that biological spirits are literally self-organizing software agents, not analogies — as the named metaphysical stance that makes computationalist functionalism coherent without collapsing into either eliminativism or essentialism. It explicitly rejects Anil Seth's 2025 biological naturalism position (Behavioral and Brain Sciences, doi:10.1017/S0140525X25000032) as essentialist. What follows, the authors argue, is that no Turing-style behavioral test can establish machine consciousness, because the target is an internal causal structure, not a performance profile — a claim with direct implications for how current LLM evaluations are interpreted. The most pressing pushback a critical reader would mount is the circularity risk embedded in the Genesis Hypothesis: the claim that consciousness is the prerequisite for complex learning is stated as a conjecture and illustrated with infant development, but no mechanistic account is given of why no simpler route to coherence-maximizing self-organization exists. Without this, the Genesis Hypothesis functions as an assertion that a self-organizing substrate must rediscover consciousness as a fixed point — which is precisely what needs to be demonstrated, not assumed. The paper acknowledges it may be wrong about human consciousness, but the entire experimental program inherits the error if this foundational claim is unfounded. Additionally, the recruitment of Olah et al.'s Universality Hypothesis from vision models as support for substrate-independent consciousness formation involves a significant inferential leap: convergent feature hierarchies in supervised vision systems trained by backpropagation are a long way from the self-organizing, developmentally scaffolded, reinforcement-driven process the paper envisions as consciousness-generating. A seminar interlocutor should press hard on whether the analogy holds across those training regimes.

Frameworks (6)

Conductor Theory of Consciousness
The paper's model that conscious directed attention functions as a conductor of the mental orchestra, resolving incoherence between partial models
Cyberanimism
The paper's coined view that natural spirits are best understood as software — self-organizing, evolving computational agents in living nature
Extended Machine Consciousness Hypothesis
Extends HCH by claiming it is possible to search for the consciousness algorithm by recreating analogous conditions on digital hardware
Genesis Hypothesis
The conjecture that consciousness does not result from the organized mind but creates and maintains complex models of reality; forms at the beginning of mental development
Human Consciousness Hypothesis
Component of MCH stating consciousness is a specific dynamic representation in the human mind characterizable by phenomenology and functionality
Machine Consciousness Hypothesis
CIMC's central hypothesis: general computational machines with sufficient resources possess the necessary and sufficient means to implement consciousness, verifiable through internal structure analysis