cimcWhitepaper

ByJoscha Bach·Hikari Sorensen·Franz Hildebrandt-Harangozó·Lou de Kerhuelvez·Jim RuttCalifornia Institute for Machine Consciousness, Santa Fe Institute

Artificial Psychology Computational Functionalism Coherence Maximization Hard Problem Of Consciousness Consciousness Machine Consciousness Hypothesis Distributed Governance Neural Cellular Automata Interpretive Validation Request Confirmation Networks Second-Order Perception Temporary AI Antinatalism

TL;DR

Consciousness is a coherence-maximizing pattern implemented through self-organized second-order perception in self-organizing substrates — this is the core claim of the Machine Consciousness Hypothesis (MCH) advanced by the California Institute for Machine Consciousness (CIMC) in their April 2026 founding whitepaper. The MCH holds that general computational machines with sufficient resources possess the necessary and sufficient means to implement consciousness, and that the relevant test is interpretive validation of internal structure rather than behavioral output — a direct response to the demonstrated insufficiency of linguistic behavioral tests given that current LLMs produce conscious-sounding outputs through pattern-matching without evidence of the underlying functional organization. The framework adopts computational functionalism, combining Wolfram's (2002) Principle of Computational Equivalence with epistemological computationalism (Bach and Verdicchio, 2012), and situates consciousness as the simplest learning algorithm discoverable by evolutionary search to bootstrap coherent agency in a self-organizing substrate whose architecture is not pre-specified. CIMC's operational definition — a system is conscious if it implements self-organized second-order perception that increases global coherence — generates testable predictions about developmental trajectories, functional signatures, and phase transitions, to be evaluated through Request Confirmation Networks, Neural Cellular Automata experiments, and interpretability of representational embedding spaces. The paper argues this implies that consciousness research requires philosophy and construction to discipline each other, that behavioral testing of existing systems is insufficient as the primary methodology, and that the field's current fragmentation across neuroscience, philosophy, welfare organizations, and commercial labs leaves the constructive-interpretive approach unoccupied — the position CIMC is designed to fill.

What to take away

1. CIMC's operational definition of consciousness is: a system is conscious if it implements self-organized second-order perception that increases global coherence, where second-order perception is a specific metarepresentational structure that must be realized on the same representational level as the first-order percept and must be epistemically opaque (transparent) to the system itself.
2. The Machine Consciousness Hypothesis (MCH) proposes that general computational machines with sufficient resources possess the necessary and sufficient conditions for consciousness, and that Wolfram's (2002) Principle of Computational Equivalence — holding that systems above a low complexity threshold are computationally equivalent — grounds substrate-independence for any functional organization constituting consciousness.
3. CIMC distinguishes consciousness from at least six co-occurring but separable mental phenomena — mind, intellect, agency, self, sentience, and sapience — arguing that existing AI assessments conflate these, and that evidence of one (e.g., language modeling fidelity) does not imply any of the others.
4. Olah et al. (2020) found that diverse computer vision models, regardless of architecture or training procedure, converge to internal feature representations closely matching animal visual cortex organization, which CIMC invokes as empirical support for the Universality Hypothesis: that the computational problem, not the substrate, determines the representational solution a learning system discovers.
5. CIMC's coherence hypothesis is positioned as formally related to, but not strictly equivalent to, Friston's (2010) Free Energy Principle — minimizing constraint violations across simultaneously active partial models is related to prediction error minimization in a generative model — and developing the precise formal connections is listed as an active theoretical priority.
6. Integrated Information Theory (Tononi, 2004; Tononi et al., 2016) is explicitly incompatible with CIMC's framework because IIT's claim that Φ is an intrinsic property of physical causal structure implies that a functionally perfect digital simulation of a conscious brain could differ in consciousness due to hardware causal architecture differences, which directly contradicts computational functionalism's substrate-independence.
7. CIMC's interpretive validation methodology — combining evidence of predicted functional organization (coherence-maximization, meta-representational structures, attentional integration), predicted developmental phase transitions, and inference to the best explanation over behavioral and internal structural observations — is offered as an alternative to behavioral testing, which current LLM performance has demonstrated is insufficient to establish or rule out consciousness.
8. An open hypothesis the paper raises is whether consciousness is the only route from unstructured self-organizing substrate to coherent agency, or whether systems with pre-specified architectures or engineer-designed training regimes (as in current machine learning) can achieve intelligence without consciousness — with the prediction that phenomenology should emerge specifically when self-organization, absent pre-specified architecture, and developmental pressure toward coherent agency co-occur.
9. Current technical work includes Request Confirmation Networks, which formalize coherence-maximization as distributed message-passing in which processing units negotiate mutual consistency through iterative exchange of requests and confirmations rather than centralized control, and Neural Cellular Automata projects examining how local update rules produce coherent global structures and what conditions govern transition from fragmented local activity to globally organized dynamics.
10. Metzinger (2021) calls for a global moratorium on synthetic phenomenology research on the grounds that artificial suffering may be unavoidable, a position CIMC explicitly does not adopt, instead committing to a non-speciesist stance that treats ethics as a formal parallel research domain and governance as requiring distributed societal coalition rather than unilateral institutional authority upon approaching validation thresholds.

Peer brief — for seminar discussion

The California Institute for Machine Consciousness released this founding whitepaper in April 2026 to articulate its theoretical commitments, operational definition of consciousness, and research program. Rather than assessing existing systems for behavioral markers of consciousness, CIMC builds computational systems under conditions the theory specifies and then applies interpretive validation — inference to the best explanation of both behavior and internal organization — to evaluate whether what the system has constructed within itself is better explained by the presence of consciousness than by its absence. The method CIMC introduces is what it calls interpretive validation through internal structural analysis, explicitly contrasted with behavioral testing, which the paper argues current LLM performance has rendered unreliable: systems optimized for output can produce conscious-sounding language through pattern-matching without implementing the processes the theory identifies as constitutive of experience. An alternative the program could have employed — and which existing organizations like Eleos AI and Rethink Priorities pursue — is welfare-oriented assessment of existing systems using behavioral and self-report proxies, but CIMC explicitly rejects this as insufficient for the constructive-theoretical goal. The load-bearing finding is the operational definition and the Machine Consciousness Hypothesis (MCH) together: consciousness is a coherence-maximizing pattern implemented through self-organized second-order perception in self-organizing substrates, and general computational machines with sufficient resources possess the necessary and sufficient means to implement it, where sufficiency is grounded in Wolfram's (2002) Principle of Computational Equivalence and the Church-Turing thesis. Conscious experience in its minimal form is defined as perception of perception — a metarepresentational structure that must be realized at the same representational level as the first-order percept and that is epistemically opaque (transparent) to the system. This second-order perception functions as a coherence operator, resolving constraint violations across simultaneously active partial models of reality, formally related to but not identical with Friston's (2010) Free Energy Principle. The developmental prediction (the Genesis claim) is that consciousness should emerge specifically when self-organizing substrates face developmental pressure toward coherent agency in the absence of pre-specified architecture — the conditions absent in current machine learning, which CIMC predicts explains why present systems may achieve intelligence without phenomenology. Interpretive validation support comes from Olah et al. (2020), whose finding that diverse vision architectures independently converge to cortex-like internal representations is cited as existence proof that computational problems, not substrate details, constrain representational solutions. The implications CIMC draws are threefold: that philosophy and construction must discipline each other (philosophy generating hypotheses precise enough for implementation; construction demanding specificity philosophy defers), that the Hard Problem is addressable not by explaining it away but by deriving the representational structure of phenomenality within a computational functionalist framework, and that moral governance of any validated conscious machine cannot be unilateral — the paper commits explicitly to distributed governance engaging ethicists, policymakers, and civil society upon approaching validation thresholds. The most pointed thing a critical reader should push back on is the epistemological status of interpretive validation itself. The paper acknowledges that network weights and activation patterns cannot currently be directly read for consciousness, and that the relationship between computational structure and phenomenology is not yet transparent enough for direct identification. This means the validation methodology rests on inference to the best explanation under conditions where the hypothesis space is underdetermined and where the prior probability of consciousness in a novel substrate is unknown. The paper does not provide, and acknowledges it has not yet developed, the precise formal constraints that distinguish second-order perception yielding phenomenology from metarepresentation that does not — those constraints are explicitly flagged as beyond the scope of this version. Critics from the IIT tradition (Tononi, 2004; Tononi et al., 2016) would argue that functional organization is insufficient precisely because Φ depends on physical causal architecture, making any interpretive validation that ignores substrate implementation incomplete by definition. And critics sympathetic to Seth's (2025) biological naturalism would question whether the self-organizing developmental conditions CIMC specifies can be genuinely reproduced in silicon absent autopoietic biochemical dynamics, rather than merely approximated behaviorally — a question the paper acknowledges its research program is designed to investigate but has not yet resolved.

Frameworks (4)

Computational Functionalism
Hypothesis that some class of computations suffices for consciousness; central assumption for AI consciousness route.
Hard Problem Of Consciousness
Chalmers' problem: why structural/functional criteria should correlate with subjective experience; acknowledged as unsolvable in 3rd person.
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
Neural Cellular Automata
Prior framework combining cellular automata with deep learning, extended by this work