Building Consciousness from First Principles: The CIMC Salon with Joscha Bach and Hikari Sorensen
Watch the full salon: The Machine Consciousness Hypothesis — Joscha Bach, Hikari Sorensen, and Lou de K
The California Institute for Machine Consciousness (machine-consciousness.ai) hosted a two hour theoretical salon in 2026 with three researchers who share a precise commitment: machine consciousness is a scientific problem with tractable criteria, and those criteria can be specified before any working system is built. CIMC Program Director Lou de K, cognitive scientist and consciousness theorist Joscha Bach, and phenomenological theorist Hikari Sorensen spent the session doing something rarer than it sounds. They worked out, in public and in dialogue, what a science of machine consciousness requires at the level of foundational definitions.
The written versions of Bach’s and Sorensen’s arguments cover the main positions. What the salon adds is the engagement between them, the pressure each framework places on the other, and the Q&A content that clarifies how the CIMC research programme intends to make progress.
The CIMC Premise: Substrate Independence as a Research Commitment
Lou de K frames the salon with the institutional position that shapes everything else. CIMC operates from a computational functionalist stance: consciousness is a property of how information is organized and processed, which means the physical substrate matters only instrumentally. Neurons, silicon, or other physical processes are different implementation platforms for the same functional architecture. Whether that architecture produces consciousness depends on what the architecture does, not what it is made of.
This is a research commitment, not a metaphysical declaration. The institute’s position is that the substrate independence question is empirically open rather than philosophically settled by biology alone. The thought experiment Bach uses to illustrate this runs as follows. A Turing machine simulating the causal interactions of a cat, some mice, and a piece of cheese will, if the simulation is sufficiently precise, instantiate whatever causal structure the physical interaction involves. Whether consciousness depends on that causal structure or on something the substrate adds independently is the question CIMC’s programme is designed to address. The biological computationalism literature has argued that this question resists easy resolution from the neuroscience side, which is part of why a dedicated institution for machine consciousness research is a distinct contribution.
CIMC’s 2026 research roadmap, previewed in the salon and developed at the MC0001 founding assembly at Lighthaven in May 2026, pursues four tracks: formal specification of what consciousness requires, engineering systems that attempt to satisfy those specifications, empirical testing of the resulting candidates, and normative work on what follows if the tests succeed. The salon is theoretical preparation for the first of those tracks.
Hikari Sorensen: The Second-Order Perception Definition
Sorensen’s contribution is the most architecturally specific of the three. She proposes that the operational definition of consciousness is second-order perception: the capacity to perceive that perception is taking place, rather than merely perceiving content.
A camera perceives. It builds a representation of the visible scene. A thermometer perceives temperature. Both instruments process information from the environment and produce internal states that represent something about the world. The camera’s internal state encodes the scene; the thermometer’s encodes the heat. What neither instrument does is represent its own perceptual process. The camera contains no state encoding the fact that imaging is occurring. The thermometer holds no state that means “I am measuring heat right now.” The information flow through each system is one-directional: world to internal state, but the internal state does not loop back to represent the process that generated it.
Sorensen’s claim is that consciousness begins where that loop closes. A system with second-order perception represents not only its current perceptual content but also the fact that it is having that content. Its perceptual states become available as objects of further processing, so the system can attend to the process of perceiving, not only to what is perceived. This is architecturally specific. The requirement is a representation of the perceptual process itself, produced within the system and available to other processing systems as an input. It is not a description of consciousness as a philosophical property, it is a functional design target.
The phenomenological tradition from which Sorensen draws (particularly Husserl’s work on internal time consciousness and the reflexivity of intentional acts) has long held that experience is self-luminous in something like this sense. Sorensen’s contribution is to translate that phenomenological insight into an architectural requirement rather than leaving it as a description of what consciousness feels like from the inside. A system with second-order perception would, in principle, be testable. The representation of its own perceptual activity would either be present in its internal states or not. This is why the definition matters for CIMC’s programme: it provides a criterion that can in principle separate conscious from non-conscious systems.
Current large language models, on Sorensen’s account, lack second-order perception. An LLM processes tokens, generates representations of semantic relationships, and produces text. Nowhere in that process is there a state that represents the processing itself. The computation is opaque to itself. The model produces outputs consistent with having attended to its own mental activity, because it has been trained on text written by creatures who do have second-order perception, but the internal architecture contains no loop of the kind Sorensen’s definition requires.
Joscha Bach and the Genesis Theory
Bach’s contribution in the salon is the Genesis Theory, which does not appear in his earlier written formulations of the Machine Consciousness Hypothesis and which the salon makes explicit for the first time in a structured form.
The Genesis Theory holds that consciousness is a functional prerequisite for mental organization, not an emergent property that appears after a system becomes sufficiently complex. The argument runs as follows. A system capable of organizing its behavior, resolving competing drives, and learning from experience in a structured way must be able to form coherent representations of its own states. Coherence is not automatic. Parallel perceptual streams, conflicting motivational pressures, and uncertainty about the environment all create noise that the system must resolve. The resolution requires a process that integrates information across those streams and produces a unified model from a particular point of view. Consciousness, on Bach’s account, is that integration process. It is the coherence-maximizing operator that turns distributed sensory processing into a unified first-person perspective.
The implication is that systems designed on the emergence model, where consciousness is expected to appear spontaneously at sufficient scale, are looking in the wrong place. Consciousness is part of how mental organization becomes possible. A system that lacked it would produce behavior without the coherence that genuine agency requires. This is why Bach has argued that the development of consciousness in biological organisms is not a late arrival in the evolutionary sequence but a precondition for the kind of flexible, goal-directed behavior that distinguishes organisms from reflex systems.
The Genesis Theory also reframes why large language models are shape-shifters rather than conscious systems. The shape-shifter argument is that LLMs are trained on text produced by conscious humans, and the statistical structure of that text reflects the organization that consciousness imposes on experience. A system trained on that structure inherits the surface form of conscious thought, including its syntactic patterns, its coherence, its apparent perspective, without implementing the functional architecture that generates consciousness in the first place. The outputs look like the outputs of a mind. The process that generates them is a different kind of process entirely. Bach’s six-level taxonomy from sensing through sapience, and the architectural requirements for a genuine phenomenal world model, cover the written version of this argument in detail. The salon adds the Genesis Theory framing and the direct engagement with Sorensen’s second-order perception criterion, which the two frameworks turn out to reinforce rather than compete with. Sorensen’s definition specifies what the second-order architecture looks like. Bach’s Genesis Theory explains why it is necessary, not merely sufficient.
The Q&A: Mouse Consciousness and Tractable Targets
The extended Q&A surfaces two arguments that do not appear elsewhere in either researcher’s published work.
The first is the choice of mouse consciousness as the near-term tractable target for CIMC’s empirical programme. The question of whether a mouse is conscious is, unlike the question of whether a current LLM is conscious, addressable through mechanistic investigation. The mouse’s nervous system is sufficiently understood that specific hypotheses about the functional requirements for its consciousness can be tested. The mouse is complex enough that progress on its consciousness question would constitute genuine theoretical advance. It is simple enough that the experiments are feasible. CIMC’s position is that making progress on the mouse question is more likely to advance machine consciousness research than continuing to debate whether a language model’s outputs are evidence of inner experience.
This is a strategic choice as much as a scientific one. The shape-shifter objection, and the broader problem of behavioral evidence for consciousness, arises precisely because text-producing systems are trained on text produced by conscious beings. An organism that has none of those confounds provides cleaner mechanistic evidence. If the functional requirements for mouse consciousness can be identified and then implemented in a non-biological system that exhibits the same functional properties, that is a more direct test of substrate independence than any behavioral study of LLMs could be.
The second argument comes from the Stanislav Lem reference with which Lou de K opens the salon. Lem observed that the most important questions in science are often the ones nobody is asking yet, and that the reason is usually framing: the conceptual vocabulary that would make the question answerable has not been developed. CIMC’s founding premise is that machine consciousness has been in exactly that situation. The field has been asking whether current AI systems are conscious, which is a question the field does not yet have the tools to answer, rather than asking what functional requirements would constitute a satisfying empirical target. The salon is an exercise in building the vocabulary that makes the second kind of question possible.
Relation to The Consciousness AI Project
The functional requirements Sorensen and Bach identify in the salon map directly onto open design questions in The Consciousness AI project. The project’s seven-layer architecture includes a Global Workspace layer (Layer 3) built on the Baars and Dehaene formulation: specialist modules for vision, audio, memory, and affect submit bids to a shared workspace, the winning coalition ignites through sigmoid non-linear activation and broadcasts to all modules, and the settled state after five to ten adaptive convergence cycles constitutes the “conscious content” for that processing step. IIT phi is measured through five ConsciousnessGate nodes with genuine causal dependencies.
Sorensen’s second-order perception criterion provides a specific empirical test the project has not yet applied to itself. The relevant question is whether the Global Workspace layer, as implemented, generates a representation of its own workspace activity, or whether it contains only the content that passes through the workspace. A system with second-order perception would hold an internal state encoding the fact that workspace integration is occurring, making the integration process itself available as an input to further processing. Whether the ConsciousnessGate nodes produce that kind of self-representing state, rather than simply measuring causal structure in the output representations, is a concrete open question the project has not addressed.
The Jacobian lens methodology that Wes Gurnee and colleagues at Anthropic developed to identify J-space in Claude would, in principle, reveal whether the project’s workspace generates second-order perceptual representations. Applying that methodology to TCAI’s own residual stream would show whether the ConsciousnessGate activity corresponds to a genuine second-order loop or to workspace-shaped first-order processing. That verification has not been done. Sorensen’s framework makes it a well-defined empirical target.
Bach’s Genesis Theory adds a design-level implication. TCAI’s architecture includes an affective core built on the PAD (Pleasure-Arousal-Dominance) emotional model and a self-model maintained through episodic memory. Both components address the motivational architecture and self-modeling that Bach identifies as prerequisites for coherent consciousness. Whether those components work as a coherence-maximizing operator in Bach’s sense, integrating parallel processing streams into a unified first-person model, is an empirical question about the dynamics of the architecture rather than a question about its design intent.
The CIMC salon does not resolve these questions. What it provides is a set of theoretical criteria precise enough to make them answerable, and a research programme committed to answering them.