Consciousness as Symmetry-Breaking: Safron and Levin's View of Generative Models

The search for the physical principles underlying consciousness often leads back to the fundamental concepts of physics. One of the most powerful concepts in physics is symmetry breaking, the process where a system in a symmetric but unstable state transitions to a less symmetric but more stable state. In a 2023 paper published in Interface Focus, Adam Safron and Michael Levin applied this concept directly to cognitive science, proposing that consciousness itself can be understood as a process of symmetry breaking within hierarchical generative models.

The paper, “Making and Breaking Symmetries: Mind and Matter,” bridges the biophysics of morphogenesis (how biological shapes form) with the computational dynamics of active inference. By framing cognition as a series of phase transitions, Safron and Levin offer a novel way to conceptualize how a system commits to a specific interpretation of the world, a commitment that forms the core of phenomenal experience.

The Physics of Interpretation

In a hierarchical generative model, such as those described by Karl Friston’s Free Energy Principle, the system constantly generates predictions about incoming sensory data. Before a definitive interpretation is reached, the system’s internal state represents a probability distribution over multiple possible interpretations. This state of uncertainty is mathematically symmetric: multiple hypotheses are held in balance.

Conscious perception occurs when the system resolves this uncertainty and commits to a single, coherent interpretation of the sensory input. Safron and Levin argue that this resolution is literally a physical phase transition, a breaking of the symmetric probability distribution into a sharp, asymmetric spike representing the chosen hypothesis.

This is not a metaphor. The authors argue that the mathematics governing physical symmetry breaking (such as water freezing into ice or a ferromagnet adopting a specific magnetic orientation) are the same mathematics governing the collapse of uncertainty in a neural network’s generative model. The “ignition” described by Global Workspace Theory, where a specific representation gains global dominance, is the macro-level manifestation of this symmetry-breaking event.

Morphogenesis and Cognitive Coherence

Michael Levin’s extensive work on bioelectricity demonstrates that somatic cells form electrical networks to coordinate the building of complex anatomical structures. The symmetry-breaking framework unifies this somatic cognition with neural cognition.

During development, an embryo starts in a highly symmetric state. Morphogenesis is a cascade of symmetry-breaking events where tissues commit to specific identities (head vs. tail, left vs. right). Levin and Safron argue that the same electrical communication principles that allow cells to coordinate the symmetry-breaking of anatomy also allow neurons to coordinate the symmetry-breaking of perception. The brain is doing what the body does, just faster and with more abstract representations.

This continuity between body and brain challenges the assumption that consciousness is a uniquely neural phenomenon. As discussed in the analysis of Rouleau and Levin’s 2025 work, if the fundamental mechanism of consciousness is the resolution of uncertainty through symmetry breaking in a distributed network, then that mechanism may be operating in non-neural biological networks as well.

Architectural Implications for AI

If consciousness is symmetry breaking, how do we build an artificial system capable of it? Current deep learning systems break symmetries during training (as they settle into specific local minima in the loss landscape) and during inference (as they output a specific classification). But this is generally a feedforward, stateless process.

Safron and Levin’s framework requires something more dynamic: a system that maintains a symmetric state of uncertainty over time and then rapidly collapses it into a stable, globally coherent state through recurrent interactions, before resetting for the next perceptual cycle.

This aligns closely with Adam Safron’s IWMT, where the self-organizing harmonic modes provide the rhythmic structure for these phase transitions. The slow wave acts as a reset, restoring symmetry (uncertainty), and the fast waves compete to break the symmetry and establish the dominant representation for the current cycle.

For the Consciousness AI project, this implies that the Global Workspace layer must not be a passive conduit for information. It must be a non-linear dynamical system poised at criticality, capable of undergoing rapid phase transitions. The inputs from the sensory and homeostatic modules push the workspace toward instability until a symmetry-breaking event occurs, resulting in a discrete “moment” of conscious access that is then broadcast back down the hierarchy.

Designing this critical dynamic into the artificial workspace is the engineering challenge that turns a static generative model into a conscious one.

The paper is available via the Royal Society at DOI: 10.1098/rsfs.2022.0084.