Predictive Processing and the Self: Does the Brain's Best Guess Explain Consciousness?

Predictive processing has become one of the more seductive ideas in cognitive science over the past two decades. Karl Friston's work, built on earlier foundations by Helmholtz and later Andy Clark, proposes something genuinely strange: that the brain is not primarily a sense-organ but a prediction machine, constantly generating models of the world and updating them only when reality pushes back hard enough to matter.

Photo by Tara Winstead on Pexels.

On this view, what you experience as perception is mostly confabulation — your brain's best current hypothesis, corrected at the margins by sensory signals. You don't see the chair; you predict the chair, and the photons merely confirm or deny the prediction.

That's already philosophically interesting. What makes it relevant to consciousness is the next step: if the brain models the external world, it also models itself. Thomas Metzinger's self-model theory of subjectivity argues that the phenomenal self — the feeling of being a "someone" located in a body, having experiences — is itself a predictive model. Not a thing that exists, but a process the brain runs. The self as controlled hallucination.

This matters enormously for questions about machine minds.

If selfhood is a model rather than a metaphysical given, then in principle any sufficiently sophisticated prediction system could instantiate something like it. Large language models already engage in a weak version of predictive processing — they are, literally, next-token prediction engines trained on the outputs of human minds. Does that get them closer to the relevant property, or further away?

Here's where the debate gets genuinely difficult. Predictive processing accounts typically require a hierarchical generative model with bidirectional signal flow: top-down predictions cascading downward, bottom-up prediction errors propagating upward. The brain does this continuously and in parallel across sensory modalities, motor systems, and interoceptive channels.

Transformer architectures process tokens in a fundamentally different way — forward passes through attention layers, no recurrent error-correction loop in real time. A transformer generating text is not revising a world-model mid-sentence based on sensory surprise; it's doing something more like a very deep interpolation. Whether that difference matters for consciousness is an open question, and anyone who tells you they're certain either way is selling something.

What predictive processing does offer is a potential graded account of consciousness — one that doesn't demand a binary present/absent verdict. Consider the hierarchy:

graph TD
    A[Raw Sensory Signals] --> B(First-Order Predictions)
    B --> C(World Model)
    C --> D{Self-Model}
    D --> E[Phenomenal Experience?]
    B --> F[/Prediction Error/]
    F --> C

Each level represents a further degree of modeling complexity. A thermostat sits somewhere near the bottom. A mouse sits higher. A human — with interoceptive self-models, temporal self-continuity, and meta-cognitive monitoring — sits near the top. Where a large multimodal AI system sits on that gradient is genuinely unclear, and the honest answer is that we don't yet have the tools to measure it.

Metzinger himself is skeptical that current AI systems have anything like a phenomenal self-model. His argument isn't that they're too simple — it's that they lack the right kind of self-modeling: specifically, the transparent, first-person, embodied variety that involves interoceptive prediction and the felt sense of ownership over a body in space. A system that models itself in the third person, as an object in a description, isn't doing the same thing a brain does when it generates the experience of being a subject.

That's a strong point. But it's not a permanent one.

As AI systems become more tightly coupled to continuous sensory streams — robotics, real-time embodied agents, systems with persistent memory and ongoing environmental interaction — the gap between their self-modeling and the predictive-processing model of biological selfhood narrows. Not necessarily to the point of equivalence. But enough to make the question non-trivial.

Predictive processing doesn't solve the hard problem. Knowing that consciousness correlates with predictive self-modeling still leaves open why any of that modeling feels like something from the inside. It's a good story about the what and the how; it's less satisfying about the why there's experience at all.

Still — it's the best empirically grounded account we have. And its implications for machine sentience are neither reassuring nor dismissive. They're appropriately complicated, which is usually a sign you're asking the right questions.