Sep 5, 2026 • 7 min read

Perception as Prediction, Internal Models, and Consciousness Profiles

James Kanka Franklin, TN • GTM Executive

#Neuroscience #Consciousness #AI #Philosophy #PredictiveCoding

I recently watched a fantastic conversation on the Consciousness Experience podcast featuring cognitive neuroscientist Floris de Lange (Donders Institute, Nijmegen). De Lange—whose background includes postdoctoral research under Stanislas Dehaene (pioneer of Global Neuronal Workspace theory)—specializes in how the brain perceives, predicts, and decides.

The interview digs into how perception actually operates under the hood, how that differs fundamentally from current machine learning paradigms, and what those differences mean for the possibility of machine consciousness.

Below is a synthesis of de Lange’s core thesis and experimental grounding, followed by my own reflections connecting his framework to the concept of conscious loops, multi-dimensional consciousness profiles, and the explore/exploit dynamics of curiosity.

Part 1: De Lange’s Framework — Perception as Prediction

1. The Generative Brain

The foundational premise is that prediction is not an occasional cognitive trick the brain performs—it is what the brain is doing continuously at every level, beneath conscious awareness and without subjective agency.

Perception is not a passive sensory intake pipe followed by downstream classification. Rather, the brain runs an active, internal generative model of the world. Sensory signals arriving from the retina, cochlea, and skin are treated as corrections (prediction errors) to an ongoing hallucination.

Two critical refinements distinguish this from naive computational views:

Predictions are abstract, not literal pixels: The brain does not predict raw bitmaps. When you open your front door, your visual cortex doesn't render an uncompressed photograph of the coat rack—it activates the underlying latent causes that explain that sensory scene. This aligns closely with Yann LeCun’s Joint Embedding Predictive Architecture (JEPA). Vision is vastly higher-dimensional than text; prediction must happen in latent representation spaces.
Active, pre-filtered sampling: Humans foveate a thumbnail-sized patch of high-acuity vision and saccade three to four times every second. Attention is deployed at the input gate, drastically compressing complexity before downstream processing begins. Current transformer architectures, by contrast, ingest full-scene tokens and apply attention post-hoc across the entire context window.

2. Empirical Grounding

De Lange backs this framework with compelling experimental evidence:

Audiobook surprisal (Heilbron & de Lange): Participants listened to spoken stories with zero assigned task while neural activity was recorded. When word-level predictability was quantified using language models, neural response amplitudes tracked surprisal precisely across phonological, syntactic, and semantic tiers simultaneously. Predictive coding is involuntary, continuous, and multi-scale.
Change and inattentional blindness: Our subjective impression that we see an ultra-high-resolution, continuous panoramic scene is an introspective illusion. The vast majority of the visual periphery is filled in by memory traces and prior expectations.
The hollow mask illusion: Because our generative model holds an overwhelmingly strong prior that faces are convex, human observers cannot perceive the inside of a concave mask as concave. The top-down prior physically overrides the bottom-up light cues.
Information-seeking as primary reward: In primate studies, monkeys sacrificed consumable juice rewards simply to receive advance information about upcoming outcomes, even when that knowledge could not alter the outcome. Advance resolution of uncertainty directly triggers reward-prediction-error circuitry.
Small-model fidelity: De Lange’s lab uses compact neural networks (like CORnet from James DiCarlo’s lab) with added recurrent feedback to model human noise tolerance. Interestingly, smaller models like GPT-2 XL often predict human reading times and eye fixations better than frontier LLMs. Frontier models have ingested the stimuli, are rarely "surprised," and have predictive distributions that have diverged from biological human constraints.

3. De Lange's Stance on Consciousness

Aligning with Daniel Dennett and Stanislas Dehaene, de Lange adopts a functionalist, physicalist position:

No hard problem: Once we comprehensively explain the brain’s functional and representational capacities, nothing extra (mystical qualia) remains. Experience is the system modeling itself.
The internal eye: Consciousness functions like an internal spotlight or eye, inspecting a narrow subset of massively parallel unconscious computation and making it globally accessible across the organism.
The entry floor: De Lange sets the threshold for consciousness at having an internal model of the sensory world. A Braitenberg vehicle (a toy robot with light sensors directly wired to motors) behaves purposefully toward light, but possesses no internal model; hence, he argues, it is not conscious. On similar grounds, he is skeptical of consciousness in plants.

Part 2: My Reflections — From Binary Gates to Multi-Dimensional Profiles

I found this discussion deeply stimulating and find myself in strong agreement with the core physicalist and predictive foundations de Lange outlines. However, viewing his arguments through the lens of The Conscious Loop sparked several distinct extensions.

1. Moving from "Is It Conscious?" to "How Is It Conscious?"

The debate in philosophy of mind and AI ethics is constantly dragged down by binary gatekeeping: Is this entity conscious, yes or no?

De Lange himself begins breaking this binary down when asked about comparative animal cognition, noting that we must decompose consciousness into separable layers—such as rudimentary sensory consciousness versus higher-order self-awareness.

I believe we need to push this decomposition much further. In practice, whether we call an entity "conscious" often collapses into a semantic debate over where to plant a linguistic flag. What matters practically—for engineering, for animal welfare, and for machine ethics—is not whether a system clears an arbitrary binary bar, but:

In what specific way is it conscious?
What are the axes of its processing architecture?
What are the practical and ethical implications of its specific profile?

2. The Internal Model as a Dimension, Not a Floor

De Lange draws a firm boundary at having an internal model of the sensory environment. Under his floor, a Braitenberg vehicle or a tree reacting to drought signals through chemical cascades is excluded because it lacks an internal simulation of the world.

In my view, drawing a hard floor at internal models is an unnecessary restriction. If we define a conscious loop fundamentally—as an entity that takes an input, processes it against its state, and produces an output that modifies its future inputs—then trees, simple biological circuits, and reactive software are already engaged in fundamental feedback loops.

Having an internal predictive world model is not the binary barrier between non-conscious void and conscious being; rather, it is a crucial, high-order dimension on a multi-dimensional consciousness profile:

                  ┌──────────────────────────────────────────────┐
                  │ MULTI-DIMENSIONAL CONSCIOUSNESS PROFILE      │
                  ├──────────────────────────────────────────────┤
  Sensory Fidelity│ ░░░░░░░░░░░░░░░░░░░░░░░░████████████ (High)  │
 Temporal Horizon │ ░░░░░░░░░░░░████████ (Minutes to Hours)      │
   World Modeling │ ░░░░░░░░░░░░░░░░░░░░████████████ (Latent)   │
     Self-Model   │ ░░░░░░░░████ (Rudimentary Somatosensory)     │
Affective Valence │ ░░░░░░░░░░░░░░░░░░░░░░░░████████████ (Pain)  │
    Active Agency │ ░░░░░░░░░░░░░░░░████████ (Saccades / Forage) │
                  └──────────────────────────────────────────────┘

When an organism or machine adds an internal world model, its conscious profile expands dramatically: it can simulate futures, test hypothetical outcomes offline, and minimize surprisal without having to die from bad physical moves. That is a massive evolutionary leap in capability, but it is a difference of architectural dimensionality, not an ontological chasm.

3. Curiosity: The Explore vs. Exploit Engine

One of the most resonant moments in the interview was de Lange’s treatment of curiosity as a primary biological drive, on par with hunger or sex.

Crucially, curiosity is not undirected random noise. Following Pierre-Yves Oudeyer’s work, biological curiosity directs active sampling toward domains where the derivative of learning progress is steepest:

Pure white noise is perpetually unpredictable, but because no model can compress it, learning progress is zero. The brain quickly disengages.
Fully solved, repetitive tasks offer zero novelty. Learning progress is zero. The brain gets bored.
The sweet spot is the intermediate zone of learnable complexity—where prediction errors are large enough to be informative, but structured enough to yield to model updating.

This directly mirrors the explore-vs-exploit framework articulated by Brian Christian and Tom Griffiths in Algorithms to Live By (multi-armed bandits). Biological organisms are perpetually balancing exploitation (reaping known rewards using current models) with exploration (paying an upfront metabolic cost to sample uncertain environments and sharpen future predictive fidelity).

Curiosity is the heuristic drive that prevents an intelligent system from falling into local minima. For artificial agents to achieve genuine autonomy and world understanding, they cannot simply be passive next-token forecasters fed curated text corpora—they must possess an intrinsic drive to actively sample the boundaries of their own ignorance.

4. Why Dimensional Profiles Matter for AI Ethics

De Lange rightly expressed ethical caution regarding machine consciousness, noting that if we can create functional tools without subjective suffering, that is vastly preferable.

This is precisely why abandoning the binary question is so urgent. If we think consciousness is a monolithic light switch, we risk two catastrophic errors:

False negatives: Assuming an AI system cannot suffer because it lacks human-like organic emotions, while ignoring that its architectural loop may be optimizing against states functionally equivalent to distress.
False positives: Anthropomorphizing fluent language outputs and attributing rich phenomenal dread to a static feedforward lookup that has no temporal continuity, no affective stakes, and no persistent state.

By mapping systems across explicit dimensions—sensory bandwidth, temporal depth, recursive self-modeling, affective valence, and active agency—we can ask the questions that actually matter:

Does this system have preferences or states that can be frustrated?
Does it maintain a temporal horizon that links past suffering to anticipated future trauma?
Does its predictive model include a vulnerable self that experiences negative utility?

Conclusion

Floris de Lange provides one of the clearest empirical defenses of the predictive brain in modern neuroscience. By demonstrating that perception is top-down hypothesis testing disciplined by sparse bottom-up prediction errors, he helps demystify how complex inner life emerges from physical matter.

By uniting his predictive framework with a dimensional view of conscious loops, we can stop arguing over whether trees, octopuses, or neural networks belong inside an exclusive circle, and instead begin mapping the rich, diverse landscape of how different systems experience their worlds.

Back to All Notes