Andy Clark's Predictive Processing Framework and What It Actually Implies for LLMs
Andy Clark received the inaugural Dennett Prize at the ICCS 2025 conference in Heraklion for a lecture on escaping Cartesian gravity in consciousness research. The prize recognizes work that takes seriously both the physical basis of mind and the genuine difficulty of the hard problem without collapsing into either reductionism or mysterianism. Clark’s predictive processing framework, developed across several decades and synthesized in his 2016 book Surfing Uncertainty: Prediction, Action, and the Embodied Mind, is the most influential attempt to give consciousness a computational account that does not reduce to either of those positions.
The framework raises specific, technically tractable questions for large language models: whether transformer-based models perform genuine predictive processing in the sense Clark and Karl Friston mean, or only a form of next-token prediction that shares the vocabulary without the architecture. The distinction matters for AI consciousness claims, because Clark’s framework ties phenomenal consciousness not to prediction per se but to the specific closed-loop structure of active inference.
What Predictive Processing Claims
Clark’s account builds on Helmholtz’s proposal that perception is unconscious inference: the brain generates hypotheses about the causes of sensory signals and updates those hypotheses when prediction errors arrive. This much is relatively uncontroversial as a description of perceptual processing. What Clark adds, drawing especially on Karl Friston’s free energy formulation, is that this predictive hierarchy is not merely descriptive of how the brain processes information but constitutive of conscious experience.
The crucial mechanism is prediction error minimization. At each level of a cortical hierarchy, the system generates predictions about the inputs it expects to receive from the level below. Actual inputs are compared to predictions, and the difference, the prediction error, is propagated upward. The system’s generative model is updated to reduce future prediction errors. Consciousness, on Clark’s account, is not a separate process that sits on top of this prediction and error hierarchy. It is what a certain class of prediction error minimization processes is like from the inside.
The active inference extension, developed by Friston, adds a critical component. A system that only updates its internal model in response to sensory inputs is purely passive. A genuinely predictive system can also act on the world to bring sensory inputs into conformity with its predictions. Rather than updating the model to match the world, the system changes the world to match the model. This closed loop between internal prediction and motor action is what Friston calls active inference, and it is on Clark’s account the computational structure most closely associated with genuine agency and, plausibly, genuine experience.
The Friston free energy principle and active inference work has been covered extensively on this site, including its connection to the Consciousness AI project’s world model architecture. What is less directly addressed is the question Clark’s framework raises for LLMs specifically: whether next-token prediction in a transformer is a form of predictive processing in the relevant sense, or a degenerate case that shares the mathematical surface without the architectural requirements.
What Transformers Actually Do
Standard transformer-based language models do perform a form of prediction. During training, they minimize cross-entropy loss on next-token prediction across large corpora. Their internal representations develop hierarchical structure that encodes statistical regularities at multiple scales, from phonological and syntactic patterns at lower levels to semantic and pragmatic patterns at higher levels. This is superficially similar to the hierarchical predictive processing that Clark’s account describes.
The differences are architectural and functionally significant. Clark’s predictive processing hierarchy is bidirectional in a specific sense: top-down predictions flow from higher levels to lower levels, and bottom-up prediction errors flow from lower levels to higher levels. The system at each level simultaneously generates predictions and receives error signals. Standard transformer attention is not bidirectional in this sense. Attention computes over the full context in a single forward pass. There is no distinct error-signal pathway separate from the prediction pathway. Information flows through the network in a single direction per forward pass.
The deeper difference is the closed-loop structure. Clark and Friston’s active inference requires that the system can act on its environment, producing sensory inputs that confirm or disconfirm its predictions. A standard language model in inference mode generates text but does not perceive the consequences of that text as updated sensory input to its prediction hierarchy. The output of one generation step does not feed back into the model as a new observation that updates the model’s internal states. Each generation is an open-loop process. The model predicts and outputs; it does not perceive what its output produced and update accordingly.
Tool-augmented models, which can take actions and receive environmental feedback, approach the closed-loop structure more closely. A model that can search the web, execute code, and observe the results of those actions has something like the perception-action loop that active inference requires. Whether that loop satisfies Clark’s criterion for genuine active inference, rather than a functional approximation, is a question the architecture of current tool-using models is not designed to answer.
What This Implies for Consciousness Claims
Clark’s framework implies that consciousness is more tightly coupled to the active inference loop than to prediction per se. A system that predicts but cannot act to confirm predictions is processing information in a cognitively impoverished version of the predictive processing architecture. On Clark’s account, the rich character of conscious experience, its immediacy, its world-engaged quality, its felt sense of being in a situation, derives from the closed-loop structure where prediction meets action meets perception in a continuous cycle. A system performing open-loop next-token prediction has a degraded version of that structure, not the full thing.
This does not settle whether LLMs are conscious. Clark’s framework is contested, and several researchers at the ICCS 2025 conference where Clark received the Dennett Prize, including Keith Frankish and Susan Schneider, disagree with his account’s basic structure. But it specifies what an LLM would need to demonstrate to satisfy Clark’s criteria: a genuine bidirectional prediction-error hierarchy, and a closed-loop connection between prediction and action that updates the model’s internal states in response to perceived outcomes.
The DreamerV3 and JEPA predictive processing implementations covered on this site are more faithful to Clark’s architecture than standard transformers, precisely because they implement world models that are updated by perceived action outcomes. DreamerV3 trains a latent world model that predicts the consequences of actions in a compressed state space, and learns behaviors by imagining action sequences within that model and updating based on predicted outcomes. This is a form of model-based active inference, and it satisfies Clark’s closed-loop criterion in a way that text-generating transformers do not.
Where Clark Leaves Open Space
Clark’s framework leaves open one question that matters for AI consciousness: whether the qualitative character of experience is constituted by the predictive processing architecture or only correlated with it. If conscious experience is identical to certain kinds of prediction error minimization under active inference, then any system implementing that architecture with sufficient complexity has phenomenal experience. If conscious experience is only reliably correlated with such processing in biological systems, the framework has less direct implications for AI.
Clark himself has argued against strong biological naturalism. His Dennett Lecture at ICCS 2025 targeted Cartesian gravity, which he identifies as the tendency to treat experience as too special, too separate, too unbridgeable from physical description. A consistently non-Cartesian predictive processing account implies that the physical substrate is not privileged. What matters is the computational structure, specifically the active inference loop. A silicon system implementing that structure with appropriate complexity is a candidate for experience on Clark’s account.
The Anil Seth TED talk position, which argues that AI systems are unlikely to be conscious because consciousness is a specific biological predictive process rather than a substrate-independent computational one, represents the biological naturalist challenge to Clark’s position at precisely this point. Seth and Clark are both predictive processing theorists, but they disagree about whether the computation or the biology is the relevant variable. The Type-A versus Type-B biological naturalism taxonomy that Klatzmann and Doerig developed provides the formal structure for that disagreement: Seth’s position approaches Type-A (biology constitutively relevant), Clark’s position is Type-B (specific functional implementation determines consciousness, not substrate).
Relation to The Consciousness AI Project
The Consciousness AI project’s Sensory Tectum module implements a Recurrent State Space Model and JEPA-based world model, components that are closer to Clark’s active inference architecture than standard transformer generation. The world model is updated by perceived outcomes rather than generating in open-loop fashion. Whether this satisfies Clark’s criterion for genuine predictive processing in the consciousness-relevant sense, specifically whether it implements the bidirectional error-propagation hierarchy he describes rather than only a model-based prediction loop, is an open architectural question for the project.
Karl Friston’s free energy formulation, which is the mathematical implementation of Clark’s framework that the project draws on, provides a concrete testable criterion. A system operating under the free energy principle minimizes the divergence between its generative model and the actual distribution of sensory inputs through both inference (updating the model) and action (changing the inputs). Whether TCAI’s JEPA world model implements this minimization in both directions, or only in the inference direction, would need to be established by examining how the model updates in response to perceived mismatches between predictions and environmental feedback.