Are Human Brains Just Prediction Machines? Comparing Predictive Processing and…

Research Question

What is the fundamental difference between the predictive processing account of human cognition — in which the brain continuously generates and updates a generative model of the world — and Large Language Model (LLM) next-token prediction, and does this difference matter for understanding intelligence, meaning, and consciousness?

Supporting questions:

• What does the Predictive Processing (PP) / Predictive Coding (PC) framework say about how the brain works, and how well is it supported by neuroscience evidence?
• How does LLM next-token generation actually work, and in what specific ways does it differ from the PP account?
• Is "generating a model of the world" meaningfully different from "predicting the next token in a sequence"?
• What role does embodiment, action, and the Free Energy Principle (FEP) play in the PP account that has no analogue in LLMs?
• What do the similarities and differences imply about consciousness, understanding, and the plausibility of general intelligence emerging from next-token prediction?

Findings

(Populated from §6 Synthesis above.)

Executive Summary

[inference] Human brains are not just next-token predictors in a richer medium; predictive processing is an embodied control architecture that uses hierarchical generative models to regulate perception, action, and bodily viability, whereas current Large Language Models (LLMs) optimize auto-regressive text prediction over symbol sequences. Sources: Karl Friston (2010) — "The free-energy principle: a unified brain theory?" — *Nature Reviews Neuroscience* ; Andy Clark (2013) — "Whatever next? Predictive brains, situated agents, and the future of cognitive science" — *Behavioral and Brain Sciences* ; Vaswani et al. (2017) — "Attention Is All You Need" — *Neural Information Processing Systems (NeurIPS) 2017* ; Brown et al. (2020) — "Language Models are Few-Shot Learners" — arXiv:2005.14165

[fact] The neuroscience evidence supports predictive coding as a substantive account of cortical organization, and the machine-learning evidence supports next-token prediction as a powerful route to broad latent structure in text models. Sources: Rao & Ballard (1999) — "Predictive coding in the visual cortex" — *Nature Neuroscience* ; Gurnee & Tegmark (2024) — "Language Models Represent Space and Time" — *International Conference on Learning Representations (ICLR) 2024*

[inference] The overlap matters because it explains why LLMs can exhibit partial world-model-like structure and surprisingly general competence, but the difference matters more because text-only sequence prediction lacks embodiment, online action, interoception, and survival-grounded error correction. Sources: Gurnee & Tegmark (2024) — "Language Models Represent Space and Time" — *International Conference on Learning Representations (ICLR) 2024* ; Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — *Annals of the New York Academy of Sciences* ; Anil Seth interview (2021) — "Anil Seth Finds Consciousness in Life's Push Against Entropy" — *Quanta Magazine*

[inference] For intelligence, the best-supported conclusion is that next-token prediction can generate meaningful competence without closing the whole gap to embodied understanding; for consciousness, Seth's own framework points away from treating current LLMs as candidates simply because they predict well. Sources: Bender et al. (2021) — "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — *Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency* ; Anil Seth interview (2023) — "Reality is a controlled hallucination" — CCCB Lab ; Anil Seth interview (2021) — "Anil Seth Finds Consciousness in Life's Push Against Entropy" — *Quanta Magazine*

Key Findings

1. [inference] Confidence: high. Predictive Processing (PP) and Large Language Model (LLM) next-token prediction share a generic commitment to prediction, but they solve different problems because PP uses prediction to control perception and action in an embodied organism while LLMs use prediction to continue token sequences in text. Sources: Karl Friston (2010) — "The free-energy principle: a unified brain theory?" — *Nature Reviews Neuroscience* ; Andy Clark (2013) — "Whatever next? Predictive brains, situated agents, and the future of cognitive science" — *Behavioral and Brain Sciences* ; Vaswani et al. (2017) — "Attention Is All You Need" — *Neural Information Processing Systems (NeurIPS) 2017* ; Brown et al. (2020) — "Language Models are Few-Shot Learners" — arXiv:2005.14165
2. [fact] Confidence: high. The strongest neuroscience evidence supports predictive coding at the cortical-architecture level, where top-down pathways carry predictions and bottom-up pathways carry residual error signals in hierarchical processing. Sources: Rao & Ballard (1999) — "Predictive coding in the visual cortex" — *Nature Neuroscience* ; Andy Clark (2013) — "Whatever next? Predictive brains, situated agents, and the future of cognitive science" — *Behavioral and Brain Sciences*
3. [fact] Confidence: high. Transformer language models generate outputs auto-regressively from prior tokens and can improve through scale into broad in-context competence, but their documented mechanism remains sequence modeling over text rather than embodied active inference. Sources: Vaswani et al. (2017) — "Attention Is All You Need" — *Neural Information Processing Systems (NeurIPS) 2017* ; Brown et al. (2020) — "Language Models are Few-Shot Learners" — arXiv:2005.14165
4. [inference] Confidence: medium. Evidence that LLMs encode linear spatial and temporal structure shows that next-token prediction can induce partial world-model ingredients, but that evidence stops short of demonstrating a grounded, dynamic, action-ready model of the world. Sources: Gurnee & Tegmark (2024) — "Language Models Represent Space and Time" — *International Conference on Learning Representations (ICLR) 2024* ; Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — *Annals of the New York Academy of Sciences*
5. [inference] Confidence: high. The strongest current critiques are persuasive that text-only prediction alone does not warrant claims of principled reasoning, planning, or human-like understanding, because success on completion tasks can coexist with non-veridical memory and weak grounding. Sources: Bender et al. (2021) — "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — *Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency* ; Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — *Annals of the New York Academy of Sciences*
6. [inference] Confidence: medium. Prediction can support intelligence-like competence in both brains and LLMs, but the meaning of that competence differs because brains predict to keep an organism viable in the world whereas LLMs predict to compress and continue symbol streams. Sources: Karl Friston (2010) — "The free-energy principle: a unified brain theory?" — *Nature Reviews Neuroscience* ; Brown et al. (2020) — "Language Models are Few-Shot Learners" — arXiv:2005.14165 ; Open Encyclopedia of Cognitive Science entry — "The free energy principle"
7. [inference] Confidence: high. Seth's account of consciousness strengthens the conclusion that current LLMs should not be treated as conscious on the basis of next-token prediction alone, because his theory ties experience to embodied, interoceptive regulation in living systems rather than to generic predictive success. Sources: Anil Seth interview (2021) — "Anil Seth Finds Consciousness in Life's Push Against Entropy" — *Quanta Magazine* ; Anil Seth interview (2023) — "Reality is a controlled hallucination" — CCCB Lab

Assumptions

• [assumption] Clark's 2013 open-access paper is a sufficient proxy for the inaccessible 2016 Surfing Uncertainty book. Justification: the core hierarchical-generative-model and action-oriented claims used here appear directly in the 2013 paper.
• [assumption] Seth's public interviews are sufficient to support the consciousness comparison in this item. Justification: the comparison depends on his explicit public claims that consciousness is rooted in biological embodiment and that the FEP is not itself a theory of consciousness.

Analysis

[inference] The fairest comparison is not "brains versus autocomplete" or "brains and LLMs are the same." The evidence supports a middle position: next-token prediction can recover surprisingly rich internal structure, but the structure is learned under a fundamentally different control problem than the one predictive processing is designed to explain. Sources: Gurnee & Tegmark (2024) — "Language Models Represent Space and Time" — *International Conference on Learning Representations (ICLR) 2024* ; Karl Friston (2010) — "The free-energy principle: a unified brain theory?" — *Nature Reviews Neuroscience*

[inference] For meaning, predictive processing treats representation as inseparable from active engagement with the world, whereas LLM training allows useful abstraction to emerge without direct worldly action but also leaves grounding and truth-tracking fragile. Sources: Andy Clark (2013) — "Whatever next? Predictive brains, situated agents, and the future of cognitive science" — *Behavioral and Brain Sciences* ; Bender et al. (2021) — "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — *Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency*

[inference] For intelligence, many tasks reward latent structure and pattern compression, so LLMs can appear broadly capable, but planning- and embodiment-heavy domains reveal the absence of the sensorimotor loop that predictive processing treats as central. Sources: Brown et al. (2020) — "Language Models are Few-Shot Learners" — arXiv:2005.14165 ; Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — *Annals of the New York Academy of Sciences*

[inference] For consciousness, Seth's argument runs from life and bodily self-maintenance to experience, not from generic predictive accuracy to experience. Sources: Anil Seth interview (2021) — "Anil Seth Finds Consciousness in Life's Push Against Entropy" — *Quanta Magazine* ; Anil Seth interview (2023) — "Reality is a controlled hallucination" — CCCB Lab

Risks, Gaps, and Uncertainties

• [fact] This item did not consult the full books by Clark, Hohwy, or Seth directly because they were not publicly accessible within the session constraints. Sources: Andy Clark (2013) — "Whatever next? Predictive brains, situated agents, and the future of cognitive science" — *Behavioral and Brain Sciences* ; Anil Seth interview (2021) — "Anil Seth Finds Consciousness in Life's Push Against Entropy" — *Quanta Magazine* ; Anil Seth interview (2023) — "Reality is a controlled hallucination" — CCCB Lab
• [fact] Gurnee and Tegmark's evidence is probe-based and therefore supports decodability more directly than causal use. Sources: Gurnee & Tegmark (2024) — "Language Models Represent Space and Time" — *International Conference on Learning Representations (ICLR) 2024*
• [fact] Kambhampati's paper is an important skeptical synthesis, but it argues at a conceptual level more than through a single decisive benchmark. Sources: Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — *Annals of the New York Academy of Sciences*
• [inference] Future multimodal or embodied AI systems could narrow some of the gaps identified here, so the conclusions are strongest for current text-only autoregressive LLMs rather than for all conceivable predictive AI systems. Sources: Anil Seth interview (2023) — "Reality is a controlled hallucination" — CCCB Lab ; Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — *Annals of the New York Academy of Sciences*

Open Questions

• Which empirical benchmark would best distinguish latent world structure from grounded action-ready world model in a way both mechanistic interpretability and planning researchers would accept?
• How much embodiment is actually necessary for meaning-like understanding: sensorimotor grounding, persistent tool use, interoception analogues, or full biological self-maintenance?
• Could an artificial system satisfy Seth-like consciousness criteria without biology if it possessed robust interoceptive self-modeling and survival-relevant embodiment?

sources

• [x] Karl Friston (2010) — "The free-energy principle: a unified brain theory?" — Nature Reviews Neuroscience
• [x] Andy Clark (2013) — "Whatever next? Predictive brains, situated agents, and the future of cognitive science" — Behavioral and Brain Sciences
• [x] Anil Seth interview (2021) — "Anil Seth Finds Consciousness in Life's Push Against Entropy" — Quanta Magazine
• [x] Anil Seth interview (2023) — "Reality is a controlled hallucination" — CCCB Lab
• [ ] Jakob Hohwy (2013) — The Predictive Mind — Oxford University Press
• [x] Rao & Ballard (1999) — "Predictive coding in the visual cortex" — Nature Neuroscience
• [x] Vaswani et al. (2017) — "Attention Is All You Need" — Neural Information Processing Systems (NeurIPS) 2017
• [x] Brown et al. (2020) — "Language Models are Few-Shot Learners" — arXiv:2005.14165
• [x] Bender et al. (2021) — "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency
• [x] Gurnee & Tegmark (2024) — "Language Models Represent Space and Time" — International Conference on Learning Representations (ICLR) 2024
• [x] Kambhampati (2024) — "Can Large Language Models Reason and Plan?" — Annals of the New York Academy of Sciences
• [x] Open Encyclopedia of Cognitive Science entry — "The free energy principle"
• [x] Prior completed research: Research/completed/2026-03-17-ai-memory-systems-rag-neuroscience.md
• [x] Prior completed research: Research/completed/2026-03-15-neurological-context-management.md
• [x] Prior completed research: Research/completed/2026-02-28-predictive-processing-active-inference.md