The one-sentence version
Critics of Large Language Models (LLMs) often say: "It is just predicting the next word." However, the highly influential Predictive Processing Theory in cognitive science suggests that the human brain is essentially also a prediction machine, constantly predicting the next moment of sensory input and only passing "prediction errors" up the hierarchy. So, what is the difference between the human brain and an LLM? This article's answer is: both share the core principle of "prediction," but they differ fundamentally in what they predict (multisensory body and world vs. text), why they predict (to maintain survival vs. to minimize training loss), whether they can test predictions through action (yes vs. essentially no), and structure (hierarchical bidirectional vs. primarily feedforward). These differences point exactly toward the potential directions for improvement in next-generation models.
Positioning
Theoretical comparison and commentary. Predictive processing theory has a wide influence in cognitive science but is also subject to criticism; analogies between neuroscience and LLMs must be handled with caution.
Layer 1: The Intuition
1.1 The Predictions You Don't Notice
- You walk down the stairs, expecting one more step, but there isn't one; your foot drops into empty space, and you jump in surprise;
- You read "Today the weather is very..." and your brain has already prepared "good" or "cold";
- In a familiar room, the clock stops, and it takes you a long time to realize "something sounds missing."
These moments demonstrate that: the brain is constantly predicting in advance, and you only "notice" when a prediction fails.
1.2 The Core Landscape of Predictive Processing
| Traditional View | Predictive Processing View |
|---|---|
| Senses send information into the brain; the brain processes and recognizes it layer by layer | The brain sends predictions top-down; senses send back only the prediction error from the bottom up |
| Perception = processing input | Perception = the brain's best guess, constantly corrected by input |
| Action happens after perception | Action is another way to make the world conform to predictions |
Layer 2: Theoretical Foundations
| Source | Content |
|---|---|
| Helmholtz (19th Century) | Perception is "unconscious inference" |
| Rao & Ballard (1999) | Predictive coding model of the visual cortex: higher levels predict lower levels, lower levels pass back errors |
| Friston's Free Energy Principle (since 2006) | Generalizing predictive processing as a general principle for the brain and even life: minimizing surprise (see "Free Energy Principle as a Training Objective") |
| Andy Clark, Surfing Uncertainty (2016) | A systematic philosophical exposition of predictive processing |
| Anil Seth, Being You (2021) | Explaining conscious experience (including self and emotion) as the brain's "controlled hallucination," emphasizing prediction of internal bodily states |
| Comparative studies of Neuroscience and LLMs | Multiple studies have found that the internal representations of language models can, to some extent, predict brain activity when humans listen to stories or read text; the human brain indeed predicts upcoming words during language comprehension |
The last point is particularly interesting: in the narrow domain of language, the predictions of LLMs show measurable correspondences with the human brain's language areas.
Layer 3: Similarities and Differences
3.1 Similarities
- Both are driven by prediction error for learning;
- Both can allow an understanding of structure (grammar, semantics, common sense) to emerge from prediction;
- Both can be viewed as constructing a generative model: an internal model capable of "generating" possible inputs.
3.2 Fundamental Differences
| Dimension | Human Brain | LLM |
|---|---|---|
| What to predict | Multisensory continuous signals such as vision, hearing, touch, and internal bodily states (heartbeat, hunger) | Text tokens (multimodal models extend this to images, audio) |
| Why predict | To maintain bodily homeostasis and survival; prediction errors have real costs | To minimize training loss; prediction errors are just gradients |
| Ability to act | Can change input and test predictions through action (active inference) | Completely passive during pre-training; only limited action exists in Agents and RL |
| Structure | Hierarchical and bidirectional (massive top-down feedback connections) | Single-pass forward computation is primarily bottom-up; "top-down" only manifests in the loop of token-by-token generation |
| Timing of learning | Lifelong, continuous learning | Frozen after training (see "Learning at Test Time") |
| Embodiment | Has a body, with an internal environment that needs maintaining | None |
3.3 What these differences mean
Inverting these differences one by one provides a roadmap for improving LLMs, which corresponds to several articles in this llm-frontier series:
| Difference | Corresponding Research Direction |
|---|---|
| Only predicts text → Predicts world states | World model post-training |
| Passive prediction → Testing through action | Interventional training, Free Energy Principle as a training objective |
| "Thinking" only at the token level → Hierarchical internal prediction | Latent space reasoning |
| Frozen after training → Lifelong learning | Learning at test time |
| No intrinsic "caring" → Goals that need maintaining | Controversial: Will this bring new security risks? |
3.4 Returning to that criticism
The phrase "It is just predicting the next word"—if predictive processing theory is correct, then "just predicting" is not inherently derogatory: the human brain "just predicts" as well. The truly meaningful questions are:
- Is its scope of prediction wide enough?
- Are its predictions linked to action and consequences?
- Do its predictions serve some real need?
In these three aspects, current LLMs still have a massive gap compared to the human brain. The gap itself is the direction.
3.5 A Warning
The analogy between neuroscience and AI is highly inspiring, but it is also easy to misuse:
- Artificial neural networks and biological neural networks differ vastly in detail;
- "The brain is a prediction machine" is itself a theory still being tested, not a settled fact;
- One cannot conclude from "both are predicting" that "both have consciousness" or "both have understanding" (see "AI Consciousness Theory Checklist").
Summary
| Layer | Key points |
|---|---|
| Theory | The brain predicts top-down and passes errors up bottom-up; action is also a way to reduce error |
| Similarities | Both learn via prediction error; both construct generative models; language prediction has measurable correspondences |
| Differences | Object of prediction, motivation, action, structure, timing of learning, embodiment |
| Significance | The list of differences = the direction for the next generation of models; "just predicting" is not inherently derogatory |