Evolutionary Blueprint (2). Your brain predicts the world, but it is not ChatGPT
Series: Evolutionary Blueprint of AI, article 2. Predictive coding offers a useful, contested account of perception. Learn where the analogy with language models helps, where it breaks, and what it means for dependable business AI.
Who should read it?
Business leaders, product teams, and curious readers without a technical background
Metapher
The anticipated landscape
Your expectations meet the world
You reach for your coffee cup without looking closely. It is where you left it, so your hand closes around the handle. Tomorrow someone moves the cup a few centimeters. Your fingers miss, you look down, and you adjust.
That small correction illustrates a question that has fascinated neuroscientists:
- "How much of perception and action depends on expectations, and how does the brain use incoming information to revise them?"
The example is illustrative, not an experiment or proof of a particular theory.
Here we look at one proposal about how the brain might combine expectations with sensory evidence. It offers a productive comparison with generative AI, provided we do not mistake a shared idea for a shared mechanism.
A theory of perception, not a settled circuit diagram
Our senses do not deliver a finished description of the world. Light reaching the eye is compatible with many possible causes:
- A shadow, a dark surface, or an object blocking the light
Prior experience and context can help us interpret what we see, especially when the evidence is ambiguous.
Predictive coding is a more specific proposal about how such interpretation might work. In this family of models, different levels of a system generate expectations about activity at lower levels. Differences between expected and observed signals, called prediction errors, help revise the current interpretation. Some formulations also allow longer-term learning. The proposed comparisons occur within processing networks, not simply inside the eyes or ears.
In their 1999 visual processing model, Rajesh Rao and Dana Ballard showed how feedback predictions and upward error signals in a simulated network could account for some response patterns associated with visual cortex. This was a model of a possible mechanism, not a demonstration that the whole brain works that way.
The distinction matters. In their 2024 review, Rowan Hodson, Marishka Mehta, and Ryan Smith described the empirical support for predictive coding as modest. Some findings are also consistent with alternative explanations. Evidence that expectations affect perception is not, by itself, proof of one particular neural algorithm.
The coffee cup example can help us imagine the proposal. You expect the handle at its usual position, but your fingers miss it. Looking down gives visual evidence that it has moved. In a predictive-coding account, the mismatch between expected and incoming signals helps update your estimate of the handle's position, and you adjust your reach. This illustration does not establish which neurons sent which messages. Nor does predictive coding imply that the brain processes only errors or ignores all correctly predicted sensory information. Sensory signals, ongoing activity, attention, and feedback all matter in the broader problem of perception.
For a business reader, this is a useful discipline in its own right. A plausible account of why a system works is not the same as evidence that it works that way. Test the specific mechanism before designing a product around the story.
The comparison with language models has limits
A language model also makes predictions, but of a different kind. Consider the unfinished sentence:
- "Please send the invoice by Friday so we can process the..."
A model estimates which token, a piece of text, might come next given the preceding context. It can repeat that operation to produce an answer.
During a common form of pretraining, the model predicts the next token and compares its prediction with the token in the training text. A method called backpropagation helps adjust the model's numerical settings, or weights, to reduce prediction error. Because the text supplies the target, this is often called self-supervised learning. The GPT-3 paper by Tom Brown and colleagues describes an autoregressive language model and its use with examples in the prompt.
The family resemblance is real but broad:
- Both descriptions involve predictions and information that can change subsequent behavior
Beyond that, the comparison becomes less certain. A theory about neural processing of sensory signals is not the same thing as an engineered training objective for text. Biological learning involves many processes and timescales. It has not been shown that the brain runs the same algorithm, objective, or backpropagation procedure as a language model. Conversely, not every generative AI system is trained only on next-token prediction.
The practical consequence is that a fluent answer is not a certificate of accuracy. A system trained to produce plausible continuations can still be useful for drafting, classification, or code assistance, but each use needs evaluation against the actual task. A supplier's claim that a model "learns like the brain" tells you little about whether it will identify the right invoice or follow your approval rules.
Perception, hallucination, and the danger of a convenient metaphor
Neuroscientist Anil Seth describes perception as a "controlled hallucination":
- In his account, experience is actively generated and calibrated by sensory signals
It is a vivid description of a proposed explanation for perception, not a claim that the external world is imaginary. The cup is still on the desk, even when your first guess about its position is wrong.
The distinction echoes Immanuel Kant's distinction between the world as experienced and things independently of our experience, though the parallel proves neither philosophy nor a theory of consciousness.
When an AI model "hallucinates”, the word normally refers to an output that is fabricated or unsupported, such as a citation to a study that does not exist. That usage does not establish that the model has perception, subjective experience, or a human-like hallucination. Training a model to predict words or video frames does not by itself demonstrate that it has a conscious inner world. Whether any particular AI system could have subjective experience remains a separate, unresolved question.
Nor should a fabricated business answer be excused as the same useful feature that supports human perception. Creative generation and factual reliability are different requirements. A fictional story can invent a company; a compliance assistant must not invent a regulation and present it as fact.
A practical response is to provide relevant source material and require checks against it. Patrick Lewis and colleagues' 2020 research on retrieval-augmented generation combined a language model with retrieved documents and reported more factual language than a baseline on the tasks they studied.
Retrieval is not a guarantee:
- Documents can be stale, irrelevant, or misread
For a policy assistant, a safer design retrieves the current approved policy, shows the source and effective date, and allows the user to escalate when the answer cannot be verified.
Adapting is not one thing
The original temptation is to contrast a brain that learns continuously with an AI model frozen forever. Both sides need qualification.
People can change their immediate interpretation of a situation, learn a new skill, and form longer-lasting memories. These are not necessarily the same process, and revising an interpretation does not prove that synaptic weights are rewritten every second because of prediction errors. Biological learning operates across multiple timescales.
AI systems can adapt in several distinct ways too:
- Context:
- Give an assistant the latest refund policy in its prompt. Its answer can change during that interaction without changing its weights. Brown and colleagues demonstrated this kind of in-context use with GPT-3, without gradient updates during the task.
- External information:
- Update a document repository or retrieval index. Future answers can use new material while the underlying model weights remain unchanged.
- Parameter updates:
- Fine-tuning or other continued training changes weights. Some research systems investigate continual or online updates; these are design choices, not universal features of deployed models.
For the refund assistant, first check that retrieval finds the applicable approved policy, instructions tell the assistant how to use it, and evaluation cases test whether it cites and applies the right version. A new policy does not automatically require retraining. If it consistently mishandles a specialized task after those checks and improvements, a controlled fine-tuning experiment may be worth considering.
Changing weights has costs and risks. A survey by Haizhou Shi and colleagues discusses continual learning and the problem of performance deteriorating on earlier knowledge or tasks, often called catastrophic forgetting. Unchecked updates can also incorporate bad or unauthorized information. Before deployment, define who approves training data, test old as well as new tasks, record model versions, and keep a rollback path. Automatic weight updates are not a prerequisite for useful adaptation or a proven route to autonomy.
Build a feedback loop you can inspect
Imagine a customer asks whether an order qualifies for a refund. In this fictional example, the policy says eligibility is governed by the purchase date. The assistant retrieves a policy that was not effective when this order was purchased and drafts an answer using it. A reviewer catches the mismatch. The team corrects the document's effective-date metadata and adds the case to an evaluation set. The next test checks both the new case and older cases that previously worked. A real workflow must follow its own policy's rules for which date determines eligibility.
This loop borrows the useful idea of comparing expectations with evidence without pretending to replicate a brain. It also assigns responsibility: someone owns the policy, someone investigates the error, and the team can see whether the change helped. Human feedback is only useful if it reaches a place where it can be checked and acted on.
Prediction is a productive lens for thinking about brains and AI. It is not a license to call their mechanisms identical. For business decisions, the better question is which feedback a particular system can access, how it changes after feedback, and who verifies the result.
Previous
Article 1: The world model gap
Next
Article 3: Why loading a dishwasher is still a hard problem for robots