Predictive Coding | Does AI Perceive the World by Making Predictions Like the Human Brain?

Predictive coding is a theory in neuroscience that posits the brain is not a passive organ that simply reacts to sensory information from the outside world. Instead, it is an active prediction engine, constantly generating and updating a model of the environment. At every moment, higher-level cortical areas generate predictions about the causes of sensory input. These predictions are sent down to lower-level sensory areas. For example, your brain predicts the feeling of your chair against your back or the sound of your keyboard as you type. This top-down predictive model is what constitutes our perception of reality. This process is remarkably efficient; rather than processing the entire firehose of sensory data from our eyes, ears, and skin every second, the brain primarily processes the differences between what it expected and what actually happened. This focus on novelty and surprise allows the brain to operate with incredible speed and energy efficiency. Perception, therefore, is not a direct reflection of the external world but a controlled hallucination or simulation, fine-tuned by sensory reality.

Learning and perception are driven by the concept of 'prediction error.' A prediction error occurs when there is a mismatch between the brain's top-down prediction and the actual bottom-up sensory information. For instance, if you expect a step to be there but it is not, the surprising lack of sensory feedback generates a strong prediction error. This error signal is then propagated up the cortical hierarchy. It serves as a crucial feedback mechanism, compelling the brain to update and refine its internal models to make better predictions in the future. It is this continuous cycle of predicting, checking for errors, and updating that allows us to adapt to new environments, learn new skills, and navigate a complex and ever-changing world. Essentially, the brain is only interested in information that violates its expectations, as this is the information that is most valuable for learning and survival.

Many modern artificial intelligence (AI) systems, particularly in machine learning, are built on a similar principle of prediction. For example, large language models like GPT are trained to predict the next word in a sentence. Generative models in computer vision are trained to generate images that are statistically similar to a dataset of real images. In these systems, the model makes a prediction, and this prediction is compared against the actual data. The difference—the error—is calculated and used to adjust the model's internal parameters through a process called backpropagation. This allows the AI to "learn" the patterns in the data and make increasingly accurate predictions over time.

Despite the functional similarity, the underlying mechanisms are fundamentally different. The brain's predictions are biologically grounded, shaped by evolution for survival, and are deeply integrated with emotion, motivation, and bodily sensations (interoception). An AI's predictions are purely statistical, derived from patterns in its training data without any genuine understanding, consciousness, or subjective experience. Furthermore, the brain's predictive processing is massively parallel and remarkably energy-efficient, while current AI models require vast amounts of computational power and data, making them far less efficient. The brain's predictions are about maintaining physiological stability (allostasis), whereas an AI's predictions are about minimizing a mathematical error function.

The predictive coding framework provides a powerful lens through which to understand various mental health conditions. For example, anxiety disorders can be conceptualized as the brain being stuck in a state of making overly strong predictions of threat or negative outcomes, even in safe environments. Chronic pain might persist because the brain continues to predict the presence of pain even after an injury has healed. In psychosis, hallucinations may arise when the brain's top-down predictions are so strong that they override contradictory sensory evidence, causing a person to perceive something that is not there. In this view, mental illnesses are not just chemical imbalances but can also be seen as dysfunctions in the brain's predictive machinery, where the internal model of the world has become maladaptive.