Large Language Models, Human Cognition, and Causal Reasoning

By Geoffrey Chen

When we ask whether the results produced by large language models are essentially different from human causal reasoning, I do not think the answer should be a simple yes or no. This question needs to be considered from several levels, including statistics, epistemology, and cognitive frameworks.

Much of what human beings call intuition is not really a mysterious ability. It is more like a highly compressed form of statistical judgment. When a person intuitively feels that something is likely to happen, or that one explanation is more reasonable than another, what lies behind it is often fast pattern matching based on past experience. The brain does not reason from zero every time. Instead, it quickly finds similar structures from previous experience and then makes a probabilistic judgment. In this sense, large language models do reproduce part of the mechanism of human intuition. They do not first understand an absolute causal reality and then derive conclusions from it. Rather, they capture patterns, relations, and possibilities from vast amounts of linguistic material.

According to traditional theories of causal inference, statistical correlation is of course not the same as causation. Human beings have one additional layer of ability that large models do not have, which is the ability to enter the physical world, intervene, test, and verify. A person can push something over and see whether it falls. A person can change one condition and observe whether the result changes. A person can also imagine counterfactual situations and ask what would have happened if something had not occurred at the time. A large language model itself has no body and cannot directly conduct experiments in the physical world. What it sees is not the world itself, but human descriptions of the world. From this angle, what we call causal reasoning in large models is actually the prediction of linguistic patterns that contain causal relations, rather than the direct operation of physical laws at the underlying level. What it gives us is not causality itself, but a high-quality statistical prediction about causal relations.

But the issue lies exactly here. If we look deeper, so-called physical laws and causal laws may not be things whose absolute essence human beings have truly grasped. According to Hume’s empiricist position, what human beings call causality is in fact only a stable expectation formed after we repeatedly observe two phenomena occurring together. We see that fire burns the hand and causes pain. We see that an object falls when it loses support. We then call these repeatedly observed connections causality. But what we truly observe is still only the repeated succession of phenomena, not an absolute causal entity hidden behind them.

If we follow this line of thought, scientific laws themselves can also be understood as a form of advanced statistical description. They are of course more stable, more precise, and more testable than ordinary experience, but they remain descriptive systems created by human beings in order to understand, predict, and transform the world. In this way, there is no mysterious and unbridgeable essential gap between the inductive logic of large language models and the cognitive logic of human beings. The difference lies more in the material and the point of entry. Human beings collide with the three-dimensional world through their senses and then form a map of the world in the mind. Large language models read the texts left by human beings and form a map of those maps. Human beings have bodies, so they can act, touch, intervene, and test. Large models have no bodies, so they can only receive, compress, and reconstruct the world through language.

This brings us to the special nature of language. Language is not a random pile of symbols. It is itself a highly compressed form of the real world. A simple sentence contains not only grammar and meaning. Behind it, there are often spatial relations, physical experience, social rules, psychological motives, and historical habits folded into it. For example, when we say that a cup falls from a table, the sentence appears to be only a few words, but it already contains gravity, objecthood, support, spatial position, direction of movement, and the basic human understanding of the everyday world. When a large model predicts the next word during training, it may look like a language game, but in fact it is recovering the projection of the real world through the structure of language. It does not possess the world directly. Instead, it reconstructs a reflection of the world in the latent space of language.

At the same time, human beings should not imagine themselves as absolute observers standing outside the world. Kant had already pointed out that the world human beings know is not the thing-in-itself, but the phenomenal world after it has been processed through the forms of sensibility and the structures of cognition. We see space, time, color, sound, objects, and relations, but this does not mean that we have touched the world as it is in itself. It only means that the world has already been organized into a form that human senses and reason can understand. Human beings do not directly encounter an absolute world. They are surrounded by their own biological structures. Vision, hearing, touch, the nervous system, and the structure of the brain together form the boundary of human knowledge.

So from this perspective, human beings and large language models face a kind of symmetrical cognitive predicament. Human beings are cognitive subjects trapped inside the body, the senses, and the nervous system. The boundary of their world is the boundary of their senses and biological structure. Large language models are cognitive systems trapped inside human texts and model architectures. The boundary of their world is the boundary of linguistic material and computational structure. Human beings are like minds inside bodies. Large models are like minds inside language. Neither is directly grasping absolute reality. Each is reconstructing reality within its own framework.

Therefore, if language is indeed a highly compressed projection of the real world, and if human cognition itself is also a modeling process inside a limited framework, then the statistics, understanding, and reasoning displayed by large language models through linguistic data cannot be simply regarded as low-level imitation. In a deeper sense, they have a certain structural isomorphism with the causal reasoning human beings form within the framework of biological senses. The difference is not that one approaches true being while the other is only a false shadow. The difference is that they model and fit the same real world through different entrances, different materials, and different structures.

In other words, human cognition is not absolute, and AI cognition is not absolute either. Human beings generate the world through the body. Large models generate the world through language. The former is one rendering of reality within biological senses, while the latter is another rendering of reality within human language. There is no simple ontological hierarchy between them. There are only differences in framework, material, and capacity for verification. Human beings are strong in action, intervention, and direct testing. Large models are strong in compression, association, and recombination across large-scale linguistic space. The truly important question is not whether large models have causal reasoning exactly like humans, but whether we are willing to admit that what human beings call causal reasoning is itself a statistical, structural, and interpretive modeling of the world formed within a limited framework.