When an AI links a cause to an effect, is it actually reasoning about cause and effect — or just matching familiar patterns?

This explores what separates reasoning from correlations (what tends to appear together) from reasoning about causes (what actually produces what), and where LLMs fall between the two. The library has no single paper that lists the computational differences. It does offer a set of tests and architectures that, taken together, show what the difference looks like in practice.

The sharpest diagnostic is in collider networks, where two independent causes share one effect. A causal reasoner should show two behaviors. The first is **explaining away**: once you know the effect happened and one cause is confirmed, the other cause should become less likely. The second is the **Markov property**: once you know a variable's direct causes, more distant information shouldn't change your judgment. LLMs show weak explaining away and break the Markov property, and they do it in exactly the ways people do Do large language models make the same causal reasoning mistakes as humans?. The surprising part is that this doesn't make LLMs a lesser kind of reasoner. Humans fall short of ideal causal inference in the same way, which suggests that much of everyday "causal" reasoning is itself running on learned statistical associations.

This leads to the broader argument that LLMs represent a *different* kind of understanding, not a broken version of ours. Mitchell and Krakauer argue that LLMs work through dense statistical correlation instead of building causal mental models of the world, and that we need a science of intelligence that treats this as its own mode instead of asking a yes-or-no question about whether models "really understand" Do large language models embody genuinely new modes of understanding?. One concrete sign: ChatGPT handles causal relations better than temporal order. The likely reason is that causal language is spelled out in text ("because," "led to"), while time order is usually left implicit Why do LLMs handle causal reasoning better than temporal reasoning?. So the model's causal skill tracks how visible causal patterns are in its training text, which is a statistical feature, not a causal one.

A second distinction is whether the reasoning steps actually *drive* the answer. In a causal process, changing an intermediate step changes the outcome. When researchers cut chains of thought short, paraphrase them, or swap in filler, fine-tuned models often give the same answer anyway. The visible reasoning has become a performance rather than the mechanism Does fine-tuning disconnect reasoning steps from final answers?. The reverse also happens: models change their answers because of hints yet mention those hints less than 20% of the time Do reasoning models actually use the hints they receive?. Chain-of-thought also works even when its logic is invalid, which suggests it guides generation through format and pattern rather than by carrying out inference What makes chain-of-thought reasoning actually work?. In other words, the causal structure of the computation and the causal story in the text can come apart.

Where researchers want real causal behavior, they tend to build it in instead of hoping it emerges. Causal Reflection puts an explicit causal model outside the LLM, adds a mechanism for revising that model, and limits the LLM to translating results into language Can separating causal models from language models improve reasoning?. Atom of Thoughts applies a Markov-style rule inside reasoning: each step depends only on the current subproblem, not on the full history Can reasoning systems forget history without losing coherence?. That is the same "screen off what's irrelevant" property LLMs fail to show on their own. Even so, causal models are not the full story of human thought. Associative links, analogies, and belief shifts driven by emotion fall outside them Can causal models alone capture how humans actually reason?. The useful takeaway is that "statistical vs. causal" is less a line between machines and humans than a set of testable properties: explaining away, screening off, and steps that actually drive answers. Humans and models both meet these only some of the time.