AI language models often make the exact same reasoning mistakes humans do — does that mean they understand cause and effect the way we do?
Can statistical correlations in LLMs function equivalently to human causal understanding?
This explores whether the pattern-matching that LLMs learn from text can do the same work as the cause-and-effect models humans use to understand the world, or whether it's a different kind of thing that only looks the same from outside.
This explores whether the statistical patterns LLMs pick up from text can do the same job as human causal understanding, or whether they only look similar from the outside. The corpus doesn't give a clean yes or no. It suggests the question rests on two assumptions that don't hold up well: that human causal reasoning is a clean, rule-following process, and that correlation-learning is one uniform thing.
Start with the surprise. When LLMs reason about causes, they make the *same mistakes humans do*: they fail to 'explain away' one cause when another is confirmed, and they break the same probabilistic rules people break Do large language models make the same causal reasoning mistakes as humans?. If LLMs were simply worse at causal reasoning, you'd expect different errors. Matching errors point to a shared source. Human causal intuitions may themselves be partly built from statistical regularities in experience and language, which is the same material LLMs train on. The human side isn't purely causal either. Formal causal models of human thinking leave out association, analogy, and emotion-driven belief change, all of which shape how people actually reason Can causal models alone capture how humans actually reason?. A related finding: LLMs fine-tuned on psychology experiment data predict human decisions better than theory-driven cognitive models do Can language models learn to model human decision making?.
The similarity has limits, though. LLMs handle causal relations better than temporal ones because causal words like 'because' and 'therefore' are frequent and explicit in text, while the order of events is usually left implicit Why do LLMs handle causal reasoning better than temporal reasoning?. So the 'causal understanding' partly reflects how people *talk about* causes, not how causes work. Simulation work shows a similar pattern: LLMs guided by causal models reliably get the *direction* of social effects right but not their *size* Can structural causal models automate social science with language models?. That fits a system that has absorbed the shape of causal talk without measuring the world.
One response is to stop asking for equivalence. Mitchell and Krakauer argue LLMs are a distinct kind of intelligence that works through correlation rather than causal mental models, and that 'does it truly understand?' is the wrong binary Do large language models embody genuinely new modes of understanding?. Interpretability research makes this concrete. Inside a model, compact principle-like circuits sit next to shallow heuristics instead of replacing them, so you get a patchwork, not a single mode Do language models understand in fundamentally different ways?. That patchwork can also be a strength. BrainBench found that the same tendency to blend patterns that causes hallucination when recalling facts lets LLMs beat neuroscience experts at predicting which experimental results actually happened Can LLMs predict novel scientific results better than experts?.
The practical upshot: when causal reliability matters, some researchers don't ask the LLM to be causal at all. They put the causal reasoning in a separate formal model and use the LLM only to translate between that model and language Can separating causal models from language models improve reasoning?. And here's the twist you might not expect. Researchers studying *LLMs* face the same correlation-versus-causation problem. Finding a pattern in a model's internals proves nothing until you intervene and show it actually drives behavior Can LLM understanding rely on just representation or causation alone?. The same test applies to model self-reports. They usually echo training data, but they become real introspection when a causal link connects the internal state to the report Can language models actually introspect about their own states?. Correlation becomes something more when it is causally connected to what it describes.
Sources 11 notes
LLMs show weak explaining away and Markov violations in collider networks, matching human error patterns exactly. This suggests shared mechanisms rooted in training data statistics rather than categorical reasoning inferiority.
Causal belief networks excel at modeling causal reasoning but cannot represent associative links, analogical mappings, or emotion-driven belief shifts. The GenMinds framework itself acknowledges this as a tractable starting point rather than a complete theory.
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
ChatGPT excels at causal relations but struggles with temporal ordering because causal connectives are explicit and frequent in training data, while temporal order is often implicit and must be inferred contextually.
LLMs guided by structural causal models can propose and test causal hypotheses across negotiation, bail, interview, and auction scenarios. Simulations reveal effect directions reliably but not magnitudes, making them useful for directional social science.
Show all 11 sources
Mitchell and Krakauer argue that LLMs represent a distinct species of intelligence that operates through statistical correlations rather than causal mental models. They propose this warrants an expanded science of intelligence that moves beyond the binary debate of whether LLMs truly understand.
Mechanistic interpretability reveals conceptual understanding (features as directions), state-of-world understanding (factual connections), and principled understanding (compact circuits). Crucially, higher tiers coexist with lower-tier heuristics rather than replacing them, creating a patchwork of capabilities.
BrainBench benchmarks show fine-tuned LLMs outperform neuroscience experts at predicting which experimental results actually occurred. The same pattern-integration tendency that causes hallucination in retrieval tasks enables genuine prediction in forward-looking scenarios.
Causal Reflection separates causal reasoning into a formal dynamic model with a Reflect mechanism for revision, relegating the LLM to structured inference and language rendering. This architecture sidesteps asking LLMs to perform causal reasoning directly, addressing both spurious-correlation failures and RL's explanation gap.
Research shows that representational analysis alone identifies correlates without proving causation, while causal analysis alone demonstrates effects without explaining function. Only paired methodology—locating candidates representationally then verifying causally—produces genuine mechanistic understanding rather than descriptive claims.
LLM self-reports usually reflect human training distributions rather than actual internal processes. However, when a causal chain connects an internal state to accurate reporting—like inferring low temperature from output consistency—genuine lightweight introspection occurs without requiring consciousness.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Do Large Language Models Reason Causally Like Us? Even Better?
- Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning
- The Debate Over Understanding in AI's Large Language Models
- Causal Reflection with Language Models
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- What Do Large Language Models Know? Tacit Knowledge as a Potential Causal-Explanatory Structure
- Mitigating Hallucinations in Large Language Models via Causal Reasoning
- Mechanistic Indicators of Understanding in Large Language Models