LLMs fail at causal accuracy. It's a fundamental problem with how they work.