A new causal audit finds that AI models' chain-of-thought text often doesn't drive what the model is actually computing.
Researchers built a metric called the CoT Mediation Index, which measures how much a model's output actually depends on its stated reasoning by patching the hidden states tied to that reasoning and comparing the resulting performance drop to a control patch. Testing multiple model families, including Phi, Qwen, and DialoGPT, across different scales, they found that reasoning's causal influence is usually concentrated in narrow layers rather than spread throughout the network. Some models scored near-zero on the index despite producing chain-of-thought text that looked perfectly plausible, meaning the explanation was cosmetic window dressing. Models specifically tuned for reasoning showed stronger, more structured dependence on their stated reasoning than larger models that hadn't been tuned that way, while separately, Mixture-of-Experts models showed a more distributed pattern consistent with how they route computation across experts.
The finding matters because chain-of-thought output is increasingly treated as a window into how a model reached an answer, including for safety monitoring and debugging. If the visible reasoning can be decoupled from the real computation, reading it tells you less than it appears to, and behavioral benchmarks alone can't catch the gap.
Chain-of-thought was sold as reasoning made visible; this audit is a reminder that fluent explanations and faithful ones are not the same claim.