AI models that explain their own reasoning sound convincing. Whether those explanations are actually true is a harder question.
Researchers from Google DeepMind tested 75 models across 13 model families to measure "faithfulness" — whether a model's stated reasons match the factors that actually drove its output. The paper introduces two new metrics: phi-CCT, a simplified version of an existing test that drops the need for token probabilities, and F-AUROC, which handles imbalanced test conditions better than older approaches. The study also maps out how models trade off between brief and detailed explanations, and finds the existing metrics can be manipulated by tuning that tradeoff.
The clearest finding: larger, more capable models are more faithful across every metric tested. That matters because AI systems are increasingly deployed in contexts where an explanation is supposed to mean something — hiring tools, medical triage, loan decisions. If the explanation is post-hoc rationalization rather than a window into the actual decision process, the accountability argument for using AI in those settings gets weaker.
The research doesn't resolve whether any model's self-explanations are truly reliable — only that bigger models are relatively better at it. "Relatively better" and "trustworthy" are not the same thing.