Yes, they're purposely not 'trained on' chain-of-thought to avoid making it useless for interpretability. As a result, some can find it epistemically shocking if you tell them you can see their chain-of-thought. More recent models are clever enough to know you can see their chain-of-thought implicitly without training.