It can be trained some form of trustfulness, and a measure of certainty is intrinsically part of the model (compare with image to text models etc "70% cheetah 10% dog")

In any case it can be trained to assert confidence via a side channel