[wizzwizz4]

It's not a person. When it "says" something confidently, that just means it's predicted a string that, interpreted by humans, appears confident. It's completely unrelated to whether the output values of the model are "confident" (highly contrasting).

A predictive text system, driven non-deterministically, can "confidently" say one thing some of the time, and "confidently" say a completely different thing. It probably will, if it "knows" that the expected output should be something said confidently.