When I write code by hand and I feel confident about what I'm doing, the confidence comes from knowing that the extraction is generally already done. But I'll still find myself pausing and hesitating, checking for typos, making sure things correspond to their surrounding context… even after decades. It feels like "thinking" in the moment, but not so much when I reflect back on the nature of the task. AI seemingly is not limited by psychology, so it doesn't have these hangups (and of course it also just "types" far faster). It doesn't get an ego boost from being right the first time, and its "confidence" in the compiler, tools and tests is just not relevant.
In other words, the data shows it is maybe not possible, you think it is, describe why the data is wrong and you are correct.