This is no different from reviewing code from actual humans: someone could have written great looking code with excellent test coverage and still have missed a crucial edge case or obvious requirement. In the case of humans, there's obvious limits and approaches to scaling up. With LLMs, who knows where they will go in the next couple of years.
It is, because a human would have used "thinking" to create this piece of code. There could be error and mistakes but at least you know there there is a human logic behind and you just have to check for things that can easy mistakes for a human.
With AI, in the current state at least, there is no global logic involved with the whole thing that was created. It is a random set of probabilities that generated a somehow valid code. But there is not a global "view" about it that it makes sense.
So when reviewing, you will basically have to do in your head the same mental process as would have done an original human contributor to check that the whole thing makes sense in the first place.
Worst than that, when reviewing such change, you should imagine that the AI probably generated a few invalid versions on the code and randomly iterated until there is something that passed the "linter" definition of valid code.
