The continuous execution features consolidated in the period pose a practical question for anyone leading a team: code review was designed assuming traceable human authorship.
When part of the work happens in the background, with the system trying, measuring and correcting on its own, the change arriving for review is the result of a process nobody watched.
That doesn't make review useless, it makes it different. The question stops being only whether the code is correct and starts including whether whoever submitted it can explain the decisions taken along the way.
In practice, the teams that handled this best adopted simple measures: limiting the size of submitted changes, requiring a description of what was tested, and reviewing error handling and edge cases with extra care.
Those are precisely the points where generated code tends to be generic, because the model solves the main path easily and handles exceptions with the most common pattern it knows.