It helps catch the most obvious bugs. Like the algorithmic bugs. If the comment string says something and the code does not do that, that's caught fine. But it is not so good at catching domain related bugs. I think that's to be expected because these AI-tools don't have the knowledge of the domain and the hidden assumptions that come from experience in a domain. It could improve in future if the AI-based tools are trained on a lot of design documents + code that solve domain problems.