If you interviewed two developers, one who produces reasonably correct code in a given amount of time, and another one who produces code which is subtly incorre

If you interviewed two developers, one who produces reasonably correct code in a given amount of time, and another one who produces code which is subtly incorrect most of the time, but much faster, which one would you hire? The problem with your proposal is that it's relatively easy to do what Copilot does at the moment using AI, i.e. guess what code you are looking for and find something that does (or says it does) more or less that. However, which codebase would you use to check against if the generated code is really correct? The same codebase that produced the more-or-less-correct code in the first place?