Does it need to solve original problems? Most of the code we write is dealing with the same problems in a slightly different context each time.

As others say in commends it might be the case where we meet in the middle. Us writing some form of tests for AI-produced code to pass.