Doesn't really work that way in that it still reduces to an overall system that doesn't understand what it's doing (someone else mentioned GANs, they are a grea

Doesn't really work that way in that it still reduces to an overall system that doesn't understand what it's doing (someone else mentioned GANs, they are a great example of how unwittingly incorrect ML model output can be - go look at the out-takes from any image generation model) A better approach is good uncertainty quantification (this is not at all trivial) to quantify when the model is and isn't qualified to be making a prediction.