>>12919560>>12919772AI works well because sometimes it is extremely wrong. I'll try to explain.
You're "training" AI on a dataset that represents your problem. If it has zero error on this set, you have run into overfitting. This basically means your AI model just learned the dataset by heart and won't generalize at all.
For generalization, there has to be some error left. This fact will cause an AI to do wildly inaccurate predictions for some situations, most of which would not have happened with a human expert. They're also impossible to debug.
Now, you may argue that although these catastrophic errors happen they're MUCH less common than human error and that's correct. The problem is accountability. If a human expert makes a mistake, he takes responsibility for it and some insurance company will likely cover his/her ass. In the case of software, who's responsible?
Truth is, AI is already widely used in healthcare, but at least in diagnostics and treatment only together with a human expert (it's used without human intervention in insurance companies, data acquisition and processing). This expert takes the final responsibility.
