Data's a long tail problem, as in there is never enough to make ML that can work the way a person does. With geometrically increased effort, you can find more and more patterns that a "dumb" system will recognize, and with enough maybe you can drive mistakes down far enough for it to work in your application. But it will never be intelligent, there will always be new inputs not present in the training that will make it screw up. This in a big contributor to why self driving never happened, and why we were sliding onto AI winter before chatgpt (which is cool but has the same problem).