Probably the best solution would be a combination of the two: the ML algorithm would also learn from human examples and corrections. I think this will be the s

Probably the best solution would be a combination of the two: the ML algorithm would also learn from human examples and corrections. I think this will be the solution in a lot of problems and industries: ML does the boring and repetitive work, and a human sanity-checks it and fixes the edge cases where ML is 'dumb'