@tewha@appdot.net I saw David Chisnall post something about this that I think fits very well: in general ML is useful when the cost of mistakes is low and the value of correct answers is high. For example using ML for branch prediction or for indexing into database tables on disk: if the ML outputs a correct answer the computer runs a little faster and if it outputs wrong answer the computer runs barely slower than if it hasn't tried guessing. I think code review fits that framework: failures to find problems don't matter much because it puts you in no worse a state than if you hadn't tried. False reports of problems only waste your time. But accurate reports of problems may save big headaches down the road.