Then we'll have the problem of us humans trying to predict the behaviour of the machine that unintuitively will be harder than predicting that of other humans. Of course the system will be much safer when there are no more humans involved in the decisions.