Some people’s job description involves things that we know are mostly computationally feasible: efficient resource allocation, task routing, and the quantitative control of complex symbolic systems. Other people’s jobs are orders of magnitude more difficult from a computational point of view, involving things like the control of physical objects and dealing with poorly structured natural language information and requests. Yet commercial applications of artificial intelligence are overwhelmingly focused on the second set of tasks.

A naive person would assume that it’s because the second set of tasks commands much higher wages, and hence have higher returns to automation. But the opposite is the case; people who allocate resources, route tasks, and so on, that is, people in the first group, are actually much, much better paid, both individually and as a class.

A paranoid person, a game theorist, or an economist studying organizations (assuming there’s much that differentiates those three disciplines) would offer an alternative hypothesis: software hasn’t made any inroads replacing people in the first group because people in the first group are the ones who decide what software gets made and bought.

Call it psychology, economics, or evolutionary biology, a high-level manager will eagerly slash middle management positions – making the same number of people do more work is one of the basics of his or her job description – but replacing middle managers with software would raise uncomfortable questions as to whether management in general can be done by software... and therein, of course, lies madness, anarchy, complete economic ruin, horses eating each other, and so on.

I guess now I should proceed to pad this column up with examples and witticisms, because if the value of what I’m doing resided mainly on suggesting an hypothesis based on publicly available information, then it could eventually be done by software, and that of course is patently impossible.

Trust me. Why would I lie about this?