Raymond's Weblog^

No free lunch for ethics

It turns out that if you don’t know anything about the outside world, it is impossible to make moral choices in a consequentialist framework in a way that outperforms randomness. This is even if you are endowed with the ability to perfectly discern all the consequences of a given action in some fixed time.

In brief: while divining the fundamental nature of ‘good’ might be a purely intellectual exercise, evaluating a decision theory for how to do good in the world is in fact a search problem, and therefore subject to the No Free Lunch Theorem.

What this theorem means in this context is that, across the set of all possible worlds (which, overall, have no structure), all decision theories based on considering an action and checking what its results are will perform equally well, including random search. This is regardless of whether your goal is to find ‘the best action’ or just one above a threshold.

In slightly more detail

There have been some number of attempts to divine morality from pure intuition, most notably by Kant. Descartes also sort of did this. But as with many moral theorists the main concern was ‘what is good’, and not the necessary followup question of ‘how should you decide what to do, such that you do good’.

I claim that this second question demonstrably cannot be solved with pure intuition, assuming that the goodness of your action is a function of its consequences, even if you can perfectly discern the consequences of your actions.

The reason is that pure intuition doesn’t give you any information about the structure of the world, or the manner in which it is structured. In the absence of this, there’s no reasonable notion of two actions being ‘similar’, so the consequences of one action provide no information about the consequences of any other actions, other than that they’ll all be different from it.

A decision theory will, at best, provide you with a protocol for checking actions, but any such protocol will be implicitly assuming a given structure for the universe, and therefore will do better if the universe is structured in this way and worse if it isn’t.


This is basically of no relevance to any practical choices. The universe is in fact structured, reasonably parsimonious, and brimming with reasonable opportunities to use inductive reasoning. But we can conclude that any attempt to reason about what you should do without reference to external experience is doomed to be exactly as effective as random action.