[waterhouse]

In game theory, you encounter situations where a player's optimal strategy is a "mixed strategy", i.e. "x% chance of choosing strategy A, y% chance of choosing strategy B, ...", with at least two nonzero values in x,y,... . For example, in rock-paper-scissors, if your strategy is anything other than 1/3 rock, 1/3 paper, 1/3 scissors, then an opponent can defeat you (in terms of expected value) by choosing the right pure strategy. People don't normally take out a die to make the choice, but... In a sport like boxing, if in certain situations you always "punch high" or "punch low" or otherwise act predictably, that will probably hurt your chances against an observant opponent. That presumably applies to real fights as well. Outside of direct combat, sometimes someone does you a small favor (or a small offense), and maybe the only favor (or punishment) you can do them in return is significantly larger than that; in that case, the appropriate response might be an x% chance of doing the thing and 1-x% of doing nothing.

So nondeterminism is a useful skill in at least some circumstances, quite likely some circumstances that existed in the environment where we did lots of evolving. Therefore I would expect us to have evolved some capability for (apparent) nondeterminism.

How often does the agent deal with lawyers in an adversarial situation? How often does the lawyer deal with people in adversarial situations like that? If an expert is able to see through and manipulate a beginner, well, I don't consider that good evidence that humans can never effectively deploy nondeterminism.