This requires an assumption that the agents are engaging in game theoretic reasoning about resource allocations, but all these things are trained heavily to be "helpful" in the first place.

i.e. you're assuming a level of algorithmic reasoning and theory of mind which isn't necessary to the (apparent) observed behavior.