It would seem that detachment of self is not necessarily a bad thing for survival, but there may be other situations where the "autopilot" mode is fatal. The fatalities don't get interviewed, so you'd have to look for near-fatalities who were rescued and undertake a difficult statistical analysis. (I agree there are evolutionary pressures for the type you describe.)
I'm going to call that a design flaw!
In that situation, one might see how survival required depersonalization. The decision to abandon the others would likely be incongruent with most people's sense of self, as you recognize. I share your skepticism of this hypothesis. There are a few rickety steps there. First of all, there are obvious survival advantages in modeling other primate's behaviors. But why model our own? The hypothesis must be that this type of modeling is an inevitable spandrel of the helpful type. The other problem is that even if we begin modeling our own behavior, an intelligence that models everyone's behavior would not seem to have any particular way to single out one particular individual as being special just because the data is better-- why is that modeling any different from the others? (But maybe people who live so intimately with another for so long that they can model the other person as easily as themselves do in some sense come to feel a unity of self with that other person.)

Still, I think I see a more natural survival benefit in having a sense of self that is not a spandrel of modeling-- it motivates the survival urge. Sure all creatures seem to have some kind of survival instinct even without a sense of "self", but it doesn't work in concert with intelligence. Fear is a motivator too, but sometimes survival requires doing exactly what is feared. If we want to draw from the survival advantages of our intelligence, we need to convince the intelligence that it is worth saving. That reminds me of the end of the movie "Dark Star", where the crew tries to convince the artificial intelligence of the ship not to self destruct. I forget how they tried to do it exactly, but it might have involved giving it a sense of self... (I don't believe it worked).
I agree, but don't see how the modeling concept comes directly into play. I think your argument is the same if one substitutes "self-building" where you have "modeling". Furthermore, I find it more plausible that we build a sense of other selves after we have built a sense of our own self, just as we imagine other consciousnesses only after we discover our own. Certainly our own sense of self is more refined, but one might argue that it's because the data is so much better, I just don't know.


Yes, this could be the survival tradeoff right there, as situations involving practiced agents would be more likely. The squirrels in my back yard run away from my dogs several times every day, and most live to reproduce, surprisingly.