The difference is that the positional sacrifice is less tangible. A space advantage, a tempo advantage, more mobile pieces, improved cohesion/coordination of pieces (Kasparov was legendary for taking this last kind of advantage and turning it into a lethal attack). It's a dynamic advantage rather than a static/permanent advantage, which also means there's a risk of that advantage dissipating as the game drags on.

These advantages aren't the kind where you can sit back and let the game play out confident of winning. It's a deliberate unbalancing of the equilibrium of the position, and one where this temporary dynamic advantage needs to be used to create a longer-lasting and static advantage.

Would it be fair to say you are trying to optimize for future positions where you aren't sure you will win, but the positions resemble certain archetypal positions/ share certain features that are advantageous (i.e. has a high probability of transforming into conventionally advantageous situations)?

I'm sure the chess AIs are full of this sort of knowledge internally, though, in the form of computation optimization algorithms. Perhaps the issue is to translate it to a human-usable format.

Indeed, chess engines do have heuristics to include positional advantage in their evaluation of a board, so they "know" in some way that a doubled pawn is disadvantageous or that development of pieces or attacking central squares is beneficial, much as humans know these things.

I've never heard experts discuss this, but I bet it's true that human beings still succeed in appreciating many of these benefits at a higher level of abstraction than machines do. An argument for this is that computers needed an extremely large advantage in explicit search depth to be able to beat human grandmasters. So the humans had other kinds of advantages going for them and most likely still do. One of those advantages that seems plausible is more sophisticated evaluation of why a position is strong or weak, without explicit game tree searches.

I looked at the Stockfish code very briefly during TCEC and it looks like a number of the evaluation heuristics that are not based on material (captures) are manually coded based on human reasoning about chess positions. But if I understood correctly, they are also running machine learning with huge numbers of simulated games in order to empirically reweight these heuristics, so if a particular heuristic turns out to help win games, it can be assessed as more valid/higher priority.

You could imagine that there are some things that human players know tacitly or explicitly that Stockfish or other engines still have no representation of at all, and they might contribute quite a bit to the humans' strength.