Generally, every model is wrong about something. That the core difference between models and the things we model. Every description is an approximation of some sort.

However, some things are tautological. Some things may be necessary in reality, even if they are not in any model. (I wake up each morning because I have no alternative, not because it is the 'right thing to do'.) Some things are very safe to assume because our human cultural context or our biological environment is stable over multiple generations of activity. And some things are true simply because assuming otherwise contradicts the assumptions needed for language to work effectively. And sometimes we do prove things, even if we lack any meaningful ability to prove what we'd like to the degree that we'd like.

The point is, we're not always wrong in a linguistic sense. Any time you have a statement that is perfectly symmetrical and valid in every context, you do not have language which can be used to make decisions. It's only meaningful to be wrong if there are things you can be right about.

How do you improve your knowledge if every path forward is still wrong?

(In ML terms, how do you apply gradient descent if every path forward is up?)

Not OP, but I think the idea is to consciously pause before taking the path that is obviously (to you) downhill. This gives you a chance to think whether it's actually downhill instead of making an assumption. It's a way to practice mindfulness

You start with a prior, but you assume that it is not the optimal solution.