My adopted working hypothesis is that we have to simulate what others observe/think in order to predict their behaviour and then adjust our own expectations and plans. It's a lo-fi model, because we don't have access to their internal workings - and yet it's a powerful model because it needs to predict the behaviour of very complex agents. If we then turn that powerful machinery onto ourselves, we can add all the extra information we didn't have of others. So it's a much richer, more detailed description of what's going on.

It's not true that we have to be conscious - we can do things on autopilot too. The same way we can be thoughtless in treating others - we can turn the machinery off.

This autopilot works very well with a car and a very well known road ( like a daily commute).

Recently I was thinking of a very hard problem and I cannot absolutelly recollect approx 30 minutes of the drive home. Apparently my body and some part of the brain took over the driving, while my cortex was busy with the problem reserving the short term memory completly. I snapped out of the situation when apparently a stronger break was needed, and this intrerupted my flow and popped my attention to the road. Fascinating!