Your players can be categorised in one of two mental models.
Some of your players are like you -- they've associated tilting right/left with the ball spinning clockwise/anticlockwise.
Others have a different mental model: They associate tilting right/left with the ball moving right/left on the screen. It's these players who, when their ball is tracking the ceiling, expect that tilting left will send the ball toward the left on the screen.
As long as the ball is only following the ground, this mistaken mental model is okay -- the two mental models both produce the same, predictable results for any particular set of control inputs. But as soon as you throw a new context into the mix -- the ball following the ceiling -- half the players will become confused, as the game's behaviour suddenly doesn't match their mental model of what the game should be doing any more.
There are a couple things you can do to handle this situation.
The first thing to try when you find yourself in this sort of situation (half your users are confused because they have an incorrect mental model of how the game works) is always teaching. You want to have a tutorial very early in the game where you explain to the user that tilting left/right causes the ball to spin in that direction, and let them practice with it in a hazard-free scenario until they feel comfortable with it.
The second thing to try is to see if you can't make both mental models work, by avoiding the situation which causes them to come into conflict. In this case, don't have the ball be able to follow the ceiling, or perhaps spin the view around so that whatever surface the ball is following is at the bottom of the screen, so that tilting left does always move the ball to the left on the screen, even if that's technically the 'ceiling'. This has the benefit of not needing a 'tutorial' scenario; it just allows people's assumed controls to work for them in every situation, no matter which mental model they've picked, because we're presenting scenarios in such a way that both mental models produce the same expected behaviour for the same set of control inputs.
If you can't solve this sort of conflict in either of the two previous ways, then in some cases, you can offer a game configuration setting to swap control styles between the two mental models, so that each player can make the game controls work according to their own intuition of how the controls ought to work.
But you probably can't do that last option in this case -- the problematic "tilt left to go left" mental model doesn't work at all if the ball is traveling along a wall (which I assume is possible in your game, since you've mentioned floors and ceilings and loops). In your case, you're probably going to need to do one of the first two options; either teach people the correct mental model, or change your presentation style so that the incongruity between the two possible mental models doesn't present itself.