There's only so much maneuvering space, and so much time a single brushing session takes; perhaps the error of double integration is small enough they can get away with dead reckoning. Especially since you mouth is a small, confined space, so they could estimate its bounds and use that to further correct their position estimate.

(I imagine a system like this would need at least two accelerometers - one close to the tip, and another one in the grip area; the second one would be used to cancel out the movement of users' head and body, allowing the first one to effectively operate in a mouth-relative coordinate system.)

It's probably not going to be accurate to a single tooth, but to a region.

Once you've moved the brush around a little bit, it can probably make a basic 3d map of you teeth to do the job.

I think it would be quite accurate if you parse based on training data, think about how we brush our teeth and the relatively consistent and unique angles and accelerations.

Maybe not from brushing one tooth but brushing the whole mouth you'd get a pretty good picture I'd say. The app would justify itself if it did give you a rough map of how you brushed your teeth and suggest areas that you're missing/if it's being done incorrectly.