It's inherent in the nature of the trigonometry, coupled with any slight inaccuracy in the calibration.
Proper calibration requires you know exactly what values it would give under zero gees of acceleration-- hard to manage on Earth's surface where the sum force of all three axes will be around 1.0 gees. A calibration routine could collect data at +1.0 gees and -1.0 gees, and assume a linear relationship. I have a routine in the works for this, but I doubt any game player would bother-- that's why the Wii never bothers.
A workaround for the trigonometry would be to figure out that the pitch was more than 45 degrees, and then start computing the pitch using the inverse formula. I haven't started that technique but it should work much smoother, say to 85 degrees. There is a limit, though. Even measuring three axes, you will run into "gimbaling," where Euler angles cannot express a coherent position near the north and south poles of the rotation space.