There was a great debate about this some years ago, after 16bit ops was added to Photoshop. I can't remember what forum it was on. The debate centered on the accusation that some Photoshop tools operated in 8bit mode even when the image was in 16bit mode.
In some examples, people claimed the brush tool and gradient tool created banding associated with 8bit processing. Some claimed that "soft" brushing on the main pixel layer looked quite different than making the same brush stroke on a mask.
One person developed a test which appeared to prove that masks were in 8bit mode even on 16bit documents. I just repeated that test to the best of my memory, and now (under CC 2014) it appears to prove that masks are indeed 16bit, although in grayscale mode (one value per pixel instead of 3 RGB values).
Here's the test:
1. Load any image, dupe the background twice, for a total 3 layers.
2. Save that in 2 versions, one in 8bit, one in 16 bit, as uncompressed TIF.
3. Add a layer mask to the top layer of each of these and save new versions.
Now, with 4 versions of the file, look at file sizes on disk and compare the differences. Look at the difference between the 8bit masked and 8bit unmasked versions. That difference is the size of the mask in 8bit. Then look at the difference between the 16bit masked and unmasked versions. Finally, look at the difference between the differences.
As I recall from years ago (version CS? CS2?) There was no difference between the differences. The size added by the mask in 8bit mode was the same as the size added in 16bit mode, which seemed to prove that all masks were 8bit. When I do this test now, there is a difference. The size added by the mask on the 16bit image is twice the size added to the 8bit image. And those sizes equate approximately to a grayscale size of my test image.
A disclaimer - I don't completely trust my memory of the debate from years ago, so neither should you.