...might be ok if you don't do very much with your image. the problem is you're going to apply a number of transformations, and then rounding errors accumulate.
take a look at an image histogram after changing the black and white points in colors->curves. that is data loss at work (a particularly graphic case, i admit). why should i put up with gaping holes in my histogram if the original image had 10 or 12 bits of depth?
the discussion is pretty much the same as with cd audio quality. a carefully mastered and dithered cd is capable of totally adequate sound quality. but its underlying format, 16bits at 44k1 sampling rate, is woefully inadequate for studio use. so what everyone does is work at higher bitdepths and possibly higher sampling rate, and only down-convert as a very last step, before the product goes out to the consumer.