I assume you meant FSL (using fsl_motion_outliers)?

Not sure how the RMS motion is computed, and this could be an artifact of the particular program you are using to compute the motion. If there were such a thing as a perfect way to measure motion, we would simply use that measure as a goal to optimize. But any measure introduces a bias to its assumptions.

But it is certainly not impossible for DTIPrep (or any motion correction, such as in FSL) to make things worse. Especially if there is no motion present, a registration process (to correct the potential motion) may correct for a larger motion than actually present (due to the high noise, low SNR). 

Martin