That's bull. You can't recover data that isn't there.
Its not fully Bull because not all the data is lost, it is just convoluted.
As lord filbuster said, imagine shooting some stars with a long exposure and you get star trails, it is somewhat trivial to remove the relative motion and return to points of stars, which is basically what this software does.
Another analogy, if you are using film an you take 1 exposure of a scene and without winding on the film you taken a second identical exposure. The data information for both photos are stored on the single frame and since one knows the exposure times were the same then simply diving the pixel values by 2 (e.g. if the frame was scanned) will get you back to the original exposure without many detrimental effects (except perhaps overexposure of highlights). Same kind of thing happens with motion blue, the image data gets recorded to the sensor but the data gets superimposed and convolved. If you know the motion trajectory function, known as the point spread function (PSF), which you can retrieve through a Fourier Transform, then you can apply the reverse function, a deconvolution.
Here is an example to show you what is possible, remember this case?
http://thelede.blogs.nytimes.com/2007/10/08/interpol-untwirls-a-suspected-pedophile/?hp
The warped face has massive distortion and one would think it would be impossible to retrieve the original image, interpol just played with photoshop to find the correct parameters and applied the twist in the opposite direction. You can try this for yourself.
Of course, it wont be anywhere near as good as getting it right in the first place, and it will only have a limited scope for improvement as well. A bit like IS/VR correcting only 2 to 3 stops worth of movement.
