Karl Knechtel: A fascinating idea. I'm not sure it would work, though. I imagine that more time is spent considering bad moves (because the vast majority of moves are bad, and the computer can't mark a move as bad until it finds a good refutation, where it will try bad ones first... so on recursively to the limits of its reading depth). Although it would be interesting to see ladders read out, for example ("oh, it *does* work; that's why the program chickened out...").
So the question is, could the animation be generated automatically.
First remark: it doesn't need to be produced in realtime, precalculated animation is perfectly fine.
Methinks, that with complete automation it would be hard to achieve good quality.
At opposite, crafting it by hand would be a tiedous task (while not impossible, with a good sgf editor one can explore many relevant branches rather quickly)
The truth probabely is inbetween. For example we could choose a set of canonic shapes as staring points and let the computer drill them. Since it's about local tactics, there are chances that the analysis will be correct (at least for tsumego)
But there are others aspects to consider: personally i know almost nothing about visual learning process, there are questions a neuro-psicho-cognito-whatever specialist could answer on general basis, not related or little related to Go.
and so on
Another big field to investigate are sound fx we could drop in, AFAIK humans have more "processing power" for sound than for video: just consider the fact that if we mix two frequences, we still hear both while two mixed colors give one thrid color. So it could be a valuable resource for enriching the experience.