The real problem here is that what is "2-3 sec on average" worth of work for an "average pc" might be virtually no work at all for a gpu/asic. What is 2-3 seconds of work for an asic might take hours on an average pc. It can't really be constant, it needs to adapt to network conditions.
It's not a problem actually, we can try to find some heavy hash alghorithm which require huge amount of memory (bye-bye asics) and which are very inconvinient for GPUs.
While I tend to agree that this would be quite a good solution, I have yet to find a reasonable way to actually execute it. If the work "mixed in" has too much impact on the physics it makes for some unpleasant experience for a human player. If it does not have enough impact then bots can simply ignore the perturbations and solve without the extra work and just "double check" their solutions for validity with the work cycles added back in.
We don't even need to make them unnoticable. Parameters can vary within the range +/- 20%. I bet it wouldn't be a problem for people even if they would need to drive bike and 5 seconds later it would transform to hellicopter.
Anyway, the final decision is yours of course...