If the FMRI signal changes were large (say 10% or more), then using a non-task (pre-stimulus) baseline would make some sense. For example, suppose that the task-state is about 1/2 the duration of the experiment. Then if the baseline is 'b' and the signal change is 'x', then the mean signal would be about 'b+x/2', so using this mean would give a fractional ("percent") signal change of x/(b+x/2) which is approximately

p = x/(b+x/2) = (x/b) * 1/(1+x/(2b)) = (x/b) * ( 1 - (x/2b) + ... ) [Taylor series expansion of 1/(1+a) for small a]

So the "true" fractional signal change is P = (x/b) and the estimated value above is off by a fraction of 1-P/2. So if the true P = 5% = 0.05, then using the mean signal as a baseline will give a result off by a fraction of 0.975 = 2.5% error on the percentage. This estimation bias is so small compared to the effects of typical FMRI noise that it is reasonably ignored. But as the true P grows, so does the error in using the mean as the baseline.

In a nutshell, the above is why we use the mean signal -- otherwise, we would require a long period before the tasks with no stimulation to get a decent estimate of the baseline. If there are only 5 TRs before the first stimulus (a very common situation), then the estimate of 'b' from those 5 points would be very noisy.

Also, note that AFNI does per-voxel time series normalization, rather than normalizing all the voxels together.

Here, I've ignored the effects of long term signal drift, which is a related issue that also needs to be addressed in the processing. I've also ignored the factoid that baseline itself fluctuates due to resting-state neural activity, so the concept of a "baseline" is a little problematic.