Just to pipe in on this topic.... If you are doing your group analyses by performing a statistical test across the individual subjects' r-to-Z transformed maps, then removing the temporal autocorrelations at the subject-level using 3dREMLfit will affect the group map. The reason for this is that the REMLfit operation reduces the subjects' r-values, which results in lower Zs. Ziad, of course, knows this is certainly the case; he was merely testing us to see if we are awake…
We've looked at this in our data and find a pretty substantial drop in the r-values using 3dREMLfit (vs. 3dDeconvolve), although the group maps overall don't look too different assuming you use a slightly lower p-threshold on the group map. So, 3dREMLfit will affect the group maps (quantitatively, if not qualitatively) if you are performing your group analysis on r-to-Z transformed data (which is what most folks do). This fact is separate from the issue of whether or not to remove temporal autocorrelations in functional connectivity analyses, and whether it might matter if your connectivity analyses are conducted in task data or task-independent (resting state) data. On that issue, Ziad makes some very good points.