I don't think getting a group mask slightly larger than the template is all that surprising really. If you have used an affine transformation method (@auto_tlrc) or a manual stereotaxic procedure (manual Talairach), the result will be a collections of subjects that are all still a little different than the template. Some brains will bow out or in a little more or less than the template, but a group mask that is the union combination of all brain masks will be larger because it includes all the brains that are a little larger than the template even if there is only one brain like that. It's difficult to assess the quality of the group alignment because the affine transformation doesn't mean that all structures overlap. You can look to see if the brains generally overlap, and the orientation is consistent. Use the afni GUI tools like overlay opacity, overlay/underlay switching and underlay edge enhancement. With segmentation at an individual level, you could assess the level of overlap among subjects and then potentially compare to another method. The nonlinear alignment method, auto_warp.py, is easier to assess because the brains generally overlap much more tightly. It's still not perfect, but it might be more like what you expect.