Did they model both cue and response in the subject-level GLM? Or did they just model cue?
From my experience with such models, it's impossible to disambiguate one part of the trial (e.g. cue) from the other parts of the trial (response, etc), at least with fMRI. Because response must always follow cue, you can't get sufficient jittering to separate them. (That is, the usefulness of jittering is limited when you can't randomize the order of the events.) If they model multiple parts of the trial (e.g. cue, response and maybe other phases), they might get something that survives some measure of statistical significance, but it's hard to see how it would be a meaningful measure of anything.
I did see one paper that claimed to separate the trial parts by using incomplete trials. So if you have a trial with cue that's truncated, you could claim that the cue-only trials measure cue; and you could get response by comparing the parameter estimates for complete trials with parameter estimates for the truncated trials. But that has an obvious weakness: at least naively one would expect that the cognitive processes during truncated trials would be different from those in complete trials.
So, on the information you provide, I'd venture to say this is a weakness in the paper you're reviewing.