Jan,
There’s another related issue, I think, with Fyfe et al. As I understand it, what they did to get the observed trend and its uncertainty was to consider 100 (IIRC) ensemble member realisations of the HadCRUT4 data. Hence, their uncertainty is really the error on the mean of all 100 realisations, not the uncertainty in the trend for a single ensemble member. Their uncertainty is therefore, I think, smaller than it would be had they used the uncertainty in the trend for a single ensemble member, rather than the error on the mean for all 100 ensemble realisations. This, I think, is consistent with what you’re suggesting which is that they’re assuming that the mean trend is the actual forced and that it is unaffected by unforced variability.
So, as I see it, all they’ve really shown is that the mean trend in the HadCRUT4 data is barely consistent with the model trends, but they haven’t really shown that this is necessarily implies anything particularly significant, since they’ve assumed no role for unforced variability.
On a similar note, it is interesting that some use the large uncertainty in the trends to argue that we can’t rule out that there’s been no warming since 1998, and then, when doing a model-observation comparison, use the error on the mean (which is smaller than the uncertainty in the trend) to argue that the models and observations are inconsistent.