I don't know if this has been clearly discussed elsewhere, but the recent publication of a couple of papers on the topic gives me a good excuse to talk about it. The first is Gillett et al, in GRL a little while ago, and there's a new paper by Gareth Jones et al in press at JGR. Both appear to perform fairly conventional D&A analyses on CMIP5 models. I've been fairly critical of the conventional D&A approach in the past, primarily on the grounds that the null hypothesis of no anthropogenic influence is always false a priori (and therefore a failure to detect an anthropogenic influence is always a matter of insufficient data). These recent papers point to another, arguably more terminal, problem. Attribution will inevitably fail as the anthropogenic effect increases!
Gillett et al present their results primarily in terms of a better constrained estimate of the transient response. What they don't point out is that they cannot actually attribute the warming to anthropogenic forcing. As conventionally portrayed (eg by the IPCC here), attribution requires that the observed changes are consistent with the modelled estimate. However, the Gillett et al estimate of the scaling factor on the anthropogenic (GHG) component is about 0.7, with a confidence interval that clearly excludes 1. The figure below is the relevant one from their paper, and the critical point to note is that the left-most red coloured bar (which represents their main analysis) does not reach 1 on the y-axis. In fact the only way they can get a result which is consistent with the model is to limit their analysis to 100y of data. So, according to the standard definition their analysis simply cannot attribute the recent warming to GHGs.
Jones et al analyse a larger set of models, and for about half of them, the confidence interval on the estimated scaling factor includes 1. Don't blame me for the horrible picture below which is all that is available in their pre-production proof (hopefully the final version will be readable). Again, the central issue is whether the red bars on the figure include the value y=1. By my count, 9 of them do but 7 of them don't. (5 of the ones that do include 1, also include zero in the ranges of both G and OA, indicating a failure to detect an anthropogenic inflence at all. This voids any attempt at attribution a priori. However, I'm not so interested in that here.)
A moment's thought should confirm that failure to "attribute" in this sense will be an inevitable consequence of gathering a sufficiently long and precise time series of data. The actual value of the scaling factor - ie the ratio of real to modelled forced response - is never going to be precisely 1, for any model, any more than the actual value of the forced response is zero. All such hypotheses based on point estimates are inevitably false, and a failure to reject them only ever meant we didn't have enough data. But now it is increasingly the case that we do have enough data to reject many of the models. And this problem will only get worse as the data will surely keep pouring in at a rate greater than model improvement can keep up.
It will be interesting to see how the D&A community addresses this problem. Atribution of the observed changes to GHG and other influences was touted as a major step forward when it was first achieved, so it would surely be rather embarassing to lose the ability to do this. It looks a bit like they are trying to just ignore it for now, but that can't really be tenable as a long-term strategy.