What I would like to see on this is a bit less discussion about the RNG capabilities, and instead a statistical analysis of the actual random threads used in the experiments. It really makes no difference how well the machine works over one length of time or another for this particular experiment; what matters is whether the data sets used were at all predictable (and no, it does not have to be consciously detected--in fact, some classical conditioning experiments would argue that a phenomenon like this might work better without conscious awareness).
In other words, for the actual trials used...did the showing of one emotional pic reliably predict a few trials without an emotional pic? If this is the case, over the long haul we should expect to see precisely the sort of "precognitive" effect show up as an experimental artifact. The nice thing is, it is perfectly testable, even in a post-hoc analysis of Bem's data here (might do this, hey, a publication for moi!). We can predict that the subjects who had data sets where the (randomly generated) trials had predictable patterns in them (and this can be determined mathematically, so we don't need raters to determine this) should show higher precognition scores. It should be fairly simple analysis. (btw, this analysis is fairly standard in the "judgments of control" literature--when we look to see how much control subjects feel they have over a random event, we must always be sure to see what the actual outcome of the RNG gave them--just in case they were right!)
