The identification of genes in newly determined vertebrate genomic sequences can range from a trivial to an impossible task. In a statistical preamble, we show how "insignificant" are the individual features on which gene identification can be rigorously based: promoter signals, splice sites, open reading frames, etc. The practical identification of genes is thus ultimately a tributary of their resemblance to those already present in sequence databases, or incorporated into training sets. The inherent conservatism of the currently popular methods (database similarity search, GRAIL) will greatly limit our capacity for making unexpected biological discoveries from increasingly abundant genomic data. Beyond a very limited subset of trivial cases, the automated interpretation (i.e. without experimental validation) of genomic data, is still a myth. On the other hand, characterizing the 60,000 to 100,000 genes thought to be hidden in the human genome by the mean of individual experiments is not feasible. Thus, it appears that our only hope of turning genome data into genome information must rely on drastic progresses in the way we identify and analyse genes in silico.