While lexicostatistics and glottochronology had been suffering a lack of prestige for a long time, the integration of stochastic methods taken from genetics has initiated an unexpected revival of these scorned disciplines. The proponents of these "new quantitative methods" in historical linguistics claim that the procedures are relatively robust regarding errors in the data (wrong cognate judgments, undetected borrowings, or wrong translations). In order to check this claim, we have investigated the differences and errors in two large lexicostatistical datasets and tested their influence on the topologies of computed family trees. Our results show clearly that the shortcomings of lexicostatistics and glottochronology have not been overcome by these new computation methods: the main problems of lexicostatistics and glottochronology, the translation of basic concepts into individual languages, and the execution of cognate judgments are still so grave that no reliable results can be drawn from these methods.