Delimitations of species are crucial for correct and precise identification of taxa. Unfortunately "species" is more a subjective than an objective concept in taxonomic practice due to difficulties in revealing patterns of infra- or inter-specific variations. Molecular phylogenetic studies at the population level solve this problem and lay a sound foundation for DNA barcoding. In this paper we exemplify the necessity of adopting a phylogenetic concept of species in DNA barcoding for tree peonies (Paeonia sect. Moutan). We used 40 samples representing all known populations of rare and endangered species and several populations of widely distributed tree peonies. All currently recognized species and major variants have been included in this study. Four chloroplast gene fragments, i.e. ndhF, rps16-trnQ, trnL-F and trnS-G (a total of 5040 characters, 96 variable and 69 parsimony-informative characters) and one variable and single-copy nuclear GPAT gene fragment (2093-2197 bp, 279 variable and 148 parsimony-informative characters) were used to construct phylogenetic relationships among the taxa. The evolutionary lineages revealed by the nuclear gene and the chloroplast genes are inconsistent with the current circumscriptions of P. decomposita, P. jishanensis, P. qiui, and P. rockii based on morphology. The inconsistencies come from (1) significant chloroplast gene divergence but little nuclear GPAT gene divergence among population systems of P. decomposita + P. rockii, and (2) well-diverged nuclear GPAT gene but little chloroplast gene divergence between P. jishanensis and P. qiui. The incongruence of the phylogenies based on the chloroplast genes and the nuclear GPAT gene is probably due to the chloroplast capture event in evolutionary history, as no reproductive barriers exist to prevent inter-specific hybridization. We also evaluated the suitability of these genes for use as DNA barcodes for tree peonies. The variability of chloroplast genes among well-defined species or population systems of a species complex is 4.82 times the figure within the groups, and the GPAT gene is twice as variable between the groups as within the groups. The number of completely divergent sites is sufficient to mark the two subsections, the two species in subsection Delavayanae, and the well-divergent species in subsection Vaginatae. But the genes currently used either from the chloroplast genome or from the nuclear genome alone cannot correctly assign samples of P. decomposita, P. jishanensis, P. qiui, or P. rockii to the species as currently defined. We conclude that (1) DNA barcoding should be based on prior phylogenetic studies to understand the evolutionary lineages and how well the taxonomic species correspond to the lineages; (2) it is unlikely to find a single short fragment as a barcode for every plant and such a fragment could result in misidentification when a chloroplast capture event happened in the evolutionary history of plants like tree peonies; and (3) we suggest striving for a universal marker at the familial level and locally universal barcodes within a family instead of looking for a universal barcode for all plants.