Due to insufficient morphological diagnostic characters in larval fishes, it is easy to misidentify them and difficult to key to the genus or species level. The identification results from different laboratories are often inconsistent. This experiment aims to find out, by applying DNA barcoding, how inconsistent the identifications can be among larval fish taxonomists. One hundred morphotypes of larval fishes were chosen as test specimens. The fishes were collected with either larval fish nets or light traps in the northern, southern and northwestern waters of Taiwan. After their body lengths (SL) were measured and specimen photos were taken, all specimens were delivered, in turn, to five laboratories (A-E) in Taiwan to be identified independently. When all the results were collected, these specimens were then identified using COI barcoding. Out of a total of 100 specimens, 87 were identified to the family level, 79 to the genus level and 69 to the species level, based on the COI database currently available. The average accuracy rates of the five laboratories were quite low: 80.1% for the family level, 41.1% for the genus level, and 13.5% for the species level. If the results marked as "unidentified" were excluded from calculations, the rates went up to 75.4% and 43.7% for the genus and species levels, respectively. Thus, we suggest that larval fish identification should be more conservative; i.e., when in doubt, it is better to key only to the family and not to the genus or species level. As to the most misidentified families in our experiment, they were Sparidae, Scorpaenidae, Scombridae, Serranidae and Malacanthidae. On the other hand, Mene maculata and Microcanthus strigatus were all correctly identified to the species level because their larvae have distinct morphology. Nevertheless, barcoding remains one of the best methods to confirm species identification.