Context. The CoRoT mission was the first dedicated to the search for exoplanets from space. The CoRoT exoplanet channel observed about 163 600 targets to detect transiting planetary companions. In addition to the search for exoplanets, the extremely precise photometric time series provided by CoRoT for this vast number of stars is an invaluable resource for stellar studies. Because CoRoT targets are faint (11 ≤ r ≤ 16) and close to the galactic plane, only a small subsample has been observed spectroscopically. Consequently, the stellar classification of CoRoT targets required the design of a classification method suited for the needs and time frame of the mission. Aims: We describe the latest classification scheme used to derive the spectral type of CoRoT targets, which is based on broadband multi-colour photometry. We assess the accuracy of this spectral classification for the first time. Methods: We validated the method on simulated data. This allows the quantification of the effect of different sources of uncertainty on the spectral type. Using galaxy population synthesis models, we produced a synthetic catalogue that has the same properties as the CoRoT targets. In this way, we are able to predict typical errors depending on the estimated luminosity class and spectral type. We also compared our results with independent estimates of the spectral type. Cross-checking those results allows us to identify the systematics of the method and to characterise the stellar populations observed by CoRoT. Results: We find that the classification method performs better for stars that were observed during the mission-dedicated photometric ground-based campaigns.The luminosity class is wrong for less than 7% of the targets. Generally, the effective temperature of stars classified as early type (O, B, and A) is overestimated. Conversely, the temperature of stars classified as later type tends to be underestimated. This is mainly due to the adverse effect of interstellar reddening. We find that the median error on the effective temperature is less than 5% for dwarf stars classified with a spectral later than F0, but it is worse for earlier type stars, with up to 20% error for A and late-B dwarfs, and up to 70% for early-B and O-type dwarfs. Similar results are found for giants, with a median error that is lower than 7% for G- and later type giants, but greater than 25% for earlier types. Overall, we find an average median absolute temperature difference |ΔTeff| = 533 ± 6 K for the whole sample of stars classified as dwarfs and |ΔTeff| = 280 ± 3 K for the whole sample of giant stars. The corresponding standard deviation is of about 925 ± 5 K for dwarfs and 304 ± 4 K for giants. Typically for late-type stars, this means that the classification is accurate to about half a class. The catalogue is only available at the CDS via anonymous ftp to http://cdsarc.u-strasbg.fr/viz-bin/qcat?B/corot