That many species of songbirds learn their songs imitatively is well established, but it is less clear why they do so. A component of the developmental-stress hypothesis posits that young males in good condition learn songs more accurately than males in poor condition and that females use learning accuracy as an honest signal of male developmental history. An unresolved problem is how females reliably assess learning accuracy when they are not certain of the identity of the male's tutor and thus the specific model from which a song was copied. We therefore investigated whether song learning accuracy assessment (SLAA) can be reliable, using evolutionary simulation models of song learning. We found that SLAA is indeed less reliable than assessment in which male signals are compared to an unlearned standard, as a result of three types of errors in matching songs to their models. In the simplest models, SLAA was particularly unreliable, but when the model is made more realistic by including features such as geographically constrained learning, repertoire complexity, and, in particular, song categorization, the reliability of SLAA increased. Our results demonstrate a range of conditions under which the assessment of song learning accuracy might be reasonably reliable and therefore likely to evolve.