It is often assumed that ecological specialization represents an evolutionary “dead-end” that limits further evolution. Maximum-likelihood (ML) analyses on phylogenies for 15 groups of phytophagous insects revealed that high transition rates both to and from specialization occurred, but that the mean ratio of rates was significantly biased toward a higher rate to specialization. Here we explore the consequences of the fact that transition rates inferred by ML are affected not only by the distribution, but also by the frequency, of character states. Higher rates to the more common state were inferred in the analyses of Nosil (2002), in similar studies published since 2002, and in a small set of simulations. Thus, the ratio of the rate toward versus away from specialization was strongly, positively correlated with the proportion of specialist species at the tips of the phylogeny and whether transitions away from specialization occur remains unclear. Here we reexamine these data using methods that do not rely on directly comparing transition rates. Maximum-likelihood analyses show that models with no transitions in one direction (e.g., irreversible evolution as predicted by the “specialist as dead end” framework) are usually strongly rejected, independent of the proportion of specialists at the tips. Ancestral state reconstruction revealed two instances where generalists were unambiguously derived from specialists. Transition rates would need to biased 100-fold and 5000-fold toward specialization to reconstruct a history where these shifts from specialization toward generalization do not occur. The general conclusions of Nosil (2002) appear to hold; transitions in either direction likely occur such that specialization does not always limit further evolution. Most generally, inferences regarding character evolution can be strengthened by comparing models of character change and examining ancestor states, rather than only comparing parameter values.