When are in-group linguistic conventions opaque to non-group members (teen slang like "rizz") or generally interpretable (regionalisms like "roundabout")? The formation of linguistic conventions is often studied in iterated reference games, where over repeated reference to the same targets, a describer--matcher pair establishes partner-specific shorthand names for targets. To what extent does the partner-specificity of these linguistic conventions cause them to be opaque to outsiders? We use computational models and experiments with naïve matchers to assess the opacity of descriptions from iterated reference games. Both human matchers and the computational model perform well above chance, suggesting that conventions are not fully arbitrary or opaque, but reflect aspects of shared semantic associations.