Similarity-induced interference or facilitation in language production reflects representation, not selection.

Researchers have long interpreted the presence or absence of semantic interference in picture naming latencies as confirming or refuting theoretical claims regarding competitive lexical selection. But inconsistent empirical results challenge any mechanistic interpretation. A behavioral experiment first verified an apparent boundary condition in a blocked picture naming task: when orthogonally manipulating association type, taxonomic associations consistently elicit interference, while thematic associations do not. A plausible representational difference is that thematic feature activations depend more on supporting contexts. Simulations show that context-sensitivity emerges from the distributional statistics that are often used to measure thematic associations: residual semantic activation facilitates the retrieval of words that share semantic features, counteracting learning-based interference, and training a production model with greater sequential cooccurrence for thematically related words causes it to acquire stronger residual activation for thematic features. Modulating residual activation, either directly or through training, allows the model to capture gradient values of interference and facilitation, and in every simulation competitive and noncompetitive selection algorithms produce qualitatively equivalent results.