The Road Not Taken: Unsaid Word Alternatives are Represented in the Brain

Abstract

Human language allows multiple ways to express the same thought, implying that several lexical alternatives may exist in parallel before a single word is spoken or heard. We test the multiple alternatives-co-activation hypothesis by combining high-density ECoG during spontaneous dialogue with behavioral paradigms and ranked next-word predictions from large language models (LLMs). Behaviorally, words that LLMs rank as more likely continuations are recognized faster in a preregistered lexical decision task and produced with shorter inter-word intervals in free speech, indicating graded anticipatory activation of alternatives. Neurally, encoding models reveal that activity in classical language regions (IFG, STG) prior to word onset is predicted by embeddings of multiple top-ranked alternatives, not only by the word ultimately used; critically, mean embeddings that pool the top candidates outperform single-candidate embeddings, and the effect persists with arbitrary (non-semantic) embeddings that control for distributional similarity. Extending beyond a handful of options, encoding strength increases as embeddings are averaged across larger top-k sets, implying that a broad cohort of lexical candidates is simultaneously represented. Finally, models trained in comprehension generalize to production (and vice versa), preserving rank order and suggesting a shared neural code for candidate sets across modalities. Together, these findings provide direct evidence that the brain co-activates unsaid alternatives during natural language use and identify parallel candidate activation as a computational principle common to human comprehension, human production, and artificial language modeling.