The Roads Not Taken: The Neural Dynamics of Lexical Alternatives During Free Language Production

Language is a flexible tool. When we want to say something, we rarely have only one way to say it. If a friend asks how a presentation went, we might choose to call it "amazing," "great," or simply "good." We usually pick the word that fits our mood and our relationship with the listener, but this choice implies that other options were available to us at the same moment. For decades, scientists have wondered if the brain holds these unchosen words in reserve, activating them alongside the one we eventually speak, or if it discards them immediately. The problem is that these unspoken words leave no trace in the air; they are silent, internal events that vanish the moment we make a sound. Because they cannot be heard or easily reported by the speaker, they have remained hidden from direct observation, known only through indirect clues like slips of the tongue or results from rigid laboratory tests.

A team of researchers has now peeled back this layer of silence, offering the first direct look at how the brain manages these invisible alternatives during real conversation. By combining high-resolution recordings of brain activity with the predictive power of modern artificial intelligence, they discovered that the mind does not wait to decide what to say before it begins to prepare. Instead, the brain simultaneously holds a broad, shifting set of potential words in a state of readiness, constantly updating this list as the conversation flows and even preparing options for words that will be spoken several seconds in the future.

To catch these fleeting thoughts, the researchers turned to a unique group of volunteers: four patients with epilepsy who had electrodes placed directly on the surface of their brains to monitor seizures. These electrodes allowed the team to record electrical signals with a precision that standard brain scans cannot match, capturing the activity of thousands of neurons as the patients engaged in natural, hour-long conversations with doctors and family members. The challenge was to know what the patients were thinking about before they spoke. To solve this, the team used a large language model, a type of artificial intelligence trained on vast amounts of text, to predict what word would come next in the conversation based on the words that had already been said.

The researchers focused on moments where the artificial intelligence guessed the next word incorrectly. In these cases, the top three guesses made by the computer were all words that the patient did not say. These unchosen guesses served as a window into the patient's mind. The team reasoned that if the brain was indeed holding multiple options in reserve, the neural activity recorded just before the patient spoke should contain traces of these specific, unspoken words. They analyzed the brain signals from a half-second before the target word was spoken, a window known to be critical for retrieving words from memory.

The results were clear and compelling. The neural activity in the brain, particularly in the left frontal region known for language processing, contained distinct patterns that corresponded to the unspoken alternatives. The researchers could decode the identity of the first, second, and third predicted words from the brain signals, even though the patient never said them. Crucially, the brain did not treat these alternatives as a vague, blurry group. Each unchosen word had its own unique signature in the neural activity, distinct from the others. This suggests that the brain maintains a parallel track of several specific candidates, each carrying its own semantic meaning and grammatical role, rather than just a general sense of "something similar."

The study also revealed how many words are held in this mental buffer. By testing how well the brain signals could be predicted when including more and more potential words, the researchers found that the neural activity improved as they added more candidates to the mix. The signal continued to grow stronger until about twenty words were included, suggesting that the brain is simultaneously juggling a surprisingly large set of possibilities—more than a dozen words at once—before settling on the one to articulate. This broad scope challenges the idea that the brain narrows its focus to just one or two options early in the process.

Furthermore, this mental list is not static; it is a dynamic, moving target. As the conversation progresses from one word to the next, the set of alternatives updates instantly. The brain does not wait for a sentence to finish before it starts planning the next part. The researchers found that while the patient was speaking a current word, their brain was already representing alternatives for the next word, the word after that, and even the one following that. These future options appear in the brain in a specific order, with alternatives for the immediate next word appearing first, followed by those for the second word ahead, and so on. This indicates that speech production is a highly forward-looking process, where the mind is constantly preparing a pipeline of choices for the words that have not yet been spoken.

The nature of these representations also offered a clue about how the brain organizes language. The unspoken words were not just floating as vague concepts; the brain activity showed that they carried specific grammatical information, such as whether they were nouns or verbs, in addition to their meaning. This supports the idea that the brain retrieves full lexical items—complete words with their grammatical roles—rather than just abstract ideas. Interestingly, the brain did not seem to prepare the sound patterns of these unspoken words in the same way it prepares the sounds of the word that is actually said, suggesting that the final step of turning a word into a sound happens only after a choice is made.

This work shifts our understanding of how we speak. It suggests that the fluidity of human conversation relies on a massive, parallel engine of preparation. We do not simply select a word from a single line of options; we navigate a wide, evolving landscape of possibilities, keeping many paths open until the very last moment. The brain is not just a speaker of words, but a constant simulator of what could be said, ready to pivot instantly as the conversation demands. By listening to the silent electrical whispers of the brain, the researchers have shown that the words we do not say are just as present in our minds as the ones we do.