Inside the Stanford Lab Building an AI That Reads Pre-Verbal Bilingual Toddlers’ Minds

The lab is located on the second floor of a Stanford research building, which has a subtle coffee and printer toner odor. This is the kind of place where someone would be writing a grant rather than attempting to read a toddler’s thoughts. However, that is essentially what is taking place here. A small group of computer scientists, neuroscientists, and a few obviously exhausted graduate students are developing what they almost reluctantly describe as a system that may one day be able to decipher a pre-verbal bilingual child’s thoughts before the child has the words to express them in either language.

It sounds like science fiction. For the most part, it isn’t. The foundation is taken from Frank Willett’s lab, where paralyzed adults’ inner speech has already been translated into text using brain-computer interfaces. Even the researchers who created it were taken aback by the results, which were published in Cell this past August. The arrays are smaller than a baby aspirin, and the algorithms were trained on phonemes. If a stroke patient is able to think the word “water” and have it appear on a screen, the next logical but somewhat unsettling question is: what happens if the speaker has no speech at all?

The Stanford team is currently probing that question, and it’s more difficult than it seems. Toddlers move around a lot. Toddlers who are bilingual do not think in neat linguistic categories, particularly those whose homes combine Mandarin and English or Spanish and English in the middle of sentences. Of course, the team isn’t implanting electrodes in kids—that would be a completely different ethical world. Rather, they are employing non-invasive imaging in conjunction with multi-modal language models—the same general family of tools that other Stanford researchers, such as Sang Truong and Sanmi Koyejo, have been testing on pediatric speech disorders. There is a reason for the overlap. From a different perspective, it’s the same issue.

Observing the team in action reveals how unsure they are. The breathless confidence that sometimes comes from AI startups is absent. A postdoc half-joked that the models are more frequently incorrect than correct, which is consistent with Truong’s group’s findings that none of the fifteen top language models tested on pediatric speech achieved even 55% accuracy in diagnosing basic disorders. The information is disorganized. The toddlers are more disorganized. Furthermore, it turns out that bilingual cognition behaves differently than monolingual training corpora predict.

Even though nobody is entirely certain of the final product’s shape, there’s a feeling as you walk through the lab that something significant is being attempted here. “What is the phoneme of a thought that hasn’t chosen a language yet?” is written on a whiteboard near the door, and no one has removed it in weeks. As expected, investors are already circling. Some have arrived by plane. Most depart with more inquiries than when they arrived.

The quiet stakes are difficult to ignore. There are not nearly enough pathologists to assist the millions of children in the United States who struggle with speech and language. The ramifications go well beyond a single lab if even a small portion of Stanford’s efforts—decoding intent, mapping early bilingual neural patterns, and identifying children who require assistance before they fall behind—are successful. It’s another matter entirely if it works. The group refuses to comment. They are still paying attention.