A brain-computer interface is a system that translates brain activity into commands for an external device. In a new study, researchers used one to explore a very human problem: communicating with words and gestures together after paralysis.

That combination matters to how I think about this story. A greeting can include a wave. Agreement can come with a nod. When we talk about communication technology, I think it is worth asking whether we are helping someone convey only a message, or more of how they want to express it.

The study, published September 14 in Nature Neuroscience, included three participants. Two used personalized avatars in real-time experiments. It is an early demonstration, not a finished communication system.

What the AI does

The AI consists of trained neural networks that classify recorded brain activity. Researchers paired neural recordings with the phrases and gestures participants were instructed to attempt. Training on both separate and simultaneous attempts improved performance across those contexts.

During one demonstration, a screen prompted a participant to attempt a phrase, a gesture, or both. Two models processed the resulting neural signals in parallel. Predicted speech appeared as text. Predicted gestures animated an avatar. The models could also select a rest state rather than produce an output.

This is a useful distinction from the AI writing tools many of us use. The task here is interpreting an attempted action, not composing a response on someone’s behalf.

What was demonstrated

In one participant’s simultaneous copy task, accuracy was 66% for gestures and 70% for speech, using ten gestures and ten phrases. Those are separate decoder results, not the percentage of complete conversations communicated correctly.

The supplementary material also describes cued exchanges with a conversational partner. These used a countdown before the participant’s response, with decoded text and avatar movement carrying the output. They should not be confused with unrestricted everyday conversation.

For me, the errors are part of the story, not a detail to hide. A communication aid has to earn the confidence of the person relying on it. Recognizing an intended message some of the time is meaningful research progress. Being dependable enough for daily use is a different standard.

The person remains central

This work depends on an implanted recording system, a restricted vocabulary, and a very small participant group. The authors also disclose relevant patent interests and a company connection.

I would want future reporting to answer practical questions alongside accuracy: How much effort does use require? How easily can someone correct a mistake? Does the person actually prefer communicating this way?

Those questions do not diminish the work. They are how I would judge whether the next version becomes useful outside a research demonstration.

What makes this worth covering in Where AI Actually Works is the purpose of the collaboration. The people involved are working toward another way for someone to express their own intent. That is a specific, human reason to build an AI system, and a good reason to follow the evidence as it develops.

Subscribe to Where AI Actually Works for one (or two) carefully researched story each week about humans and AI solving real problems.

---

This article was researched and written in partnership with AI. Every load-bearing figure traced back to the primary source and verified by a human before publication. The judgment about what to include, and what to leave out, is my own. Writing a series about humans and machines working together, it would be a little strange to pretend otherwise.

---

Sources:

1. Brosler, Liu, Silva and colleagues. “Simultaneous speech and gesture decoding for multimodal communication in paralysis.” Nature Neuroscience, September 14, 2026. Primary study, methods, results, and competing interests.

  [Read the study](https://www.nature.com/articles/s41593-026-02446-2)

2. The study’s supplementary information. Includes additional methods, the clinical protocol, and descriptions of the demonstration videos. The article’s supplementary section also provides the videos themselves.

  [Read the supplementary information](https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41593-026-02446-2/MediaObjects/41593_2026_2446_MOESM1_ESM.pdf)

These materials belong to the same research project, not independent replications. The numerical result in the article refers to one specified task and participant.