A man who lost the ability to speak sat in a research lab, silently imagining words inside his head. On a nearby screen, those words appeared as text. No mouth movement. No whisper. No physical effort of any kind. A tiny array of electrodes implanted in a brain region called the supramarginal gyrus had picked up the firing patterns of individual neurons as he thought each word, and a decoding algorithm translated those patterns into readable language in real time.
The results, published in May 2025 in Nature Human Behaviour, represent the clearest demonstration to date that purely imagined speech, the internal voice people use when thinking in words, can be captured by a brain-computer interface and converted into communication. A Nature news analysis described the device as the most successful yet at decoding internal speech, a milestone that researchers in the field have pursued for more than a decade.
Why imagined speech is a harder problem
Most high-profile speech neuroprostheses work by decoding attempted speech, where a paralyzed person tries to move their lips, tongue, and vocal tract even though no sound comes out. Those systems tap into motor cortex signals tied to the physical mechanics of speaking. They have produced impressive results: a separate clinical trial published in the New England Journal of Medicine showed that an implanted neuroprosthesis could decode attempted speech into words displayed on a screen and used in real conversation, with roughly 30 minutes of calibration before the participant could begin communicating. The NIH confirmed the trial’s key parameters, including implant location, vocabulary size, and reported accuracy.
But attempted-speech systems have a fundamental limitation. They require the user to generate motor signals, even faint ones. For patients with the most severe forms of paralysis, including locked-in syndrome, where a person is fully conscious but unable to move any voluntary muscle, those motor signals may not exist. Imagined speech bypasses the motor system entirely. The supramarginal gyrus sits in the parietal lobe, a region associated with language processing rather than movement planning. Recording from it means capturing the neural signature of a word a person merely thinks, not one they try to physically produce.
That distinction is what makes the Nature Human Behaviour result significant. The participant was asked to imagine speaking specific words from a defined vocabulary set during controlled laboratory sessions. The decoding algorithm matched neural firing patterns to intended words and produced text output. The National Institute of Neurological Disorders and Stroke issued a press release providing error-rate ranges and vocabulary context, and framed the advance as opening a path to communication that requires no physical effort at all.
What the system cannot do yet
The published data describe controlled sessions with cued word lists, not open-ended conversation. The participant imagined specific words when prompted. Whether the system can scale to the thousands of words needed for natural, free-form dialogue has not been demonstrated. This was a proof of concept, not a finished product.
Long-term durability remains an open question. The Nature Human Behaviour study reports short-term vocabulary and error-rate figures, but multi-month or multi-year accuracy logs from the supramarginal gyrus implant have not been released. Neural signals are known to drift over time as brain tissue reacts to implanted electrodes, and it is not yet clear how stable the decoding remains across sessions separated by weeks or months.
No direct statements from the participant about his experience using the device appear in the available research record. That absence is common in early-stage neuroscience trials, but it leaves a gap: readers have no way to know how the technology felt from the inside, whether it was effortful or intuitive, frustrating or liberating.
Safety oversight documentation is also limited in the public record. Only general NIH funding acknowledgments appear in the citation trails. The raw neural data and calibration logs from both the Nature Human Behaviour study and the NEJM trial remain unpublished, which means independent researchers can evaluate only summary statistics, not the underlying signals.
The privacy problem no one has solved
If a device can decode words a person is silently thinking, an uncomfortable question follows: what happens when it decodes words a person did not intend to share? Research on inner speech signals recorded from motor cortex has flagged this risk directly, noting the need for what scientists call intent gating, a software mechanism that activates decoding only when the user deliberately signals readiness to communicate.
No standardized protocol for intent gating exists. Regulatory frameworks have not caught up to the technology. The scenario is not hypothetical: anyone who has ever had an intrusive thought understands that the voice inside your head does not always say what you would choose to say out loud. A brain-computer interface that cannot distinguish between deliberate communication and mental noise would create problems that go well beyond engineering.
As of June 2026, no federal agency has issued formal guidance on mental privacy protections specific to speech-decoding brain implants. The conversation is happening in bioethics journals and at academic conferences, but it has not yet produced enforceable rules.
Where the field stands now
The trajectory of speech neuroprosthetics over the past several years has been steep. Earlier work, including a 2023 Nature study that used cortical surface recordings to drive both text and synthetic avatar outputs, established that brain signals tied to speech can be decoded with increasing speed and accuracy. The NEJM trial pushed attempted-speech decoding into clinical-grade territory. The Nature Human Behaviour result now extends the field into a fundamentally different signal type: neural activity that occurs without any motor component.
Other companies and research groups are working on related problems. Neuralink has implanted its N1 device in human participants and is testing cursor-control and communication applications. Synchron’s Stentrode, a stent-based electrode array inserted through a blood vessel, has been used in early trials for text-based communication. Neither company has published peer-reviewed results on imagined-speech decoding specifically, but both are expanding the range of what brain-computer interfaces can do.
For the patients who stand to benefit most, the people who cannot speak, cannot move, and cannot generate even the faintest motor signal, the supramarginal gyrus approach offers something the other systems do not: a way to communicate using thought alone. The gap between a controlled lab demonstration with a small vocabulary and a device that lets someone hold a real conversation is still enormous. But the gap between nothing and a first proof of concept is the one that matters most.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
