Can a brain implant read the voice in your head?

There's a voice in your head right now, reading these words. You're not moving your mouth and you're not making a sound. For almost the entire history of neuroscience, that inner voice was private in the strongest possible sense, locked inside the skull with no way out.

A study published in Cell on August 21, 2025, by Erin Kunz, Frank Willett, and their colleagues at Stanford and the BrainGate project chips away at that. Working with four people who'd lost the ability to speak clearly, from ALS or a brainstem stroke, the team read words they were only imagining, straight from electrodes in the motor cortex, the strip of brain that plans movement.

Key takeaways

In the motor cortex, imagining a word produces almost the same pattern as trying to say it out loud, just weaker. Inner speech is a scaled-down copy of the plan to speak, and a decoder trained on attempted speech can read it.
Read freely from a 125,000-word vocabulary, it's rough. Word error rates ran from about 26 to 54 percent across participants, far from a clean transcript but well above chance. On a fixed 50-word set it fell to 14 to 33 percent.
The team showed inner speech can leak, then built safeguards, including a keyword “mental password” that worked 98.75 percent of the time.
This is invasive, early, proof-of-concept work in four people. It reads the motor plan for speech, not free-form thought, and it isn't mind reading.

What they actually did

The four were enrolled in the BrainGate2 trial, and each had thin arrays of microelectrodes sitting in the speech-movement part of the motor cortex, along the precentral gyrus. It's the same class of implant behind the earlier devices that let people with paralysis produce text by trying to talk. Three had ALS and one had a brainstem stroke, and all of them had lost clear speech. The task set up a clean contrast. In one condition the participant tried to physically speak, working the mouth and throat even when little or no sound came out. In the other they only imagined saying a word, with no attempt to move. The team recorded the firing of the neurons in each case and trained software to turn those patterns back into words.

Inner speech is a faint copy of trying to speak

The first result is the one that makes everything else possible. When a participant imagined a word, their motor cortex produced almost the same pattern as when they tried to say it out loud. The paper states it plainly.

“The representation of inner speech was highly correlated with attempted speech, though we also identified a neural ‘motor-intent’ dimension that differentiates the two.”
Kunz et al., Cell, 2025

So imagined speech isn't a separate secret code. It's a scaled-down version of the same motor plan, quieter but the same shape, weaker than a real attempt but pointing the same way. It wasn't a perfect copy, though. Alongside the shared pattern the team found a separate signal, that “motor-intent” dimension, marking whether you're truly trying to speak or only picturing it. Hold onto that difference, because it's what makes the privacy fix possible later. The voice in your head, it turns out, runs on the machinery for speaking, just turned down.

When someone only imagined a word, their motor cortex showed the same pattern as trying to say it out loud, just weaker. Inner speech isn't a separate thought, it's a quieter copy of the plan to speak. Schematic of the population pattern, after Kunz et al., Cell, 2025.

They decoded imagined sentences, imperfectly

Because inner and attempted speech share so much, a decoder trained only on attempted speech can already pick up imagined speech. That cross-over is the whole trick.

“When evaluating inner speech with a decoder trained solely on attempted speech, performance was above chance for all participants … indicating that a speech BCI trained only on attempted speech signals can decode inner speech.”
Kunz et al., Cell, 2025

How well? Here's the honest picture, and it's a range, not a headline. Reading freely imagined sentences from a vocabulary of 125,000 words, close to the full working vocabulary of English, the word error rate ran from about 26 to 54 percent across participants. One example block landed at 52 percent, with a 95 percent confidence interval of 42.1 to 61.8. On a smaller fixed set of 50 words it did better, missing 14 to 33 percent. That's far better than the near-total error you'd get guessing at random on a vocabulary that size, but it's nobody's clean transcript.

How often the decoder got an imagined word wrong, across participants. On the enormous 125,000-word vocabulary the error ran from about 26 to 54 percent. On a small fixed set of 50 words it fell to 14 to 33 percent. Both sit far from the near-total error of guessing at random. Data from Kunz et al., Cell, 2025.

It helps to see what a run at that error rate actually looks like. These are real decodes of sentences one participant only imagined, on the full 125,000-word vocabulary.

ImaginedI think it has the best flavor.

DecodedI think it has the best player.

ImaginedWe don't have a real strict budget.

DecodedWe don't have a drill press but.

ImaginedThat person rushes you.

DecodedDuring red you.

ImaginedI don't know how long you've been here.

DecodedI don't know how long you've been here.

ImaginedSome of the experiments fail.

DecodedSome of the instruments they.

Real-time decodes of sentences participant T16 only imagined, on the 125,000-word vocabulary. Words the decoder got wrong are marked. Sometimes it lands the sentence, more often it slides into something plausible but wrong. Recreated from Figure 3 of Kunz et al., Cell, 2025, licensed under CC BY 4.0.

The privacy problem, answered in the same paper

Here's the twist that traveled fastest. If a decoder built to help someone talk can also pick up imagined speech, it might quietly transcribe things the person never meant to say. The team showed the leak was real, decoding inner speech while participants silently counted or ran through a memorized sequence in their heads, with no intention of speaking at all. They named the worry directly.

“A concern raised by both researchers and potential users is ‘mental privacy’ … whether a speech BCI would ‘be able to read into thoughts or internal monologues of users.’”
Kunz et al., Cell, 2025

Their answer came in two parts, both leaning on that motor-intent signal from earlier. The first trains the decoder to treat inner speech as silence, so it stays tuned to real attempts to speak and lets private thoughts pass. The second is a keyword, a kind of mental password. The system stays locked until the user imagines a chosen phrase, in the study a deliberately silly one, “chitty chitty bang bang,” and only then begins decoding. In real-time tests it caught that keyword 98.75 percent of the time. It's rare to see a paper raise an obvious fear about its own technology and answer it in the same experiments.

Illustration of the keyword safeguard. On the left, a locked brain-computer interface does not decode the thought 'I am tired.' On the right, imagining the keyword unlocks the interface, which then begins decoding. A table lists keyword detection accuracy of 98.75 percent. — The keyword lock, from the study itself. Until the user imagines the unlock phrase, the interface stays locked and decodes nothing. Figure 7E from Kunz et al., Cell, 2025, licensed under CC BY 4.0.

What this does and doesn't mean

“Brain implant reads your inner voice” invites a far bigger claim than the study supports. This isn't mind reading. The device isn't pulling arbitrary thoughts, images, or feelings out of the brain. It reads the motor cortex's plan for speech, the pattern tied to moving the mouth, whether you speak, try to speak, or imagine speaking. Anything that never takes the shape of words on its way to the muscles stays beyond its reach. The authors are blunt about the ceiling.

“It was not possible to accurately decode complete, intelligible sentences during free-form thinking.”
Kunz et al., Cell, 2025

And the caveats stack up. This runs on electrodes surgically placed in the brain, in four people, on a limited vocabulary, and even then it gets a quarter to more than half the words wrong. Nothing here works from outside the head, and nothing here reads a person who hasn't chosen to take part. It's early, proof-of-concept work.

What it does show is real, and it matters for the people it's built for. For someone who's lost the ability to speak, imagining a word takes less effort than straining to force it out. The authors found the inner-speech route “required less effort, offered improved comfort, and bypassed physiological constraints” that slow attempted speech, and it points toward a device that could listen to that quieter attempt instead. It also flags, early and in the open, that such a device has to be built to listen only when its user wants it to.

The honest open question

The study shows that inner speech lives in the speech-motor system as a quieter copy of attempted speech. It doesn't show that all inner experience works this way, and it doesn't settle where the rest of thought, the part not headed for your mouth, actually lives. Even the privacy fixes are, in the authors’ own words, “initial explorations,” a first pass at a problem that gets sharper the moment these devices leave the lab.