There is a difference between what we say and what we think, because we generally don't say everything that comes into our heads. However, this may be what could happen with brain implants. These devices implanted directly in contact with the brain can read the neuronal activity of a region. This is particularly the case for implants to decode speech, which allow paralyzed people to communicate by predicting what the person is trying to say.
And that's the difference: the user must attempt to speak, because the implant reads the activity of the neurons that control the muscles needed for speech. This is therefore a voluntary action by the user, and only what they want to say will be said. Except that a new study has just shown that these implants can also "read" the users' internal thoughts, that is, what they say to themselves, but do not necessarily want to say to others. This discovery was presented in the journal Cell August 21, 2025 by researchers at Stanford University, USA.
Inner speech could be used to decode real speech
This "internal speech" is, for example, what you hear when reading silently, or when you try to record information by repeating it in your head. This little voice is extremely intimate and should remain so... but it also has the advantage of not requiring the effort of trying to speak, which can be difficult for paralyzed people. It would therefore be useful for decoding speech more effectively.
Both types of speech activate the same neurons
The researchers tested the possibility of decoding this internal thought in four participants who already had a speech-decoding implant (BrainGate), placed in contact with the motor cortex. To differentiate between internal speech (what the users think) and "real" speech (what they want to say), they implemented strategies to activate the former. In the first exercise, the participants watched words appear on a screen, associated with colors that indicated whether they should just see them, read them in their heads, or try to say them.
These three tasks activated the same neurons, in the same region of the cortex read by the implant, but at different levels. Internal speech and silent reading resulted in lower activation. It was as if there were a threshold of activity at which these neurons understood that the user really wanted to say these words aloud, and only from this threshold would they send this information to the muscles involved in speech. Users reported that it was easier for them to activate the implant with this internal speech than with real speech, probably due to this lesser effort.
Several activities engage this internal thinking
In a second exercise, participants had to think or speak sentences, composed of a vocabulary limited to 50 words. The implant was able to decode the thought sentences with an error rate ranging from 14 to 33 % depending on the participant. Then, the same exercise, but with a vocabulary of 125,000 words, resulted in an error rate ranging from 26 to 54 %. The error rates were lower when the participant attempted to speak the sentences, showing that the higher level of neural activity during real speech makes the implant's reading more reliable.
To confirm that this was indeed internal thinking, the researchers used other experiments. For example, users were asked to remember a sequence of movements to perform with a joystick (up, left, up, for example), which they had to reproduce a few seconds later. To remember, the participants involuntarily repeated these instructions in their heads, which was visible to the implant. Or they were asked to think about something else from their lives (their favorite food, for example), which caused the same neurons to activate more than if they were told to think about nothing.
A password to protect thought
The implants were therefore able to decode this internal speech of the participants, which was useful for facilitating speech decoding... but with the risk that the users' intimate thoughts would be made public. To avoid this possibility, the researchers implemented a security system activated by a password. The user had to think of this password to activate the implant's decoding system, and if they didn't think of this keyword, no speech was decoded. This security system worked with a success rate of 99 %, showing that it would be a reliable option to ensure that the user maintains control over what they want to share with the implant. However, more studies are needed to be sure, since the risk is enormous.
English 
French