Artificial intelligence can already convert certain patterns of brain activity into text, computer commands and synthetic speech. In laboratory studies, people with paralysis have used brain–computer interfaces to communicate, operate computers and reproduce voices they lost after a stroke or neurological disease.
That sounds like genuine mind reading, but the phrase creates the wrong impression. Current systems cannot secretly search through a person’s memories, discover political beliefs or reveal every passing thought. They normally decode a narrow, intentional activity, such as trying to speak, imagining specific words or moving a cursor.
The breakthrough comes from combining sensitive brain-recording equipment with AI models trained to recognise patterns too complicated for a human researcher to interpret directly.
Brain Activity Does Not Contain Written Sentences
The brain does not store a sentence as a neat line of text waiting to be extracted. When a person speaks, imagines speaking or moves a hand, populations of neurons produce rapidly changing electrical patterns.
A brain–computer interface records part of that activity. Machine-learning software then searches for repeatable relationships between the signal and an action the participant is performing.
In a speech system, researchers may ask someone to attempt particular words or sentences while electrodes record activity from areas involved in controlling the lips, tongue, jaw and vocal tract. The AI gradually learns which neural patterns are associated with smaller speech components, such as phonemes, before combining them into likely words.
Stanford researchers explain that their systems recognise recurring neural patterns associated with phonemes and then use computational models to assemble them into sentences. The AI is therefore interpreting intended speech activity rather than finding text physically written inside the brain. (Stanford’s explanation of inner-speech decoding describes how these signals are captured and translated.)
Implanted Electrodes Produce the Clearest Signals
The most accurate brain-reading systems currently require sensors placed directly in or on the brain. These invasive interfaces record much clearer activity than equipment positioned outside the skull.
In one major 2026 study, a man with amyotrophic lateral sclerosis used an intracortical brain–computer interface independently at home for nearly two years. The system decoded neural activity associated with attempted speech and movement, allowing him to communicate and control a computer despite severe paralysis.
He used it for more than 3,800 hours, generated over 1.96 million words and communicated at an average rate of 56 words per minute. During structured tests involving prompted words, the system reached more than 99% accuracy with a 125,000-word vocabulary. Results during unrestricted personal conversations were less consistently accurate, and the study involved only one participant. (The full study in Nature Medicine documents its long-term home use and limitations.)
This achievement is remarkable, but it is not unrestricted mind reading. The system was trained around one participant and primarily decoded his deliberate attempts to speak. It also required surgically implanted electrodes, wired connections and daily setup assistance.
AI Can Recreate a Voice From Intended Speech
Some interfaces now move beyond displaying words on a screen. They generate audible speech directly from neural activity.
Researchers from the University of California, Berkeley and UC San Francisco developed a system that produced speech in near real time for a woman who had lost the ability to speak after a stroke. Earlier versions introduced delays of several seconds, making natural conversation difficult. The newer system began producing sound within approximately one second of detecting an attempted utterance and continued generating speech as the participant tried to talk.
The AI was also able to produce words that had not appeared in its original training vocabulary, suggesting that it had learned components of speech sound rather than simply memorising complete recordings. (Berkeley Engineering’s report on the brain-to-voice neuroprosthesis explains the streaming system.)
Such technology could eventually restore not only words but also elements of identity. Researchers are exploring whether pitch, emphasis, emotional tone and loudness can be decoded so that synthetic speech sounds more like the individual rather than a generic computer voice.
Inner Speech Brings AI Closer to Thought Decoding
Attempted speech still involves the intention to move speech muscles, even when paralysis prevents movement. Inner speech is different. It is the silent voice a person experiences while imagining words without trying to say them aloud.
In 2025, Stanford researchers studied four people with severe speech and movement impairments who had microelectrode arrays implanted in motor areas of the brain. They found that imagined speech produced recognisable neural patterns similar to attempted speech, although the signals were weaker.
The AI could decode enough information to demonstrate that an inner-speech communication system was possible. Researchers believe this approach may eventually be less tiring for people who find repeated attempts to speak physically exhausting.
The study also raised an obvious concern: could a communication implant reveal words the user never intended to share?
The researchers developed a neural password system. Inner-speech decoding remained disabled until the participant imagined an unusual activation phrase. They also demonstrated a method for training systems designed around attempted speech to ignore unrelated inner speech.
These protections show that mental privacy is not merely a distant philosophical concern. Researchers are already designing technical boundaries between a thought intended for communication and one intended to remain private.
Non-Invasive Systems Can Decode Meaning Without Surgery
Brain implants offer stronger signals, but non-invasive approaches are advancing quickly.
A 2023 study used functional magnetic resonance imaging and an AI language model to reconstruct the general meaning of stories people heard or imagined. The output was not a word-for-word transcript. It produced sequences that often captured the central meaning using different language.
The decoder could also generate descriptions while participants watched silent videos. However, it needed many hours of personalised training inside an MRI scanner. It did not transfer effectively to an untrained person, and successful use required the participant’s cooperation. Deliberately thinking about something else could interfere with the decoding. (The study in Nature Neuroscience includes its privacy and cooperation experiments.)
An MRI machine is large, expensive and unsuitable for everyday communication. The experiment nevertheless demonstrated that AI could reconstruct aspects of meaning from non-invasive measurements rather than only detecting basic movement commands.
Meta Is Turning Brain Signals Into Typed Sentences
Meta is developing another non-invasive approach called Brain2Qwerty. It uses magnetoencephalography, or MEG, to measure extremely weak magnetic fields generated by brain activity.
Brain2Qwerty v2 was trained using approximately 22,000 sentences from nine volunteers. Each participant spent about 10 hours inside an MEG system while actively typing memorised sentences. The AI learned to translate the recorded signals into text.
Meta reported an average word accuracy of 61%, with its strongest participant reaching 78%. More than half of that participant’s sentences contained no more than one incorrect word. (Meta’s official Brain2Qwerty update provides the training and accuracy figures.)
The experiment did not decode random private thoughts. Participants were intentionally typing known sentences, and the model used activity related to language production and movement. MEG equipment is also large, costly and sensitive to environmental interference.
The significance lies in demonstrating how AI and language models can turn noisy signals recorded outside the skull into increasingly coherent language.
Language Models Help Fill Gaps in Noisy Signals
Brain recordings are incomplete and variable. Two attempts to say the same word may not create perfectly identical signals, and patterns differ significantly between people.
AI language models improve decoding by using context. When the neural data suggests several possible words, the model can consider which one makes the most sense within the sentence.
This creates both power and risk. Context can correct noisy signals and make communication faster. It can also cause the system to generate a plausible word the user never intended.
A brain decoder must therefore distinguish what came directly from neural evidence from what the language model predicted. Without that transparency, a fluent sentence could appear more certain than the underlying signal justifies.
Mental Privacy Is Becoming a Human-Rights Issue
As decoding improves, neural data could become one of the most sensitive forms of personal information. It may reveal intended movements, attention, emotional responses or language-related activity.
In 2025, UNESCO adopted the first global recommendation devoted to the ethics of neurotechnology. It calls for prior, informed consent, strict protection of neural data and safeguards against manipulation or interference with freedom of thought. It also warns against using neurotechnology to monitor workers or create behavioural profiles without appropriate protections. (UNESCO’s Recommendation on the Ethics of Neurotechnology treats mental privacy as fundamental to identity and autonomy.)
Future regulation may need to determine who owns neural recordings, whether they can be sold or used for advertising, and whether employers, insurers or law-enforcement agencies should ever be allowed to request them.
AI Is Reading Intentions, Not the Entire Mind
Today’s systems can recognise specific forms of brain activity under tightly controlled conditions. Their greatest near-term value is medical: restoring communication and digital independence to people affected by ALS, stroke, spinal injury or paralysis.
They do not provide silent access to an unwilling person’s complete inner world. Implant systems require surgery and individual training, while non-invasive systems usually depend on bulky equipment, long recording sessions and active cooperation.
The technology is nevertheless advancing rapidly. AI is becoming better at translating noisy neural signals into words, voices and digital actions. As accuracy improves, the scientific challenge will no longer be the only one that matters.
The deeper challenge will be ensuring that a machine capable of interpreting intended thoughts remains under the control of the person producing them.
