From brain activity to communication

A study published in Cell on August 21, 2025, examines how inner speech—words imagined without being spoken aloud—is represented in the motor cortex. The research considers both the use of these signals for assisted communication and the risk of decoding thoughts that a user does not intend to share. [S1]

The study analyzed microelectrode recordings from four adult participants enrolled in the BrainGate2 clinical trial. Researchers compared several activities, including attempted speech, inner speech, listening, and silent reading. [S1]

The central finding: attempted, inner, and perceived speech share neural representations in certain regions of the motor cortex, although inner-speech signals are generally weaker. [S1]

What the researchers achieved

1. Decoding imagined sentences

The team developed a brain–computer interface that converts neural activity into phoneme probabilities and then uses a language model to produce words and sentences. The system decoded self-paced imagined sentences in real time from three participants with severe dysarthria. [S1]

In tests using a 50-word vocabulary, word error rates ranged from 14% to 33%. In large-vocabulary evaluations involving 125,000 words and two participants, word error rates ranged from 26% to 54%. [S1]

Participants preferred inner speech because it required less physical effort and avoided the vocalizations that could occur while they attempted to speak. [S1]

2. Detecting forms of spontaneous inner speech

Researchers also identified signals associated with inner speech during activities in which participants were not explicitly instructed to imagine words. These included remembering sequences and silently counting colored shapes. [S1]

In another task, the system produced more words while participants responded to verbal-thinking prompts than when they were asked to clear their minds. The authors did not publish the decoded text because its relationship to the participants’ actual thoughts was uncertain and raised mental-privacy concerns. [S1]

Protecting mental privacy

The study describes a neural motor-intent dimension that helps distinguish an attempt to speak from inner speech. Using this distinction, the researchers tested two safeguards. [S1]

  • Training the system to treat inner speech as silence sharply reduced unintended outputs while largely preserving offline attempted-speech decoding performance. [S1]
  • Unlocking the system with an internal keyword allowed users to activate decoding intentionally; in a real-time test, correct keyword detection or rejection reached 98.75% accuracy. [S1]

What the study does not demonstrate

The research does not show that technology can completely and accurately “read” a person’s thoughts. The team decoded selected aspects of inner speech during controlled tasks, but it did not produce complete, intelligible sentences from free-form thinking. [S1]

The findings are based on only four participants, and the authors note that the use and strength of inner speech may vary between individuals. The proposed privacy safeguards are promising but require further study. [S1]

Why this matters in education

The study offers a timely example of the intersection of neuroscience, medicine, artificial intelligence, and ethics. It shows that technological progress must be accompanied not only by performance testing, but also by clear mechanisms that keep users in control of their communication and neural data. [S1]

For students, the research can open discussions about brain function, language models, assistive technology design, informed consent, and the right to mental privacy. [S1]