Updated 2026-07-21
Not in the way the phrase suggests. Current systems decode a narrow, trained-for set of intentions from brain signals. They do not read language you have not agreed to produce, and accuracy falls apart the moment conditions change.
A brain-computer interface does not receive thoughts. It records electrical activity, usually from the scalp, and a model learns to associate patterns in that activity with a small set of categories it was trained on: imagining moving your left hand versus your right, attending to one flickering light versus another, naming a pictured object.
That last one is worth sitting with, because it is the closest thing to reading words. Work on decoding semantic categories during picture naming asks whether meaning is recoverable from EEG while somebody speaks aloud. Recovering the category of a word a person is deliberately producing, in a lab, wired up, is a real result and a long way from reading unspoken thought.
The honest limitation in this field is not accuracy on a good day, it is what happens on a different day. Signals drift between sessions, between people, and between recording setups, and a decoder trained on one distribution degrades on another.
This is why so much current work is about adaptation rather than raw performance: methods that realign an implanted decoder as signals shift, and approaches for adapting across datasets when the original training data cannot be shared for privacy reasons. A system that needs recalibrating every morning is not reading your mind, it is being retaught your handwriting.
There is a recurring claim that one decoding pipeline is broadly best. Tested properly across many configurations and hundreds of thousands of subject-level results, average rankings turn out to hide per-subject reality: a method can top the table overall and still be the wrong choice for a great many individuals.
For a technology worn by one person at a time, that distinction is the whole game. The average does not use the device.
The useful applications are the ones that accept a narrow vocabulary and make it count: assistive control for people who cannot move, rehabilitation after stroke or in cerebral palsy, and clinical monitoring such as detecting burst suppression in intensive care or spotting seizures in EEG.
Those are not diminished goals. Restoring a reliable yes and no to somebody who has lost speech is worth more than a machine that guesses at your inner monologue, which is not what any of this currently does.
7 papers, each explained in plain language.
Decoding Semantic Categories from Picture-Naming EEG
arXiv (neural decoding)Average Rankings Mask Per-Subject Optimality: A Friedman-Nemenyi Benchmark of EEG Motor-Imagery BCI Decoders
arXiv (BCI)Test-Time Adaptation of Spiking Neural Networks for Intracortical Neural Decoding using Membrane Potential Alignment
arXiv (BCI)An Enhanced Source-Free Unsupervised Domain Adaptation Framework for Cross-Dataset EEG Emotion Recognition via Predictive Coding and Test-Time Training
arXiv (EEG)Evaluation of EEG Foundation Models for Event-Based Burst-Suppression Detection in ICU
arXiv (EEG)What Holds Back Brain-Computer Interfaces? Uncovering Challenges and Opportunities in BCI-controlled Games for Cerebral Palsy Rehabilitation
arXiv (BCI)Sensory Restoration via Brain-Computer Interfaces: A Unified 2 x 2 Framework and Convergence Roadmap
arXiv (BCI)