When you read that scientists “decoded” someone’s thoughts or turned brain activity into text, the first question worth asking is: how were they reading the brain in the first place? The method matters, because no single technique captures everything. Three dominate the field: EEG, fMRI, and ECoG. Each one trades away something the others keep.
In short, you are juggling three properties: how precisely you can locate activity (spatial resolution), how quickly you can track it (temporal resolution), and how far you have to go into the body to measure it. EEG is fast but blurry. fMRI is sharp on location but slow. ECoG is excellent on both, but it takes surgery.
Three methods, side by side
| EEG | fMRI | ECoG | |
|---|---|---|---|
| What it measures | Electrical activity (scalp) | Blood-flow changes | Electrical activity (brain surface) |
| Spatial resolution | Low | High | High |
| Temporal resolution | High | Low | High |
| Invasive? | No | No | Yes |
EEG: fast, cheap, and blurry
Electroencephalography measures the brain’s electrical activity through sensors placed on the scalp. Its great strength is speed: it captures changes millisecond by millisecond, so it is very good for tracking when something happens in the brain. It is also cheap, portable, and completely non-invasive, which is why it shows up everywhere from hospitals to consumer headbands. The weakness is location. Because the signal passes through the skull before it reaches the sensors, EEG is poor at pinpointing exactly where activity comes from. It is like knowing a sound happened without knowing which room it came from.
EEG has a long history. The German psychiatrist Hans Berger recorded the first human electroencephalogram in 1924, and the basic idea has barely changed since. Sensors, usually placed using the standardised “10-20” layout, pick up tiny voltage fluctuations produced by large groups of neurons firing together. Researchers group those fluctuations into frequency bands: delta (below about 4 Hz), theta (about 4 to 8 Hz), alpha (about 8 to 13 Hz), beta (about 13 to 30 Hz) and gamma (above about 30 Hz). Different mental states tend to show up in different bands. Alpha rhythms, for example, rise when you close your eyes and relax.
The same speed that makes EEG useful also makes it noisy. Blinking, jaw clenching and head movement all produce electrical signals that can swamp the brain activity you want, and the skull smears whatever does get through. For brain-computer interfaces, that means EEG systems usually rely on a few strong, repeatable signals instead of reading detailed thoughts: a response that appears when you notice a flashing letter (the P300), changes in motor-cortex rhythms when you imagine moving a hand, or a flicker-following response in visual cortex when you stare at a pulsing target. Those are enough to move a cursor or spell a message slowly, but not to transcribe free-form thought.
fMRI: precise location, sluggish clock
Functional MRI works on a completely different principle. Rather than reading electrical activity, it tracks blood flow, on the logic that active brain regions draw more oxygenated blood. That gives it excellent spatial resolution. It can localise activity to small structures deep inside the brain, which EEG cannot reach. The drawback is time. Blood-flow changes lag behind neural activity by seconds, so fMRI is slow; it tells you where beautifully but when only roughly. It is also large, expensive, and immobile: you go to the scanner, not the other way round. Much of the “mind-reading” research, including our piece on mind captioning that turns fMRI scans into text, relies on exactly this precision.
The signal fMRI measures is called BOLD, short for blood-oxygen-level-dependent contrast. After a group of neurons becomes active, local blood flow rises and the balance of oxygenated and deoxygenated blood shifts, which the scanner picks up through its magnetic properties. The response typically peaks several seconds after the neural activity that caused it, and it is recorded in small cubes of tissue called voxels that are often only a few millimetres wide. That is why fMRI can tell you which part of a structure lit up, but not the order of events inside a fraction of a second. Work that recorded electrical and BOLD signals together in monkeys (Logothetis et al., 2001) found that BOLD tracks local synaptic input and processing more closely than the spikes of single neurons. In other words, fMRI is an indirect measure.
Decoding language is a good test of what the technique can do. In a 2023 study in Nature Neuroscience, Tang and colleagues trained a decoder on each participant’s fMRI recordings and used it to reconstruct the gist of continuous language from perceived speech, imagined speech and even silent videos. The output recovers meaning rather than exact words, and the team also reported that the decoder only worked when the participant cooperated, both when it was trained and when it was applied. That finding matters for the mental-privacy debate, but it also shows the limits: fMRI decoding is a laboratory tool built around extensive scanning of each person, not something that can be pointed at a stranger.
ECoG: speed and precision, at a cost
Electrocorticography places electrodes directly on the surface of the brain, under the skull. With no bone to blur the signal, it delivers both high spatial and high temporal resolution. That combination makes it the reference method for high-performance brain-computer interfaces and for detailed clinical work. The obvious cost is that it is invasive: it requires surgery, so it is used only when the medical or research justification is strong.
ECoG began as a clinical tool. Surgeons place grids of electrodes on the brain’s surface to record seizures in people with epilepsy, which helps locate where the seizures start before an operation. Because the electrodes sit under the skull and skip the blurring effect of bone, they can pick up fast activity in the high-gamma range that scalp EEG largely misses, and that activity is closely tied to speech and movement.
That is why recent speech BCIs have leaned on it. In a 2021 study in the New England Journal of Medicine, Moses and colleagues implanted a high-density electrode array over the speech-related sensorimotor cortex of a person who had lost the ability to speak after a brainstem stroke. Over 22 hours of recorded sessions the participant attempted to say words from a 50-word vocabulary, and deep-learning models decoded full sentences from the cortical activity. Two years later, Metzger and colleagues reported a participant using surface recordings to produce text at a median of 78 words per minute with a median word error rate of 25%, while also driving synthesized speech and a facial avatar. Everyday conversation runs at roughly 150 words per minute, so there is a long way to go, but these results are far beyond what scalp EEG has shown for speech.
Which one powers today’s BCIs?
It depends on the goal. Non-invasive, consumer, and many research BCIs run on EEG because it is safe and practical, even if coarse. The most capable, high-bandwidth interfaces, the ones aiming at fluent speech or fine motor control, lean toward ECoG or deeper implants, precisely because they need that spatial and temporal clarity. fMRI, meanwhile, is used less as a real-time interface and more as a research tool for decoding and mapping. To see how these feed into working devices, our guide on how brain-computer interfaces read your thoughts explains how they fit together.
Beyond the big three
These three are the headline methods, but they are not the whole toolbox. Magnetoencephalography (MEG) measures the tiny magnetic fields produced by neural currents. It offers millisecond timing with better localization than EEG, but it needs a magnetically shielded room and costly equipment. Functional near-infrared spectroscopy (fNIRS) uses light to track blood oxygenation through the scalp. It is portable and tolerant of movement, but its resolution is coarse and it shares fMRI’s blood-flow lag. At the invasive end, microelectrode arrays penetrate the cortex to record from individual neurons, the approach behind Neuralink and Paradromics, while Synchron places its sensors inside a blood vessel instead. Our comparison of Neuralink, Synchron and Paradromics covers how those designs differ.
How to choose a method
There is no best technique, only a best fit for the question you are asking:
- You need timing and portability: EEG. Tracking when attention shifts, spotting seizure activity or running a low-cost BCI all suit it well.
- You need to know where: fMRI. Mapping which regions support language or memory in a healthy volunteer is where it works best.
- You need both, and surgery is justified: ECoG. Restoring communication for someone with severe paralysis is the main example.
- You need to follow a person through daily life: EEG or fNIRS, because neither requires a scanner room.
What none of them can do yet
Decoding brain signals is different from reading a mind. Every decoder mentioned here is trained on one person’s data and usually has to be recalibrated, because signals drift from day to day and electrode positions are never identical. Most results come from a handful of participants in controlled settings, and error rates are still high enough that a person has to confirm what the system produced. The decoders also tend to recover what someone is trying to say or what they are perceiving, not private thoughts they never meant to express. Those limits are part of why the field is moving cautiously, and why the privacy questions raised by the fMRI work deserve attention before the technology improves. For a closer look at the trade-off between surgery and signal quality, see our guide to invasive vs non-invasive BCIs.
Key takeaways
- EEG is fast, cheap and safe, but blurry. fMRI is precise on location but slow. ECoG delivers both, at the price of surgery.
- The best speech results so far come from ECoG and other implanted electrodes, not from scalp or scanner methods.
- Every method decodes from one person’s own data, and none reads private thoughts on demand.
Sources
- Moses et al., “Neuroprosthesis for Decoding Speech in a Paralyzed Person with Anarthria,” New England Journal of Medicine, 2021. doi.org/10.1056/NEJMoa2027540
- Metzger et al., “A high-performance neuroprosthesis for speech decoding and avatar control,” Nature, 2023. doi.org/10.1038/s41586-023-06443-4
- Tang et al., “Semantic reconstruction of continuous language from non-invasive brain recordings,” Nature Neuroscience, 2023. doi.org/10.1038/s41593-023-01304-9
- Logothetis et al., “Neurophysiological investigation of the basis of the fMRI signal,” Nature, 2001. doi.org/10.1038/35084005
- National Institute of Biomedical Imaging and Bioengineering, Magnetic Resonance Imaging (MRI).
