Mind Reading? Not Quite: What I Learned About P300 Brain-Computer Interfaces (By Corbin Lu)

When I first became interested in brain-computer interfaces, one description appeared again and again: mind reading. I can understand why people use it. It is much easier to imagine a computer reading a thought than to explain how electroencephalography (EEG) records small changes in electrical activity from the brain, why the same choice may need to appear several times, and how software uses the resulting signals to estimate what the user was focusing on.

The problem is that the phrase started to become less convincing the more I read about how one relatively established type of BCI actually works. I am a mathematics student, not a neuroscience researcher, so I originally approached the subject as a beginner. What interested me was the gap between the complicated research papers I was reading and the much more dramatic descriptions of BCI that often appear outside research.

A visual P300 system gives a good example of that gap. In one of the early P300 spelling systems, letters and commands were arranged in a matrix. Rows and columns of characters flashed on a screen while the user focused on the one they wanted. When the target appeared, it could produce an event-related potential, which is a brief change in the EEG that happens in response to a specific event or stimulus. Researchers can use these responses to study how the brain reacts to things such as attention, sounds, or images. In this case, the response of interest is the P300, which tends to appear around 300 milliseconds after the target is noticed. The computer then combines information from several flashes to estimate the intended character (Farwell & Donchin, 1988).

The process is technically complicated, but one detail seems especially important: the possible choices already exist on the screen. The computer is not searching through everything in a person’s mind. The person is deliberately paying attention to one of the options that the interface provides.

I found it easier to understand by reducing the example even further. Imagine that the screen only has two choices, YES and NO. A person decides which answer they want and pays attention to it as the choices flash. Electrodes on the scalp record electrical activity, but the signal is not a clean message saying “YES.” It contains other brain activity, eye movement, muscle activity, and electrical noise. The system therefore has to process the signal to reduce the noise and work out which choice the user was most likely focusing on.

That description is much less exciting than saying a computer has read someone’s mind, but it is more interesting in another way. A weak and noisy biological signal can still become useful information after several stages of processing.

It also explains why accuracy is not automatically the same for everyone. Guger et al. (2009) tested a P300 interface with 100 participants and found that many could achieve high accuracy after a relatively short period, but performance varied and some participants could not use the system successfully during the test. This variation is easy to miss when only a single successful demonstration is shown.

The way people use the technology also highlights its limitations. Nijboer et al. (2008) studied a P300 communication system with people with amyotrophic lateral sclerosis (ALS), a condition that progressively affects the nerve cells controlling voluntary movement and can lead to severe muscle weakness. Their results showed that communication was possible under the study conditions, including for participants with severe motor impairment, although it was much slower than speaking or typing and was not always accurate.

However, these limitations do not make the technology unimportant. This was one of the things I had to reconsider while reading. My first instinct was to evaluate communication by comparing it with normal typing or speaking speed. For someone with very limited voluntary movement, however, even a slower selection system could create a communication route that otherwise would not exist. Technology does not have to look like science fiction to matter.

At the same time, it is important not to turn that situation into an inspirational story in which the technology simply solves everything. A P300 system may require calibration and repeated trials, and performance can be affected by the user and the particular interface. A recent systematic review also shows how much visual P300 systems vary in their design, equipment, classification methods, and applications (Kalra et al., 2023). Different interfaces may present choices in different ways, use different numbers or positions of electrodes, or process the EEG signals differently, which can affect the speed and accuracy of the system.

The term “mind reading” creates another problem because it can describe very different levels of inference. Brown (2024) argues that a system may infer something from brain activity without actually gaining reliable access to the full meaning of a person’s thoughts. A system that estimates whether someone selected YES instead of NO is doing something very different from retrieving an unspoken sentence exactly as it exists in that person’s mind.

That said, I am still not sure that the term “mind reading” is always completely useless. It can introduce the basic idea that information is being inferred from brain activity. But without an explanation after it, the phrase probably creates more confusion than it solves.

For me, a P300 BCI became easier to understand when I stopped imagining it as a machine listening to an internal voice. It is closer to a chain of steps: the system presents choices, a user deliberately attends to a target, EEG records a response, the signal is processed, and the software estimates which option the user intended to choose.

That process is already unusual enough. It does not really need the extra claim that a computer can freely read whatever someone is thinking.

-- Corbin Lu

Corbin Lu is an undergraduate student in the Faculty of Mathematics at the University of Waterloo, studying mathematics with an interest in scientific machine learning. Corbin is especially interested in neuroscience, brain-computer interfaces, and how computational methods can be used to better understand the brain.

References

Brown, C. M. L. (2024). Neurorights, mental privacy, and mind reading. Neuroethics, 17(2), 34. https://doi.org/10.1007/s12152-024-09568-z

Farwell, L. A., & Donchin, E. (1988). Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalography and Clinical Neurophysiology, 70(6), 510–523. https://doi.org/10.1016/0013-4694(88)90149-6

Guger, C., Daban, S., Sellers, E., Holzner, C., Krausz, G., Carabalona, R., Gramatica, F., & Edlinger, G. (2009). How many people are able to control a P300-based brain–computer interface (BCI)? Neuroscience Letters, 462(1), 94–98. https://doi.org/10.1016/j.neulet.2009.06.045

Kalra, J., et al. (2023). How Visual Stimuli Evoked P300 is transforming the Brain–Computer Interface landscape: A PRISMA compliant Systematic Review. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 31, 1429–1439. https://doi.org/10.1109/TNSRE.2023.3246588

Nijboer, F., et al. (2008). A P300-based brain–computer interface for people with amyotrophic lateral sclerosis. Clinical Neurophysiology, 119(8), 1909–1916. https://doi.org/10.1016/j.clinph.2008.03.034