If what happened with Terry seemed like telepathy ( check out the previous post if you missed it), an inevitable question arises: How does a machine manage to know what a person is "thinking" and what they are trying to say? The answer begins with something that happens inside the brain constantly; even before moving a hand, uttering a word, or clicking on something, groups of neurons are already generating electrical activity linked to that intention.

For virtually all of human history, that information remained trapped within the body; to turn an intention into action, the brain had to rely on biological pathways—neurons, nerves, and muscles. However, brain-computer interfaces aim to create a secondary pathway, and this is where technologies like Neuralink come into play. Tiny electrodes record the activity generated by populations of neurons. The system doesn't find a written word inside the brain, nor is there a specific neuron signaling "right," "left," or "I want to speak"; instead, there are patterns of activity, and algorithms are trained to recognize them.

Imagine a person trying to move their hand to the right; even if an injury prevents the arm from responding, specific regions of the brain can still generate activity associated with the intention to make that movement. An AI repeatedly observes these signals as the person attempts the movement; the system records the brain activity, the person tries again—gathering more data and more patterns—until the algorithm begins to establish a link between specific neural activity and what the person intended to do.

This is how an intention can begin to transform into a digital command, which explains why the initial results from brain-computer interfaces seem relatively simple—moving a cursor, browsing the internet, controlling a video game, or moving a robotic arm. AI does not need to understand everything a person is thinking; it only needs to learn to recognize a specific set of signals. However, speech makes the problem far more complex.

When Terry attempts to say a word, an extremely rapid sequence of neural activity occurs, corresponding to the movements normally required to articulate or produce it—involving the tongue, lips, jaw, larynx, and breathing. The goal of Voice is to extract enough information from that activity to reconstruct what Terry intended to say. As the algorithms improve, so does the amount of information we can glean from those signals; this is the point where the nature of the technology begins to shift. Controlling a cursor essentially means translating brain activity into a limited set of possibilities—left, right, click—whereas reconstructing a conversation requires distinguishing a vast array of combinations in real time: words, phrases, rhythm, intonation, and eventually even elements of the voice itself.

This distinction helps explain why what we are witnessing now is so significant. We are not learning to "read" the entire brain; rather, we are becoming increasingly adept at decoding specific types of information it generates—and there are various ways to achieve this. Some interfaces utilize sensors placed outside the head, others position electrodes on the membranes enveloping the brain, and more invasive systems—such as Neuralink—place electrodes in close proximity to, or directly within, brain tissue. There is a reason for this: the closer the electrodes are to the neurons, the more detailed the signal can be, though this also increases the complexity of the medical procedure.

This trade-off between the volume of information and the level of invasiveness is giving rise to several distinct strategies. This is important because, while the United States—through Neuralink—is attempting to place thousands of contact points extremely close to neurons, another country is pursuing different paths; in one instance, it has achieved something that even Neuralink has not yet managed: transforming one of these chips into a product approved for commercial release to its citizens. But that is a topic for the next post.