Teaching a computer to talk to someone’s brain

Proposed research direction · November 2026

Teaching a computer to talk to someone’s brain

A brain-computer interface can send as well as listen. The proposal is a model that learns to talk to a person through the implant: it sends a signal into the brain, and at first the same message is also written on the screen. Over a long ladder of levels the text fades until the signal alone carries the message, and the person can act on it. The model learns which signals the person can learn to read; the person learns to read them. Nothing about reading the person’s thoughts: this is the other direction.

Proposal only · no implant, no participant, no collaborator · the ladder can be prototyped with a stand-in channel such as a vibration band first

The setup

The person works through ordinary computer tasks: pick the right target, move a cursor, answer a question. Each task is a level, and there are many of them, say a thousand. Before each action the model sends a signal into the brain that carries the instruction, and the screen shows the same instruction in text. At level 1 the text is complete, so the person can do the task from the words alone and the signal is just along for the ride. As the levels climb, the text thins: a sentence becomes a word, a word becomes a hint, a hint becomes a blank. At level 1000 there is no text at all. If the person still does the task right, the signal taught it.

The model is trained on one thing: whether the person did the task correctly. It has to find signals the person’s brain can tell apart and come to understand, and it has a thousand levels of attempts to find them. This is scaffold fading, a curriculum where the support is removed as competence grows, and it has a second effect that a fixed signal set would not: the person and the model adapt to each other. The model settles on signals the person reads reliably; the person learns the model’s vocabulary.

The trap, and how to stay out of it

While the text is on screen, the person can read the text and ignore the signal, and the model gets credit for a correct action it had nothing to do with. Three rules keep the signal load-bearing from the first level:

• The signal is always sent. Text is the help; the signal is the lesson. No level is ever scored on text alone.
• Score every level twice. Some trials show the text; some, unannounced, do not. Accuracy on the quiet trials is the score. The gap between the two is how much of the person’s answer still comes from the screen.
• Fade on evidence, not on a schedule. The text thins only when quiet-trial accuracy holds above a bar. A fixed thousand-step schedule would fade the text whether or not the signal had been learned; the gate makes each level earn the next one, exactly as the gated level test does.

What to measure

The curve that matters is quiet-trial accuracy against level. If the ladder works, that curve rises as the text fades, and the gap between quiet and text-shown trials closes toward zero. Two more checks guard against fooling ourselves. Revisit early messages late in the ladder, with their text restored and then removed, to see whether the person still reads them or the model has drifted to a new vocabulary; that is the same forgetting question the gated level test asks. And run sessions on separate days, because a vocabulary that only works on the day it was learned has not been learned.

What would prove it wrong

Quiet-trial accuracy that stays flat while the text fades: the person was reading the screen all along and the model never found a signal the brain could take up. Or accuracy that collapses as soon as the text thins: the signal was being noticed but not understood, and more levels would not have fixed that. Either result says the scaffold was holding the weight.

First experiment, without an implant

None of this needs neural hardware to test as a training method. Replace the signal into the brain with any channel the person has never read before and must learn: patterns of vibration on a wristband, a sequence of tones, a pattern of lights at the edge of the screen. The model has to invent a vibration language the person can learn, with the same ladder: text fades on the quiet-trial gate, every level is scored twice, early messages are revisited. If the quiet-trial curve rises and the gap closes with a wristband, the method is worth bringing to an implant. If it does not, nothing about an implant would have saved it.

Scope: this is a proposal about how to train the model, not a medical device, a clinical protocol or a claim about any company’s hardware. What an implant may safely send into a brain is the device’s question and comes before any of this.