---
title: "AI Taught to 'Draw' Vision: How a Neural Network Controls Brain Stimulation for the Blind"
description: "Researchers from the US, Switzerland, and Spain published experimental results in Neuron: a neural network controls a 96-electrode implant in the visual cortex, allowing a blind patient to see flashes and outlines with lower energy consumption."
date: 2026-11-06T14:26:58.442Z
lang: en
url: https://xab.info/en/posts/ai-controls-brain-stimulation-for-blind-vision
tags: [ai, brain-implant, bionic-eye, neural-interface, deep-learning, vision-restoration, neuron-journal]
author: "Alexander Buran"
publisher: "XAB.info"
---

# AI Taught to 'Draw' Vision: How a Neural Network Controls Brain Stimulation for the Blind

![A man sits in front of a monitor with an electrode array on his head for AI-controlled vision.](https://xab.info/media/2026/11/06/ii-uchili-upravlyat-stimulyatsiey-mozga-slepym-videt/ii-uchili-upravlyat-stimulyatsiey-mozga-slepym-videt-1.webp)

## Key Points

- Deep learning AI controlled a 96-electrode implant in a human visual cortex in real time for the first time
- A 27-year-old trauma patient saw phosphenes — flashes and outlines — thanks to neural-network-adjusted stimulation
- The main challenge is non-linear cortical response: a fixed pattern doesn't work, requiring continuous AI adaptation
- The technology serves as a proof of concept for future commercial bionic eyes

An international team of researchers from the US, Switzerland, and Spain has presented the results of an experiment in Neuron magazine, in which deep learning algorithms took over the control of human visual cortex electrical stimulation for the first time. During the trial, a 27-year-old man who lost his sight due to an injury was able to see flashes of light and object outlines — known as phosphenes — thanks to an AI-controlled implant. According to RBC-Ukraine, the technology makes it possible to form sharper visual images with lower energy consumption compared to previous bionic vision approaches.

### How the Implant Works and What the Patient Sees

An array of 96 electrodes was temporarily implanted into the subject's visual cortex. The device delivers electrical impulses directly to the cortex, bypassing the damaged optic nerve. As a result, the patient perceives not a full picture, but individual flashes of light — phosphenes — from which, according to the developers, a complete image can eventually be assembled. During the trial sessions, the man described the characteristics of the flashes he saw: their brightness, size, shape, and color. This subjective data served as feedback for the neural network, allowing it to precisely adjust the parameters of further stimulation.

### Brain Non-linearity: The Main Engineering Challenge

The key difficulty of the development, as the authors of the study emphasize, lies in the fact that the visual cortex reacts to electrical impulses non-linearly. Engineers aim to use each electrode as a 'pixel' on a virtual screen, but the brain is not a static display: its state is constantly changing, and neighboring electrodes interfere with each other, creating crosstalk. This is precisely why a fixed stimulation pattern calculated once does not work — the brain 'drifts' away from the set mode, and without continuous adaptation, image quality rapidly degrades.

### The Role of AI: On-the-Fly Adaptation

This is where deep learning comes in. Artificial intelligence continuously analyzes the individual brain reactions of a specific person and adjusts stimulation parameters in real time. According to the researchers, a useful visual prosthesis cannot operate on a pre-set template: every brain is unique, and the algorithm must account for its current state, fatigue, attention, and other variables. Such an approach, the authors estimate, makes it possible to obtain sharper visual images with lower energy costs than with classical implant control methods.

### Who the Technology Is For and What's Next

The developers emphasize that the technology will be a salvation for people who lost their vision in adulthood due to trauma, chronic diseases, or hereditary eye pathologies. At the current stage, the work is a successful proof of concept: it has confirmed that AI is capable of effectively managing the complex non-linear 'brain-implant' system and improving the quality of perceived images. Scientists point out that the results obtained will form the basis for future commercial bionic eyes, but significant work on miniaturization, implant durability, and expanding the number of electrodes to form a full image still lies ahead before a serial product and widespread clinical implementation, by their estimates.

## Fact-Check & Sources

- [ИИ научили управлять активностью человеческого мозга: как технология помогает видеть](https://www.rbc.ua/ukr/news/shi-navchili-keruvati-aktivnistyu-lyudskogo-1793971948.html) - Primary source: study in Neuron, international team (USA, Switzerland, Spain), 27-year-old patient, 96 electrodes, phosphenes, cortical non-linearity, proof of concept. All key facts of the article match the source text.

## FAQ

### What exactly does a patient see with such an implant?

The patient sees phosphenes — flashes of light and object outlines. This is not a full image, but basic visual elements from which a more complex picture is planned to be assembled in the future.

### How many electrodes are installed in the patient's brain?

An array of 96 electrodes was temporarily implanted into the visual cortex of the 27-year-old man.

### Why is artificial intelligence needed to control the implant?

The visual cortex reacts non-linearly to impulses: the brain constantly changes state, and neighboring electrodes influence each other. A fixed stimulation pattern does not work, so AI analyzes brain reactions in real time and adjusts stimulation parameters on the fly.

### When will the technology become available to patients?

At the current stage, the work is a proof of concept. Scientists indicate that the results will form the basis for future commercial bionic eyes, though specific timelines for clinical implementation are not named in the source.