AI Guides 96-Electrode Brain Implant in a Blind Man’s Visual Cortex
AI picked stimulation patterns for a blind man's brain implant that hit target neural activity with less current, a study in one man found.
A deep-learning system chose electrical stimulation patterns for a 96-electrode implant in the visual cortex of a blind man, and the patterns it picked produced the intended brain activity more accurately than comparison methods while using less current. The proof-of-concept study came from teams at UC Santa Barbara, ETH Zurich and Miguel Hernández University and was published online in the journal Neuron (DOI 10.1016/j.neuron.2026.07.006), according to ScienceDaily and UC Santa Barbara.
The man’s visual experiences stayed limited to flashes and shapes of light. The technology is experimental, and the results come from a single participant.
How did researchers test the AI?
In 2024, researchers from Michael Beyeler’s Bionic Vision Lab at UC Santa Barbara traveled to Hospital IMED Elche in Spain. They worked with a 27-year-old man who lost his sight after a traumatic brain injury. He had an implant with 96 electrodes in his visual cortex, the part of the brain that processes what the eyes send. The device was meant to stay in for six months and then be removed.
When the electrodes fire, the participant sees phosphenes, which are small dots, bursts or outlines of light. Producing them is possible. Making them consistent is harder, because identical stimulation does not always give the same response.
The team trained a deep neural network to predict how the cortex would respond to different stimulation settings. The model also used the brain’s resting activity just before each pulse as a snapshot of its current state. From those predictions, the researchers picked combinations of signals most likely to create a chosen pattern of activity. The implant could both stimulate and record, so the model learned from real neural responses. The man also reported whether he saw a phosphene and, in some experiments, its shape, size, brightness and color.
What did the results show?
The AI-selected stimulation hit the intended patterns of neural activity more accurately than the comparison methods, and it needed less electrical current. Recorded brain activity also predicted what the man saw better than the stimulation settings did.
That means making the brain do something and knowing what a person will see are separate problems. Beyeler said engineers tend to treat phosphenes like pixels, but electrodes interact and neural responses fluctuate, so adding electrodes does not simply build a fuller picture.
Who might a cortical prosthesis help?
Some experimental devices stimulate the retina, and others target parts of the visual pathway farther along. Cortical prostheses skip the eyes and optic nerves and send signals straight to the visual cortex. That approach could suit people whose visual pathways are damaged but whose cortex still responds, such as those who lost sight to strokes, neurodegenerative disease or traumatic brain injury. Beyeler said people who could see earlier in life may want some vision back very strongly.
What is still unclear?
The study involved one participant with a temporary implant, part of a larger feasibility trial in Spain. The findings show better communication between an implant and the brain, not restored useful sight. The paper was led by co-first authors Pehuén Moure of ETH Zurich, Jacob Granley of UC Santa Barbara and Fabrizio Grani of Miguel Hernández University, with Beyeler, Shih-Chii Liu of ETH Zurich and Eduardo Fernández of Miguel Hernández University supervising.
Why it matters for vision prostheses
The brain’s response to electrical signals can change over time, including from one day to the next. The AI accounted for that by using resting activity, which suggests future prostheses may need to monitor the brain and adjust stimulation as they go instead of relying on fixed settings. According to Beyeler, a useful device has to learn how an individual brain responds.
Source: ScienceDaily
