The sensations produced by electrical stimulation of the visual cortex

TL;DR: In 1968 Brindley and Lewin wired 80 radio-driven electrodes onto the visual cortex of a blind woman and made her see spots of light at predictable places—write access to the human visual system, demonstrated once and then stalled for decades. The two things that killed it, too few channels and no model of what the cortex does with the current you inject, have both collapsed; the modern move is to treat the cortex as a system to be inverted with a learned encoder, and that is now a tractable engineering problem.

The idea, as they had it

The retina and optic nerve are one failure point in blindness; the visual cortex is another, downstream target. Brindley's bet was simple and audacious: if you can't fix the eye, stimulate the cortex directly and let the brain do the rest.

The 1968 implant was, for its time, a marvel of engineering. Eighty electrodes sat on the surface of the right occipital pole, each driven by its own radio receiver—no wires through the skull. An external transmitter, an array of oscillators the patient could activate, sent power and signal inductively. Turn on one channel, and she reported a "phosphene": a small, sharp spot of white light at a fixed position in her left visual field.

The results read like a founding document for the whole field:

• Retinotopy holds. Weak stimulation produced phosphenes whose positions matched the classical cortical maps derived from WWI head wounds (Holmes). The cortex is a map of visual space, and they were writing to it.
• Points are separable. Electrodes 2.4 mm apart gave distinguishable spots. Fire several at once and the patient saw predictable simple patterns.
• The percept is retinal-coordinate, not world-coordinate. During voluntary eye movements the phosphenes moved with the eyes; during vestibular (reflexive) eye movements they stayed fixed in space. Exactly what you'd expect if you're painting onto a retinotopic map.
• No flicker fusion. Cortical phosphenes don't fuse into steady light the way a flickering LED does—a hint that you are talking to the cortex in a language it doesn't natively use.

Point 9 is the thesis of the paper and of everything since: with a better prototype, this could be a useful prosthesis.

Why it could not work then

Three hard limits, in rough order of severity.

Channel count and electrode scale. Eighty surface electrodes at ~2.4 mm pitch. Surface (epicortical) electrodes sit on the cortex and recruit large, poorly controlled populations of neurons, so each phosphene is coarse and the currents needed are large (milliamps), which spread and can trigger pain or seizures. For comparison, a usable low-resolution image—say, reading large print—wants on the order of hundreds to thousands of well-separated pixels. Cortical magnification helps (the central few degrees of vision occupy a disproportionately large slab of V1), but 80 blobs is closer to a status light than a display.

No model of the cortical code. This is the deeper limit, and the one Brindley couldn't have solved with any hardware. He implicitly assumed phosphenes are like pixels: independent, additive, stable. They are not. Fire two nearby electrodes and you do not reliably get two dots; you get a merged smear whose brightness, size, and even position depend nonlinearly on current, frequency, and history. His own observations—no flicker fusion, phosphenes persisting up to two minutes after strong stimulation—are symptoms of a system with rich temporal dynamics that a static pixel map ignores.

Bulky, biocompatible-limited hardware. The radio-receiver array was ingenious but large, and long-term biocompatibility of chronic cortical implants was barely understood. Materials, hermetic sealing, and power delivery all limited how many channels you could realistically pack and keep alive.

The first two are the ones that matter, and both have moved by orders of magnitude.

What changed

Electrodes. Penetrating microelectrode arrays—the Utah array (roughly 96–100 shanks at ~400 μm pitch, tips inside cortex) and, more aggressively, Neuralink-class systems with on the order of a thousand contacts on flexible threads—put you far closer to individual columns, at microamp currents. Penetrating means less current, sharper phosphenes, and more channels in the same tissue volume.

A working theory of the code—and current steering. The pivotal modern result is Beauchamp, Yoshor and colleagues (Cell, 2020): instead of poking static dots, they swept stimulation across the electrode array in a trajectory, "drawing" shapes on cortex the way you'd trace a letter with a finger. Sighted and blind subjects could read letters (W, S, Z, etc.) from these dynamic patterns—something static multi-electrode stimulation had never reliably achieved. The lesson is exactly the one Brindley's flicker and persistence data hinted at: the cortex responds to spatiotemporal structure, so you should stimulate in trajectories, not snapshots.

Compute and generative models of visual cortex. We now have image-computable models of V1—from classic Gabor/energy models up to deep networks fit to neural data—that predict population responses to arbitrary images. Once you have a differentiable forward model r = f(s) mapping a stimulation pattern s to a predicted neural/perceptual response r, you can invert it: find the s that produces the response corresponding to a desired image. That is the piece Brindley was missing entirely.

What a 2026 revival looks like

Reuse Brindley's core findings—retinotopy, retinal-coordinate percepts, separability of nearby sites—and his framing that this is fundamentally a write-access problem. Replace the pixel assumption with a digital twin and a learned encoder.

The architecture:

1. Forward model (the twin). During an initial mapping session, stimulate systematically—single sites, pairs, trajectories, varied current/frequency—and record what the patient reports (position, size, brightness, merging). Fit a differentiable model f_\theta that predicts the perceived image from a stimulation pattern s, including the nonlinear interactions and temporal dynamics Brindley saw.
2. Inverse encoder. Train a network g_\phi that maps a target image x to a stimulation pattern s = g_\phi(x) by minimizing a perceptual loss \lVert f_\theta(g_\phi(x)) - x\rVert under safety constraints (charge per phase, total current). This is the analogue of the "encoder" that drives every retinal-prosthesis-quality proposal, but grounded in the patient's own cortex.
3. Trajectory-native output. Because dynamic sweeps beat static dots, s should be a spatiotemporal current program, not a frame. Beauchamp's letter-tracing becomes a special case the encoder discovers.
4. Closed-loop calibration. Phosphene maps drift; the twin must be re-fit online from sparse patient feedback (does this look more like the target?), which is a small active-learning problem, not a full remapping.

The honest hard part is that the "ground truth" is subjective report, which is slow and noisy—you cannot backprop through a person. So the twin is really a psychophysical model fit from thousands of trials, and the whole loop is bottlenecked by how fast a patient can report percepts. That, not electrode count, may be the 2026 limit.

Has it been vindicated?

Partly, and the trajectory is clear. The lineage runs Brindley → Dobelle's cortical implants (1970s–2000s, famously letting a blind man navigate crudely) → Second Sight's Orion cortical device and the Argus II retinal implant → the Utah-array intracortical work (Fernández et al. restoring crude form perception in a blind person) → Beauchamp/Yoshor's dynamic current steering. The specific "learned encoder inverting a cortical model" step is well established in the retinal prosthesis literature (encoder models that reproduce ganglion-cell spike patterns) and is the obvious next move cortically; I'm not aware of a full patient-specific cortical digital-twin-plus-encoder deployed at scale yet, so treat that as the open frontier rather than a done deal.

What's still open: getting to hundreds of simultaneously usable channels without cross-talk; long-term stability of penetrating arrays in cortex; whether V1 forward models fit to macaque recordings transfer to a human patient's idiosyncratic phosphene map; and the report-bandwidth bottleneck above. None of these is the kind of wall Brindley hit—they're engineering and modeling problems with active work on each.

Where to read

The paper: Brindley & Lewin, "The sensations produced by electrical stimulation of the visual cortex," J. Physiol. 196 (1968) — doi:10.1113/jphysiol.1968.sp008519. Bibliographic details verified in OpenAlex; the specific figures I quote (80 electrodes, 2.4 mm spacing) are from the abstract and are reliable, but I have not re-checked every number against the full text.

Read alongside: Beauchamp et al., "Dynamic Stimulation of Visual Cortex Produces Form Vision in Sighted and Blind Humans," Cell (2020), for the trajectory insight; Fernández et al. (2021) on intracortical Utah-array vision restoration; and the retinal-prosthesis encoder work (e.g. Nirenberg & Pandarinath) for the clearest existing example of inverting a neural code into a stimulation program. Together they show Brindley's 1968 conclusion—"it will be possible, by improving our prototype"—was right, and that the improvement was mostly software all along.