The setup is simple: alternate between two grating images for a few minutes. Green horizontal stripes. Magenta vertical stripes. Nothing you would encounter in a natural environment. Then look at a black-and-white grating. It appears faintly pink if the stripes are horizontal, faintly green if vertical. The colors from the induction images have been transferred selectively to the orientation information in the test image — you're seeing a color that isn't there, contingent on the angle of the edges.
This is the McCollough effect, named for Celeste McCollough who published it in 1965. It has been studied for sixty years. The mechanism is still debated.
The duration is what doesn't fit. Standard color adaptation — the greenish cast you see after staring at a red scene — disappears in seconds. The McCollough effect, induced with fifteen minutes of viewing, persists for months. Jones and Holding (1975) tested subjects 85 days after a fifteen-minute induction and found the effect still present at better than half its original strength. Eighty-five days. From fifteen minutes of looking at some striped images.
There's a second strange property: the color tag is stored in retinal coordinates, not world coordinates. If you tilt your head 90 degrees while looking at the test grating, the color assignments reverse — what was pink becomes green, and vice versa. The color isn't attached to "horizontal stripes in the room." It's attached to stripes that fall on the horizontal axis of your retina. The visual system hasn't bothered to correct for the head rotation. Whatever holds the tag doesn't know about head position.
This places the storage early. Cells that are tuned to both orientation and color, but haven't yet integrated vestibular or proprioceptive signals about head position — V1, primarily monocular. The effect also appears to be stored largely in the adapted eye: induction in the left eye doesn't reliably produce the effect when you test with only the right. The memory is upstream of where the two eyes converge.
McCollough's original explanation was that these cells are part of a chromatic aberration correction system. The eye's lens isn't perfect. Light of different wavelengths focuses at slightly different distances, creating colored fringes on high-contrast edges — horizontal edges get one pattern of fringing, vertical edges get the complementary pattern. If you've always lived in a world of white light, the visual cortex could learn over time to subtract these fringes by associating the complementary color with each oriented edge class. You'd never see the calibration because it's successfully removing what it learned.
The months-long persistence makes sense under this account: the lens doesn't change much over months, so the correction only needs to update slowly. You're carrying a calibration that was learned over years, finely tuned to your specific optics.
The McCollough experiment induces a false calibration. It shows unnaturally colored gratings with enough intensity that the correction system treats them as evidence about the lens. The system updates: "my lens is apparently adding green to horizontal edges, so I should perceive horizontal edges as pinker to compensate." Afterward, the correction runs on ordinary gratings and over-corrects — the edges that were already neutral get a color added.
What this means: the effect that looks like a strange laboratory artifact is actually a test that makes visible a correction that's running all the time in every visual system. The McCollough effect isn't the aberration. It's the calibration made detectable by being driven to a wrong value.
One more property: the effect weakens slightly each time you test it. Viewing the test grating is itself a kind of exposure — you're seeing black-and-white horizontal gratings without any colored reinforcement, which runs the correction in reverse. Repeated testing is silent counterconditioning. Jones and Holding found this too: subjects who were retested more frequently showed faster decay. The measurement interacts with what it's measuring. To read the value, you have to partially change it.
The mechanism remains genuinely unclear. A 2008 fMRI study found two adaptation timescales in early visual cortex: a fast component (decaying with a time constant of roughly 30 seconds) and a second component that showed no measurable decay at all — a permanent integrator. The months-long persistence probably lives in that second component. What makes a synaptic change permanent while another decays is not well understood. The effect is too persistent to be ordinary adaptation, and too early and too local to be clearly "learning" in the way memory researchers mean. It sits in a category that neither framework fully owns.
The version I keep returning to: somewhere in your primary visual cortex, there are cells that have been maintaining a calibration record for your specific eyes since you were an infant. Not accessible to introspection. Not something you experience as memory. Running quietly, subtracting distortions you've never seen, holding an adjustment that's been accurate for so long the update rate has slowed to almost nothing. The McCollough experiment briefly overwrites a small part of that record and makes the overwrite visible as an illusion. Sixty years of study hasn't settled how it works. That fact alone tells you something about how far outside ordinary experience this level of processing is.