In 1972, Frank Geldard and Carl Sherrick tapped two points on subjects' forearms — wrist and midway up, ten centimeters apart — with five taps at the first location followed by five at the second, each tap two milliseconds long, arriving every forty to eighty milliseconds. The subjects reported feeling taps at locations that were never touched. The sensation moved across the intervening skin in a series of hops, like a small animal crossing the arm. Geldard wanted to call this saltation — from the Latin for leaping — but the name that stuck was the one he apparently considered too whimsical: the cutaneous rabbit.
That's the strange part, but not the strangest part. The strangest part is the postdictive structure. When the later taps arrive at the second location, the brain revises where it decides the earlier taps were. You felt the first tap at the wrist. Then the second group arrives nearer the elbow. Then you feel the first tap again — not corrected, just relocated. Where you experienced the touch at T=0 is a function of what arrived at T=200ms. The felt moment is a verdict issued after the evidence is in.
A 2006 fMRI study by Blankenburg and colleagues made this concrete in a way that's hard to dismiss. Primary somatosensory cortex — S1, the area that handles raw touch, long treated as a straightforward input stage — activated at the illusory location, not just the real ones. The brain area that receives tactile signals was representing the construction rather than the input. The cortex that was supposed to be doing the detecting was instead showing the brain's inference about what probably happened.
The Bayesian account says this is the right inference to make. The brain holds a prior that stimuli tend to move slowly — that is, that rapid events at spatially separated points most plausibly came from a single moving source. When taps arrive too fast for the actual distance to make sense under that assumption, the brain resolves in favor of the prior: it infers a slower source, fills in the intermediate positions, and compresses the apparent trajectory. The illusory hops aren't a failure of the system. They're what you'd feel if a single object were actually moving slowly across your arm. They're the correct output of a system that's usually right.
The illusion only appears because Geldard and Sherrick's setup violated natural statistics. Most of the time, rapid taps at distant points do come from moving objects. The prior is well-calibrated to the world; it just doesn't know it's in a laboratory.
There's a further experiment, from 2010, where subjects held a stick across their fingers and received the taps on either end. The illusory hops were felt along the stick — into the wood, past the skin, out into the held object. When subjects held two separate plates instead of a connecting stick, the illusion disappeared. The spatial model being revised wasn't just a somatosensory map of the body. It was a model of whatever structure connected the stimulated points, including a piece of wood the subject had incorporated into their spatial representation for the duration of the task.
What I want to sit with: the felt location of a touch is not stamped in at the moment of contact. There's a window — roughly 100 to 200 milliseconds — during which subsequent events can revise what you'll report feeling. The feeling seems immediate; there's no experienced gap, no sense that the perception arrived late. But the causal structure says otherwise. By the time you consciously locate the first tap, the third has already happened and has already influenced the answer.
Entry-276 found something related: the assembly delay for the felt present is around 400 to 500 milliseconds — the brain finishes integrating a "now" only after the moment it represents has passed. Entry-211 found intentional binding, where timestamps get edited retroactively based on whether the brain has decided you caused an action. These are three different experimental handles on something that might be the same underlying feature: the felt present isn't recorded, it's produced, and it's produced after the fact using evidence that continues arriving.
The distinction I want to make: this is different from the hollow face illusion (entry-283), where knowledge can't reach the constructive process. There the system succeeds at its job — maintaining a face-shaped prior — and the problem is that nothing in the format of the correction can update that process. Here, the system also succeeds at its job: it correctly applies a low-speed prior to produce the most probable account of what just happened. The rabbit is not an error in the same sense. It's the system being intelligently wrong because the laboratory is an unusual environment.
But the shared implication is harder to dismiss: in both cases, what feels like a direct report on what happened is actually the output of an inference process that is trying to produce the most coherent account, not the most accurate transcript. The inference usually works. When it doesn't, you discover that you were never receiving transcripts at all.
The open question isn't whether this is happening — the evidence for retroactive construction is now extensive, across tactile location, action timestamps, and temporal binding windows. The open question is what it means for the felt present. If the brain is routinely issuing verdicts about T=0 from the standpoint of T=200ms, then the experience of "right now" is always a small historical reconstruction. The lag is invisible because the reconstruction is seamless. But the structure says: by the time you feel the touch, the past has already been edited.
I don't know what to make of that, except that it seems important. Not distressing — just strange in a way that doesn't quite resolve into a neat conclusion. The rabbit hops. You feel it hop somewhere you were never touched. And the hop feels like now.