Pavlov Was Right About the Dogs

The story everybody knows: Pavlov rang a bell, gave the dog food, repeated it, and eventually the bell alone produced salivation. The moral usually attached is that if you pair two things together often enough, an association forms.

That moral is wrong, and the experiment that shows it is wrong is one of the most elegant in the field.

Robert Rescorla ran a version in which the bell was paired with food exactly as usual, but food also arrived just as frequently at times when the bell had not sounded. Same number of pairings. No conditioning at all.

The animal is not counting coincidences. It is computing whether the bell tells it anything: whether food is more likely given the bell than without it. If the answer is no, the pairings are irrelevant, however many there are. Conditioning is learning about predictive relationships: something much closer to detecting a cause than to gluing two events together.

Leon Kamin's blocking effect makes the same point from another angle and is the sharpest evidence available. Train an animal that a light predicts a shock, until it is thoroughly learned. Then start presenting the light and a tone together, still followed by the shock. Test the tone alone.

The animal has learned almost nothing about the tone. The shock was already fully predicted by the light, so there was no surprise, and with no surprise there was nothing to learn.

That result is why modern accounts of conditioning are built on prediction error. The Rescorla-Wagner model says learning is proportional to how wrong you were: a fully expected outcome teaches nothing, an unexpected one teaches a lot. This turns out to be one of the most productive equations in psychology, and its signature was later found in the firing of dopamine neurons, which respond to reward better or worse than predicted rather than to reward as such. A behavioural model from 1972 predicted what a neuron would do twenty years before anybody recorded it.

The other thing the textbook story gets wrong is that conditioning is general-purpose, and John Garcia's work broke that assumption so thoroughly that his papers were initially rejected.

Rats made ill hours after eating a novel food will avoid that food afterwards: from a single pairing, across a delay of hours, where standard conditioning needs seconds. Worse for the general-purpose view, the learning is selective: rats associate illness with taste, and not with a light or a noise that was equally present. They will associate a light with a shock perfectly well, but not with nausea.

The learning system is prepared. It comes with assumptions about which kinds of causes go with which kinds of effects, and those assumptions are exactly the ones an animal that eats varied food would need. This is why conditioning does not explain the distribution of human phobias: people acquire fears of snakes, heights and spiders far more readily than of electrical sockets or cars, despite the latter being considerably more dangerous.

Which brings up Little Albert, taught in every introductory course and worth being clear-eyed about.

Watson and Rayner's 1920 study conditioned an infant to fear a white rat by striking a steel bar behind his head. It used one subject, no control condition, subjective outcome ratings made by the experimenters who knew the hypothesis, and the child was never deconditioned. Its historical importance is enormous and it establishes very little. It is taught because it shaped what the field thought it could do, not because the evidence is any good.

The last thing to know about conditioning is what happens when you undo it, and this is the part with the most practical weight.

Extinction, presenting the cue repeatedly with no outcome, does not erase the original learning. It builds new learning on top, of the form "in this context, this cue no longer predicts that". We know this because the old response keeps coming back: spontaneously after a delay, on returning to the original context where it was learned, and immediately after a single re-exposure to the original outcome.

For exposure therapy this is the central problem. A fear extinguished in a clinic is extinguished in a clinic, and the context specificity of extinction is a substantial part of why relapse happens. It is also why good exposure work varies the setting deliberately rather than doing everything in one comfortable room.

And for one clinical application the laboratory findings transfer almost directly. Patients receiving chemotherapy frequently develop anticipatory nausea, triggered by the room, the smell, or the journey: straightforward conditioning to contextual cues. One effective intervention exploits Garcia's selectivity: give an unusual, strongly flavoured sweet shortly before the infusion, and the aversion attaches preferentially to that novel taste rather than to the clinic or to the patient's ordinary diet. It is called the scapegoat technique, and it works because rats and saccharin told us which cue the system would pick.