Backpropagation worked in machines for forty years before biology figured out if it works in brains.

Rumelhart, Hinton, and Williams published the canonical backprop paper in Nature in 1986. The algorithm trained perceptrons, then deep networks, then transformers. It became the dominant learning algorithm of modern AI. The whole field rests on it.

Through that entire period, neuroscience said backprop almost certainly does not happen in the brain. The weight transport problem alone seemed disqualifying. A synapse making a weight update needs to know its weight, the postsynaptic error signal, and the symmetric weights on the return path. Biology has no clean mechanism for the symmetric return path. Hinton himself has said publicly for decades that backprop is unrealistic as a brain algorithm.

The standard story would have run the other way. Biology discovers a learning rule, ML adapts it. That is the story of Hebbian learning, of Hopfield networks, of spiking neural networks.

Backprop ran in reverse. It worked in silicon first. Then, slowly, neuroscientists started looking for biologically plausible approximations that could produce backprop-like learning under realistic synaptic constraints. Predictive coding networks. Target propagation. Equilibrium propagation. Each preserves the result that gradient information flows backward through a hierarchy, while obeying biological constraints the textbook algorithm violates.

The current best guess is that brains implement something approximately backprop-like, derived from completely different mechanisms. Engineering pulled biology toward an answer biology had not yet thought to look for.

The arrow of inspiration does not have to point from carbon to silicon. Sometimes the engineering arrives first.