Skinner’s wartime experiment

In 1943 B.F. Skinner ran an unlikely wartime project that began with birds. Frustrated by the difficulty of steering missiles with then-available technology, Skinner turned to animal behavior. After finding crows unwieldy, he trained ordinary pigeons to peck targets on projected images and planned to place them in the nose cone of a missile to guide it by pecking. The military never used the so-called Project Pigeon, but the experiments convinced Skinner that pigeons were reliable instruments for studying learning.

From behaviorism to reinforcement learning

Skinner expanded on Ivan Pavlov’s classical conditioning by developing operant conditioning, the idea that behavior is shaped by consequences. Actions that lead to reward are reinforced and therefore more likely to recur. Skinner taught pigeons to perform chains of actions by rewarding successive approximations of the desired response.

Although behaviorism fell out of favor in psychology by the 1960s, its basic mechanisms found new life in computer science. Richard Sutton and Andrew Barto translated ideas from animal learning into reinforcement learning, a formal framework in which agents search for actions and store associations between situations and rewards. That framework later became foundational for many modern AI breakthroughs.

Pigeons in the era of large compute

Early AI researchers had tried to model human reasoning directly, producing symbolic systems that struggled with perception and pattern recognition. Pigeon studies suggested an alternative: complex behaviors can emerge from simple associative rules, learned through repeated trial and error rather than by explicit rules.

A striking 1964 study showed pigeons could learn to discriminate photographs with people from those without people after being rewarded for correct pecks. They learned despite partial occlusion and visual noise, demonstrating that associative learning could support surprisingly flexible classification.

As computing power surged, reinforcement learning scaled up. Systems trained to maximize simple numeric rewards mastered games, navigation, and other tasks. AlphaGo Zero is a prominent example: starting tabula rasa and using only reinforcement signals, it rediscovered centuries of Go knowledge and developed novel strategies within weeks. The same core idea drives many modern agents that rely on search and memory rather than modeled human reasoning.

Rethinking animal intelligence

The success of reinforcement learning in machines has spurred biologists and comparative psychologists to reevaluate associative learning in animals. Johan Lind and Ed Wasserman argue that what was once dismissed as a crude mechanism may explain more complex animal behaviors than previously thought. Wasserman’s experiments, for example, trained pigeons to detect medical anomalies in scans at a level comparable to trained physicians, and pigeons learned subtle categorization tasks that human students struggled with.

These results do not deny the roles of instinct, emotion, or species-specific adaptations. Rather, they suggest that associative learning is a powerful cognitive tool across species and that some impressive behaviors attributed to higher cognition could arise from chains of reinforced associations.

Ethical and philosophical implications

Drawing parallels between pigeons and AI invites uncomfortable questions. Reinforcement learning shows that very simple associative mechanisms can produce behavior that looks intelligent, both in animals and in machines. But similarity in mechanisms does not imply equivalence of experience. Pigeons are sentient creatures that can feel pain and suffer, while current AI lacks subjective experience.

The parallels also highlight research priorities. Some philosophers and scientists call for more investment in animal cognition research to better understand sentience and to avoid conflating competent performance with true subjective awareness. The same associative principles that power modern AI could help us better understand how minds evolved and how to attribute cognition and moral consideration across species.

A humbling lesson

If artificial intelligence advances by scaling up pigeonlike associative rules, then our intuition that humanlike cognition is the natural model for machines is misleading. Machines may reach superhuman performance by amplifying simple learning principles rather than by emulating human reasoning. That is a humbling and clarifying lesson: the pigeon brain, modest and practical, has shaped both our machines and our understanding of intelligence.