An animal adjusts its behaviour based on experience—that is, it learns—when experience at one time provides information that will be useful at a later time. Viewed in this light, learning is seen as a tool for survival and reproduction because it helps an animal to adjust its behaviour to the particular state of its environment. An animal needs to know such things as what food is good to eat, when and where to find it, whom to avoid and approach, with whom to mate, and how to find its way home. When these things are not genetically preprogrammed—because they depend on the particular circumstances of an individual’s time and place—the animal must learn them.

Consider, for example, a female digger wasp called the bee wolf (Philanthus triangulum) who has finished excavating a tunnel in a sandy bank. She then digs a small outpocket where one of her young will develop, and she stocks this cell with worker honeybees (Apis mellifera), which she has paralyzed by stinging and which will serve to provision her young. After laying an egg on one of the bees, she closes off the cell with sand and starts work on a new cell. To provision the cell, she must fly out to hunt more honeybees; however, after crawling out of her nest burrow, closing its entrance hole, and launching into flight, she does not immediately depart the area. Instead, she hovers just over her nest site, inspecting the ground and flying in wider and wider arcs to scan an ever-increasing area. During this elaborate departure flight, the wasp memorizes the specific configuration of landmarks—sticks, tufts of grass, and trees—around her burrow. Later, when she returns, she will use the information to pinpoint her nest’s location. Her genes cannot provide her with knowledge of the landmark array around her nest, so she must learn it.

One of the clearest indications of the falseness of the old dichotomy between innate and learned behaviour is the fact that in most cases animals are genetically predisposed to acquire only specific information in developing their behaviour. One might say that most of the learning performed by animals is instinctive learning. This phenomenon is conspicuous in the flower-learning behaviour of honeybees (A. mellifera). Since at least the time of the Greek philosopher and scientist Aristotle (384–322 bce), it has been known that worker bees show “flower constancy,” a specialization by individual bees on a single species of flower. Flower constancy occurs in spite of the fact that honeybees are generalist foragers capable of exploiting many flower species. The flowers have much to gain from bees that remain faithful to them; specialist bees will be carrying the appropriate species of pollen. Therefore, the colours and odours of flowers probably evolved as conspicuous signals for the bees to learn. In turn, specialization benefits the bees by reducing flower-handling time and facilitating the collection of nectar.

Early in the 20th century, Austrian biologist Karl von Frisch demonstrated experimentally that honeybees are able to learn and distinguish a single floral odour from among at least 700 others. In addition, he found that they could distinguish colour from yellow into the ultraviolet across the electromagnetic spectrum. One striking feature of this type of colour and odour learning is the rigid programming of the timing. Research has revealed that a bee learns the flower’s colour only during the final few seconds before beginning to feed, and odour learning occurs during feeding. It is as if bees possess a set of switches that turn colour and odour learning on and off at specific times during the foraging process. The time course of this learning program is highly adaptive, being restricted to times when a bee is alighted on a rewarding flower. In this manner, its learning is focused on the colour and odour of the flowers of this rewarding species rather than on the hues and scents of any nearby flowers of unrewarding species.

Is this machinelike learning of bees fundamentally different from the learning processes in vertebrates? Until the mid-1960s, psychologists generally believed so. Studying mainly birds and mammals, they developed an approach known as “general process learning theory,” which attempted to account for learning with a single set of principles, namely unconstrained “associative learning” as studied in instrumental (operant) conditioning and classical (Pavlovian) conditioning. Associative learning is said to occur when an animal changes its behaviour upon forming an association between an environmental event and its own response to the event. In operant conditioning, the animal learns to associate a voluntary activity with specific consequences. In classical conditioning, the animal learns to associate a novel (conditioned) stimulus with a familiar (unconditioned) one. For example, in his study of classical conditioning, Russian physiologist Ivan Petrovich Pavlov demonstrated that by consistently exposing a dog to a particular sound (novel stimulus) and simultaneously placing meat powder (familiar stimulus) in its mouth the dog could be made to salivate upon hearing the sound even without the meat stimulus. Initially, salivation was the unconditioned response, whereas the food stimulus was the unconditioned stimulus. Once the dog learned to associate the sound stimulus with the food stimulus, salivation became the conditioned stimulus to sound—that is, a stimulus that previously did not trigger a response.

The popularity of general process learning theory peaked in the 1940s and ’50s. In the mid-1960s, however, American psychologist John Garcia discovered several puzzling phenomena that indicated adaptive limits on learning and contradicted the supposedly general principles of conditioning. One of the most important of these anomalies was flavour aversion learning. When rats (Rattus norvegicus) and many other vertebrates, including humans, sample a flavour and later become ill, they learn to avoid consuming that flavour in the future. This phenomenon has two remarkable properties. First, it occurs despite delays of several hours between experiencing the flavour (the conditioned stimulus, or CS, in the Pavlovian conditioning paradigm) and experiencing the illness (the unconditioned stimulus, or US); it does not require the brief delay specified by the general principles of conditioning. Second, in rats, learning with the US being illness is limited to flavours. This response was revealed in an experiment in which rats experienced a flash of light and the sound of a buzzer each time they took a drink from a tube of flavoured water (hence “bright noisy tasty water” became the CS). Some of the rats were made ill (nauseous) after drinking (hence illness became the US for them), whereas others were shocked through the feet shortly after they began drinking (hence pain became the US for them). After conditioning, the rats were tested with the noise plus the light alone or with the flavour alone. Those rats that had been made ill avoided drinking only the “tasty water,” whereas the rats that had been shocked avoided drinking only the “bright noisy water.” In other words, the rats could learn to associate a taste with an illness but not a visual and auditory stimulus. Conversely, the rats could learn to associate a visual and auditory stimulus, but not a taste, with pain.

These findings attracted tremendous skepticism when they were first reported because both the long delay between CS and US and the CS-US specificity contradicted the idea of general laws of learning. Both findings, however, make considerable sense in light of the problems faced by rats living in nature. If they consume a new food and become ill even hours later, they will not eat the food again and thus not suffer the illness associated with the food. Moreover, it is adaptive that rats learn to associate a taste cue, not an auditory or visual cue, with illness-causing food because rats discriminate foods best using chemical cues sensed by taste, olfaction, or both. In contrast, something that causes pain is best recognized from a safe distance. Therefore, it is adaptive that rats learn to associate auditory and visual cues with painful experiences. Thus, these “anomalies” for general process learning theory can be understood by considering the functions that the rats’ learning has evolved to serve.

There is now compelling evidence that humans also possess adaptive predispositions in learning abilities. Consider, for example, the curious anthropological discovery made in 1926 by Finnish sociologist Edward Westermarck that arranged marriages between children that grow up together (whether biological siblings or not) are far more likely to fail than arranged marriages between individuals not raised together. The failures most often result from sexual incompatibilities. Evidently, children are genetically guided to learn to treat as siblings all individuals with whom they are raised together. And because siblings tend to avoid sexual contact, presumably due to a long evolutionary history of detrimental consequences associated with inbreeding, marriages between these individuals tend to fail.

Today it is widely recognized that the general-purpose psychological approach to learning had overlooked its biological significance and that animals possess learning mechanisms that are specialized for solving the problems they face in the natural world. This view of learning explains the psychologists’ observations of the limits of learning by animals in laboratory settings. It also makes sense of ethological reports of special forms of learning, such as imprinting (that is, the rapid identification of parents by newborn animals triggered by following the first object they see moving away from them), which have been studied in naturalistic settings. To a large extent, this picture of instinctive learning has brought a constructive end to the centuries-old debate about whether “nature” (genes) or “nurture” (experiences) is the source of adaptive behaviour of animals. Animals are shaped by their experiences; however, the interpretation of each experience is governed by a collection of rules (Darwinian algorithms) set by the genes in each species.

The general-purpose view of learning that prevailed during most of the 20th century was based on two assumptions: (1) the ability to learn is always beneficial, and (2) animal learning abilities are like human learning abilities, which seem to be of completely general and unlimited applicability. Neither assumption is correct.

First, there are costs as well as benefits to learning, so learning abilities will be beneficial, and favoured by natural selection, only when the benefits outweigh the costs. The costs include those involved in building and maintaining the required neural circuitry and also the time and mistakes involved in learning while the animal is fine-tuning its behaviour to the current or likely future state of its environment. When learning is a matter of life or death—as in geese (Anser and Branta), sheep (Ovis), and antelopes (family Antilopinae), where newborn young must keep up with mobile parents—the advantage of rapid learning (that is, staying together) and the danger of slow learning (that is, lagging behind) are both extremely high. By considering both the fitness costs and the benefits of different forms of learning, one can readily appreciate the reasons why imprinting occurs in these species, rather than the slower process of trial-and-error learning.

Second, as described earlier, the learning abilities of animals, including humans, are not completely general; learning abilities are adaptively specialized so that, in any particular context, animals take in only the most relevant information. Late in his career, Lorenz referred to “the innate schoolmarm,” a phrase that picturesquely expresses the reality that animals possess adaptive predispositions in their learning.