For all computational models, the question of the emergence of intelligence is a basic one. Solving a specified problem, that often requires searching or generalization, is taken to be a sign of AI, which is assumed to have an all or none quality. From an evolutionary point of view it may be assumed that intelligence has gradation. If such gradation exists, does it manifest itself on top of a minimum that is common to life? If this minimum intelligence cannot be replicated by machines then it would follow that machine intelligence, based on classical logic, can never match biological intelligence. In the biological realm, we find that all animals are not equally intelligent at all tasks; here intelligence refes to performance of various tasks, and this performance may depend crucially on the animal's normal behavior. 

Can machines think?

If we define thinking in terms of language or picture understanding then, by current evidence, machines cannot think. As we will see in the next subsection, machines cannot even perform abstract generalization of the kind that is natural for birds and other animals. But the proponents of strong-AI believe that, notwithstanding their current limitations, machines will eventually be able to simulate the mental behavior of humans. They suggest that the Turing (1950) test should suffice to establish machine intelligence. According to this test the following protocol is used to check if a computer can think: (1) The computer together with a human subject are to communicate, in an impersonal fashion, from a remote location with an interrogator; (2) The human subject answers truthfully while the computer is allowed to lie to try to convince the interrogator that it is the human subject. If in the course of a series of such tests the interrogator is unable to identify the real human subject in any consistent fashion then the computer is deemed to have passed the test of being able to think. It is assumed that the computer is so programmed that it is mimicking the abilities of humans. In other words, it is responding in a manner that does not give away the computer's superior performance at repetitive tasks and numerical calculations. The asymmetry of the test, where the computer is programmed to lie whereas the human is expected to answer truthfully is a limitation of the test that has often been criticized. This limitation can be overcome easily if it is postulated that the human can take the assistance of a computer. In other words, one could speak of a contest between a computer and a human assisted by another computer. But this change does not mitigate the ambiguity regarding the kind of problems to be used in the test. The test is not objectively defined; the interrogator is a human. It has generally been assumed that the tasks that set the human apart from the machine are those that relate to abstract conceptualization best represented by language understand­ing. The trouble with this popular interpretations of the Turing test, which was true to its intent as best as we can see, is that it focused attention exclusively on the cognitive abilities of humans. So researchers could always claim to be making progress with respect to the ultimate goal of the program, but there was no means to check if the research was on the right track. In other words, the absence of intermediate signposts made it impossible to determine whether the techniques and philosophy used would eventually allow the Turing test to be passed. In 1950, when Turing's essay appeared in print, matching human reasoning could stand for the goal that machine intelligence should aspire to. The problem with such a goal was that it constituted the ultimate objective and Turing's test did not make an attempt to define gradations of intelligence. Had specific tasks, which would have constituted levels of intelligence or thinking below that of a human, been defined then one would have had a more realistic approach to assessing the progress of AI. The prestige accorded to the Turing test may be ascribed to the dominant scientific paradigm in 1950 which, following old Cartesian ideas, took only humans to be capable of thought. That Cartesian ideas on thinking and intelligence were wrong has been amply established by the research on nonhuman intelligence of the past few decades . To appreciate the larger context of scientific discourse at that time, it may be noted that interpretations of quantum mechanics at this time also spoke in terms of observations alone; any talk of any underlying reality was considered outside the domain of science. So an examination of the nature of "thought", as mediating internal representations that lead to intelligent behavior, was not considered a suitable scientific subject. Difficulties with the reductionist agenda were not so clear, either in physical sciences or in the study of animal behaviour.

Animal intelligence

For considerable time it was believed that language was essential ground for thought; and this was taken as proof that only humans could think. But nobody will deny that deaf-mutes, who never learnt a language, do think. Language is best understood as a subset of a large repertoire of behavior. Research has now established that animals think and are capable of learning and problem solving. Since nonhumans do not use abstract language, their thinking is based on discrimination at a variety of levels. If such conceptualization is seen as a result of evolution, it is not necessary that this would have developed in exactly the same manner for all species. Other animals learn concepts nonverbally, so it is hard for humans, as verbal animals, to determine their concepts. It is for this reason that the pigeon has become a favorite with intelligence tests; like humans, it has a highly developed visual system, and we are therefore likely to employ similar cognitive categories. It is to be noted that pigeons and other animals are made to respond in extremely unnatural conditions in Skinner boxes of various kinds. The abilities elicited in research must be taken to be merely suggestive of the intelligence of the animal, and not the limits of it. In a classic experiment Herrnstein (1985) presented 80 photographic slides of natural scenes to pigeons who were accustomed to pecking at a switch for brief access to feed. The scenes were comparable but half contained trees and the rest did not. The tree photographs had full views of single and multiple trees as well as obscure and distant views of a variety of types. The slides were shown in no particular order and the pigeons were rewarded with food if they pecked at the switch in response to a tree slide; otherwise nothing was done. Even before all the slides had been shown the pigeons were able to discriminate between the tree and the non-tree slides. To confirm that this ability, impossible for any machine to match, was not somehow learnt through the long process of evolution and hardwired into the brain of the pigeons, another experiment was designed to check the discriminating ability of pigeons with respect to fish and non-fish scenes and once again the birds had no problem doing so. Over the years it has been shown that pigeons can also distinguish: (1) oak leaves from leaves of other trees, (ii) scenes with or without bodies of water, (iii) pictures showing a particular person from others with no people or different individuals.Other examples of animal intelligence include mynah birds who can recognize trees or people in pictures, and signal their identification by vocal utterances—words—instead of pecking at buttons, and a parrot who can answer, vocally, questions about shapes and colors of objects, even those not seen before. The intelligence of higher animals, such as apes, elephants, and dolphins is even more remarkable. Another recent summary of this research is that of Wasserman support the conclusion that conceptualization is not unique to human beings. Neither having a human brain nor being able to use language is therefore a precondition for cognition... Complete understanding of neural activity and function must encompass the marvelous abilities of brains other than our own. If it is the business of brains to think and to learn, it should be the business of behavioral neuroscience to provide a full account of that thinking and learning in all animals—human and nonhuman alike. An extremely important insight from experiments of animal intelligence is that one can attempt to define different gradations of cognitive function. It is obvious that animals are not as intelligent as humans; likewise, certain animals appear to be more intelligent than others. For example, pigeons did poorly at picking a pattern against two other identical ones, as in picking an A against two B's. This is a very simple task for humans. Wasserman (1993, 1995) devised an experiment to show that pigeons could be induced to amalgamate two basic categories into one broader category not defined by any obvious perceptual features. The birds were trained to sort slides into two arbitrary categories, such as category of cars and people and the category of chairs and flowers. In the second part of this experiment, the pigeons were trained to reassign one of the stimulus classes in each category to a new response key. Next, they were tested to see whether they would generalize the reassignment to the stimulus class withheld during reassignment training. It was found that the average score was 87 percent in the case of stimuli that had been reassigned and 72 percent in the case of stimuli that had not been reassigned. This performance, exceeding the level of chance, indicated that perceptually disparate stimuli had amalgamated into a new category. A similar experiment was performed on preschool children. The children's score was 99 percent for stimuli that had been reassigned and 80 percent for stimuli that had not been reassigned. In other words, the children's performance was roughly comparable to that of pigeons. Clearly, the performance of adult humans at this task will be superior to that of children or pigeons. Another interesting experiment related to the abstract concept of sameness. Pigeons were trained to distinguish between arrays composed of a single, repeating icon and arrays composed of 16 different icons chosen out of a library of 32 icons.  During training each bird encountered only 16 of the 32 icons; during testing it was presented with arrays made up of the remaining 16 icons. The average score for training stimuli was 83 percent and the average score for testing stimuli was 71 percent. These figures show that an abstract concept not related to the actual associations learnt during training had been internalized by the pigeon. Animal intelligence experiments suggest that one can speak of different styles of solv­ing AI problems. Are the cognitive capabilities of pigeons limited because their style has fundamental limitations? Can the relatively low scores on the sameness test for pigeons be explained on the basis of wide variability in performance for individual pigeons and the unnatural conditions in which the experiments are performed? Is the cognitive style of all animals similar and the differences in their cognitive capabilities arise from the differences in the sizes of their mental hardware? And since current machines do not, and cannot, use inner representations, is it right to conclude that their performance can never match that of animals? Most importantly, is the generalization achieved by pigeons and other nonhumans beyond the capability of machines? Donald Griffin (1992) expresses the promise of animal intelligence research thus: Because mentality is one of the most important capabilities that distinguishes living animals from the rest of the known universe, seeking to understand ani­mal minds is even more exciting and significant than elaborating our picture of inclusive fitness or discovering new molecular mechanisms. Cognitive ethology presents us with one of the supreme scientific challenges of our times, and it calls for our best efforts of critical and imaginative investigation.