AI is a rapidly evolving field, so understanding what we mean by intelligence is an important first step. Before we can compare human intelligence with artificial intelligence, we need to think about what intelligence actually is. Psychologists, cognitive scientists, computer scientists, philosophers, educators, and many others have debated this question for centuries. Although there is no single universally accepted definition, most people agree that intelligence involves the ability to learn, reason, solve problems, adapt to new situations, and make decisions.
Many people associate intelligence with an intelligence quotient (IQ) score. While IQ tests measure certain kinds of reasoning and problem-solving ability, they do not capture every aspect of human intelligence. One influential, but controversial, educational theory, proposed by Howard Gardner, suggests that people possess several different kinds of intelligence rather than a single general intelligence. Gardner’s theory includes intelligences such as linguistic, logical-mathematical, spatial, musical, interpersonal, and bodily-kinesthetic abilities. Although psychologists continue to debate this theory, it has had an influence on education by encouraging educators to recognize that people learn and demonstrate their strengths in different ways.
Gardner’s Theory of Multiple Intelligences is widely known in education, but it is not universally accepted within the field of psychology. Many psychologists argue that intelligence is better explained by a smaller number of underlying cognitive abilities. Throughout this book, we use the word “intelligence” in its broad everyday sense rather than adopting any single psychological theory.
Rather than thinking of intelligence as one specific skill, it is helpful to think of it as a collection of abilities that allow an individual to understand the world, learn from experience, solve problems, and adapt to changing situations. Humans demonstrate these abilities in many different ways, and modern AI systems can now perform some of them remarkably well. However, AI and human intelligence are not the same, and understanding those differences is one of the goals of this book.
Because intelligence is difficult to define precisely, early AI researchers looked for practical ways to evaluate it. In 1950, Alan Turing posed a provocative question: Can machines think? Rather than diving into philosophical definitions, Turing reframed the question in practical terms with his now-famous Turing Test, introduced in his paper “Computing Machinery and Intelligence.”
The concept is straightforward: a human judge engages in a text-based conversation with both a human and a machine, without knowing which is which. If the judge cannot reliably distinguish the machine from the human, the machine is said to have passed the test—demonstrating behavior indistinguishable from human intelligence. Turing’s approach shifted the focus from the elusive concept of “thinking” to observable interaction.
Turing actually presented three versions of this imitation game. The 1950 version is the most well-known: a human interrogator has simultaneous text conversations with both a computer and another human, and must judge which is which. The game is unrestricted, meaning the interrogator can communicate on “almost any” subject. Turing said that if a computer does well in this game, it can be considered to be thinking.
However, it is important to understand what the test does and does not claim. For Turing, doing well in the game is a sufficient condition for thinking—meaning a machine that passes can be said to think—but it is not a necessary condition. A machine that fails the test might still be intelligent. Turing also did not intend the test as a strict definition of thinking. In fact, he stated: “I don’t want to give a definition of thinking.”
The test has several key features. It assesses a machine’s ability at open-domain conversation, rather than in narrow, specialized fields. It is qualitative, allowing for natural dialogue, and it does not rely on the tricky, adversarial questions that programmers often use to try and expose chatbots. Significantly, the test is also architecture-independent, meaning it does not dictate what kind of machine or technology must be used—it could be a neural network, a DNA computer, or any other system.
For decades, the Turing Test shaped how many people thought about artificial intelligence. Today, however, AI systems perform many impressive tasks that have little to do with carrying on a conversation. As a result, researchers now evaluate AI in many different ways depending on the problems they are trying to solve.
One reason artificial intelligence has become such a common topic of conversation is the rapid growth of generative artificial intelligence (GenAI). Unlike many earlier AI systems that focused on recognizing patterns or making predictions, generative AI creates new content. Depending on the system, that content may include text, images, audio, video, computer code, or three-dimensional designs.
Generative AI learns from large collections of examples known as training data. During training, the model identifies patterns and relationships within the data rather than memorizing a fixed collection of answers. When given a prompt, it uses those learned patterns to generate a new response. Although the output is influenced by its training data, it is not simply retrieved from a database or copied from a single source.
The same underlying idea applies regardless of what the system produces. A language model can generate essays, conversations, or computer programs. An image model can create original illustrations. Other systems generate music, speech, video, or scientific designs. While the outputs differ, they are all created by learning patterns from data and using those patterns to generate something new.
Like all AI systems, generative AI has both strengths and limitations. It can produce creative and useful results in seconds, but it can also make mistakes, reflect bias present in its training data, or generate information that is inaccurate or misleading. Learning to understand these strengths and limitations is an important part of becoming an informed user of AI.
In later chapters, we will examine how generative AI systems learn from data, why they sometimes make mistakes, and how they are changing the way people learn, work, and solve problems.
To offer a formal standard for evaluating whether a machine can be said to possess genuine understanding.
Turing did not propose the test as a formal standard for genuine understanding. He deliberately avoided giving a definition of thinking, stating: “I don’t want to give a definition of thinking.” The test evaluates behavior, not internal understanding.
To provide a systematic framework for determining the specific cognitive abilities that any intelligent system must demonstrate.
Turing’s test does not specify which cognitive abilities an intelligent system must demonstrate. Instead, it focuses on observable behavior in open-domain conversation, allowing any system that can imitate human conversation to pass, regardless of its internal architecture.
To establish a practical method for assessing machine intelligence based on observable behavior rather than abstract definitions.
Correct! Turing shifted the focus from the elusive concept of “thinking” to observable interaction. He proposed a practical test based on whether a machine could successfully imitate human conversation, avoiding the need for a strict definition of intelligence.
To design an interrogation protocol that would reliably show the limitations of early rule-based systems.
Turing’s test was not designed to expose limitations or to be adversarial. In fact, the test disallows the tricky, adversarial questions that programmers often use to try and expose chatbots. The goal was to evaluate imitation, not to reveal weaknesses.
