What makes something intelligent?

What Is Intelligence?

Here is a question nobody has fully answered: what is intelligence? Psychologists say it is the capacity to reason and learn. Biologists link it to brain complexity. Philosophers argue about consciousness and self-awareness. Engineers define it as efficient problem-solving. After thousands of years of studying this, no one has settled on a single definition.

That is not a failure. It reflects how genuinely complicated intelligence is.

For this course, here is a working definition that applies equally to humans, animals, and machines: Intelligence is the ability to perceive an environment, learn from experience, and take actions that achieve goals.

Notice what this definition does not require. It says nothing about having a brain, being conscious, or feeling emotions. It describes a functional capability. That is intentional, because it lets us think clearly about artificial intelligence without getting stuck in debates about machine consciousness.

The Turing Test: A Clever Shortcut

In 1950, British mathematician Alan Turing published a paper with a memorable opening question: "Can machines think?"

Instead of getting lost in philosophy, he proposed something cleverer. Imagine a human judge typing messages back and forth with two participants in separate rooms. One participant is human. The other is a machine. The judge can ask anything they want, for as long as they want. If the judge cannot consistently identify which participant is the machine, the machine has passed the test.

Turing called it the Imitation Game. Today we call it the Turing Test.

The genius is what Turing was doing: sidestepping the philosophy entirely. You do not have to define what thinking means. You just ask whether the machine behaves in a way indistinguishable from a thinking human. If it does, then for practical purposes it is intelligent.

Does GPT-4 Think?

Modern large language models can pass informal versions of the Turing Test. In short conversations, many people cannot tell they are chatting with an AI. GPT-4 has been shown to score in the 90th percentile on the Uniform Bar Examination, pass the US Medical Licensing Exam, and achieve near-perfect scores on the GRE Verbal Reasoning section.

So does GPT-4 think?

Most researchers say: not in the meaningful sense. GPT-4 was trained to predict which text comes next, given all the text that came before. It has seen so much human writing that it learned to produce text statistically indistinguishable from how humans write. That is a profound capability. But it is a capability about patterns in text, not necessarily about understanding the world.

The gap between producing convincing human-like text and genuinely understanding things is one of the deepest open questions in AI today. It does not take away from what these systems can do. But it matters for understanding their limits.

The Three Levels: A Map You Need

AI researchers divide possible machine intelligence into three categories. You will see this framework in every serious discussion about AI, so it is worth internalizing now.

• •Narrow AI (also called Weak AI) is designed for one specific task. It can be extraordinary at that task, often far better than any human. But it cannot transfer its skills to anything else. Every AI system that exists today is narrow AI. ChatGPT, DALL-E, AlphaGo, Siri, your spam filter, Face ID, and every Netflix recommendation you have ever received. All narrow AI.

• •Artificial General Intelligence (AGI) would be an AI capable of performing any intellectual task a human can, with the same flexibility, context-awareness, and ability to learn new things from scratch. AGI does not exist yet. Building it requires solving problems we do not currently know how to solve, including reliable common-sense reasoning, physical understanding of the world, and genuine generalization from small amounts of new information.

• •Artificial Superintelligence (ASI) would exceed human capabilities across every domain simultaneously. This is entirely theoretical. Nobody has a working approach to building it, and nobody knows whether it is physically achievable. It remains a philosophical and scientific debate, not an engineering roadmap.

Why a Chess AI Cannot Write a Poem

Here is a concrete example of why the distinction matters.

Stockfish is the strongest chess engine in the world. It plays at a level so far above any human that grandmasters lose to it on every single game. It evaluates millions of chess positions per second and makes moves that world champions cannot fully understand in real time.

Ask Stockfish to write a haiku. It returns an error.

Ask it to translate a sentence into French. Error.

Ask it to explain what it is doing and why it made a particular move. It cannot. It has no concept of what it is doing. It processes numbers representing board positions and outputs numbers representing moves.

That is narrow AI. Superhuman inside its domain, completely helpless outside it. The capabilities do not transfer.

GPT-4 looks more versatile because its domain is language tasks in general. But it is still narrow AI. It processes text and produces text. Give it a physical task, ask it to form a genuine goal and pursue it over weeks without external prompting, or ask it to learn a new concept the way a child picks up language: it fails. Its domain is narrow, even if that domain happens to be very large.

Where We Stand Today

We are living through an extraordinary period in the history of narrow AI. AlphaFold solved a 50-year-old problem in biology. GPT-4 passes professional licensing exams. DALL-E 3 creates photorealistic images from a sentence. Autonomous vehicles navigate real streets. AI reads medical scans with radiologist-level accuracy.

All of this is narrow AI. All of it is real, and all of it is transforming industries right now.

The gap to AGI remains wide. Nobody knows how to bridge it, or even precisely what bridging it would require. That is actually exciting: you are entering this field at a moment when the biggest questions are still open, the answers are not written yet, and the tools are powerful enough to start doing real work with.

The rest of this module gives you the context to understand how we got here. First stop: a quick tour through 70 years of AI history.