What is Artificial Intelligence, really?

An Honest Definition to Start

If you look up "Artificial Intelligence" in ten different books, you'll get ten different answers. That's not laziness on the authors' part, the term really is fuzzy. The most useful explanation goes roughly like this:

Artificial Intelligence is the attempt to get machines to solve tasks that we would normally need human intelligence for.

Understanding language, recognizing images, making decisions, learning from experience. When a computer does something like that, we like to say "the AI did it". What matters here: it's not about the machine "thinking" like a human. It's about the fact that, in the end, there's a result that used to be considered a purely human achievement.

The Term Is Much Older Than You Think

Many people associate AI with ChatGPT and believe the topic was suddenly invented in 2022. In fact, the term is over seventy years old. Back in 1950, the British mathematician Alan Turing asked in a famous essay: "Can machines think?" He devised a small test, known today as the Turing test, in which a human writes with a machine over a screen and is not supposed to notice that no human is replying.

In 1956, a summer workshop took place at Dartmouth College in the USA, where the term "Artificial Intelligence" was used officially for the first time. The researchers there believed at the time that they could solve the most important problems in a few weeks. They were off by a few decades.

Since then there have been waves: phases full of enthusiasm, in which everyone believed the breakthrough had finally arrived, followed by phases of disillusionment known as "AI winters". What we're experiencing right now is the biggest wave so far, but just one in a long series.

Weak AI and Strong AI, the Most Important Distinction

If you want to remember just one term from this chapter, make it this one: there is a huge difference between weak AI and strong AI. The two are cheerfully mixed up in everyday life and in the media, and that's exactly where most misunderstandings come from.

Weak AI, What Actually Exists

Weak AI (also called narrow AI) is AI that's good at exactly one specific task. A spam filter recognizes junk mail. A translation AI translates text. A language model like ChatGPT generates sentences. Each of these AIs is genuinely good within its narrow area, but as soon as you take it out of that area, nothing works anymore. The spam filter can't drive a car. The translation AI can't recognize images.

Everything you see labeled "AI" today is weak AI. Really, all of it. Even though ChatGPT sometimes seems as if it understands everything, at its core it solves a very narrow task (predicting the next word) and uses an extreme amount of training to do it.

Strong AI, the Science-Fiction Image

Strong AI (also called general AI or AGI) would be a machine with a kind of general intelligence like us humans. It could cook in the morning, write a book at noon, play chess in the evening, and learn to practice piano on the side, all without anyone having trained it specifically for any of that. This form of AI does not exist. It's a research topic, it's speculation, it's Hollywood. But it's not what's inside your smartphone.

Everything you call AI today is weak AI. Strong AI is a promise, not a product.

Where You've Already Been Using AI Today

You've probably already dealt with AI several times today without thinking about it. Here are a few examples from everyday life:

• Spam filters in your email program: an AI has learned from millions of examples what a real message looks like and what a junk email looks like.
• Navigation and route planners: behind "the fastest route in 14 minutes" are predictions about traffic, traffic lights, and the behavior of other drivers.
• Netflix, Spotify, YouTube recommendations: "You might also like this" is nothing more than an AI system comparing you to millions of similar users.
• Face unlock on your phone: the camera sends your image through a neural network that checks in a split second whether it's you.
• Autocorrect and keyboard suggestions: this is also AI, it guesses which word you want to type next.
• Bank and credit card security: if an unusual payment gets blocked, an AI system has usually flagged it as suspicious.

The Decisive Difference: Programming vs. Learning

Until now, software always worked like this: a programmer writes down rules. If the user clicks this button, then open this window. If the date is a Sunday, then show "closed". That's classic programming, and it works excellently for clear, unambiguous tasks.

But for many interesting tasks, rules don't get you very far. Try writing down how you'd recognize a cat in a photo. Four legs? Dogs have four legs too. Pointy ears? Some cats have round ones. Fur? Sphynx cats have almost none. You'll never manage to create a complete list of rules. The human brain does it differently, it has seen thousands of cats and developed a feel for what a cat is.

And that's exactly where modern AI comes in. Instead of programming in rules, you show the system thousands or millions of examples and let it figure out on its own which patterns matter. An AI is not directly programmed, it is trained.

This has two enormous consequences. First: AI can take on tasks for which there simply are no clean rules. Recognizing images, understanding language, gauging moods. Second: you can no longer "debug" AI in the classic sense. If the system classifies a cat as a dog, there isn't one broken line of code you can fix. You have to adjust the training, or live with the error.

So What Do "Machine Learning" and "Deep Learning" Mean?

You'll hear these two terms often. They all fall under the umbrella of "AI", but they are not the same thing:

So these terms aren't synonyms, they're nested inside each other like Russian dolls: generative AI is an application of deep learning, deep learning is a form of machine learning, machine learning is a part of AI.

What AI Is Not

Finally, a short list of things AI is not, even if films, advertising, or headlines often make it sound otherwise:

• AI is not conscious. There is no "I", no feeling, no intent. Even if the answers sometimes sound that way.
• AI does not understand anything in the human sense. It recognizes patterns and reacts to them. That's impressive, but it's something different from understanding.
• AI is not objective. It picks up biases from the data it was trained on. Garbage in, garbage out.
• AI is not infallible. It makes mistakes confidently, and often doesn't even notice. More on this in the "Hallucinations" chapter.
• AI is not magic. Behind every result is traceable mathematics, a lot of data, and a huge amount of computing power.

If you remember this, you already have a better grasp of AI than a lot of people who talk about it every day. In the next chapter, we'll look at how an AI actually arrives at its answers, and why it can rightly be called a "very good guessing system".