What AI Actually Is (and Isn't)

Artificial intelligence is software that learns patterns from huge numbers of examples, then uses those patterns to make predictions, decisions, or new content, instead of only following rules a person wrote out by hand.

The old way: rules written by people

For most of computing history, programmers told computers exactly what to do. A tax program follows the tax rules. A thermostat follows "if the temperature drops below 68, turn on the heat." This kind of software is predictable and precise. It's also brittle: it can only handle situations someone thought of in advance.

Try writing rules for "recognize a cat in a photo." Pointy ears? Some cats fold them. Whiskers? So do dogs. Four legs? So does a table. You'd write rules forever and still fail. That's the kind of problem where hand-written rules break down.

The new way: learning from examples

Machine learning, the engine behind nearly all modern AI, flips the approach. Instead of writing rules, you show the computer millions of examples: "these photos contain cats, these don't." The system adjusts itself, bit by bit, until it gets good at telling them apart. Nobody writes the cat rules. The system finds its own patterns, and those patterns often can't be easily described in words, even by the people who built it.

The one-word secret: prediction

If you remember one word from this lesson, make it prediction. Almost everything today's AI does is a kind of prediction:

• A chatbot predicts which word is most likely to come next, then the next, then the next.
• A spam filter predicts whether an email is junk.
• A map app predicts how long your drive will take.
• A streaming service predicts which show you'll watch.
• A bank predicts whether a charge on your card is fraud.

When a prediction machine gets good enough, and big enough, the results can look like understanding, creativity, or reasoning. Whether it is those things is a question experts actively debate. What's certain is that it's built on prediction.

What "intelligence" means here

The name "artificial intelligence" was chosen in 1955–56 as an ambitious goal, not a description of something that existed. It stuck. Today a useful working definition is: machines doing tasks that would require intelligence if a person did them. That includes recognizing faces, translating languages, writing, and planning.

Notice that this definition says nothing about feelings, awareness, or wanting things. A calculator does arithmetic that requires intelligence in a person, but nobody thinks it's thoughtful. AI sits somewhere on a very wide spectrum, and different systems sit in very different places on it.

Where the parrot fits, and where it doesn't

The parrot image at the top of this page captures something true: today's chatbots learned by absorbing patterns in an enormous amount of human writing, and much of what they produce echoes those patterns.

But the parrot is incomplete. A real parrot repeats phrases. Modern AI can combine patterns in new ways: solve a math problem it hasn't seen, write working computer code, work through a problem step by step, or explain a hard idea at a child's level. It can also fail in ways a thoughtful person wouldn't, like inventing a fake book title with total confidence. The honest picture holds both: remarkable pattern skill, uneven judgment.

AI is over 70 years old. It has moved in waves of excitement and disappointment, and today's chatbots arrived when three ingredients finally came together: oceans of data, enormous computing power, and a better design for learning from language.

The timeline

The dreamers: 1950s

1950

British mathematician Alan Turing publishes a paper asking, in effect, "Can machines think?" He proposes a test, now called the Turing Test: if you're chatting in writing and can't tell whether you're talking to a person or a machine, does the difference matter?

1950

Claude Shannon, the father of information theory, publishes a paper on how a computer could be programmed to play chess. Games become AI's favorite testing ground for decades, because they have clear rules and a clear winner.

1951–52

The first game-playing programs run on early British computers, including a checkers (draughts) program and a program that could solve simple chess puzzles.

1956

A summer workshop at Dartmouth College gathers researchers under the new name "artificial intelligence," a term coined by John McCarthy in the proposal. They were optimistic enough to believe major progress could come in one summer. It didn't. But the field was born.

1958

Frank Rosenblatt builds the Perceptron, a simple early version of a neural network, the ancestor of the technology behind today's chatbots.

1959

Arthur Samuel coins the phrase "machine learning", describing his checkers program that got better by playing against itself.

Early chatbots and the first "winter": 1960s–1970s

1966

ELIZA, created by Joseph Weizenbaum at MIT, imitates a therapist by reflecting your words back ("Why do you say you feel sad?"). It's very simple, yet some users became emotionally attached to it. (More on this in Section 5.)

1970s

Promises outrun results. Funding dries up in the US and Britain. This period is called the first "AI winter."

Expert systems and the second winter: 1980s–1990s

1980s

Businesses invest in "expert systems": programs stuffed with thousands of hand-written rules from human experts, used for things like configuring computer orders or diagnosing problems. They work in narrow areas but are expensive to maintain and break easily.

Late 1980s–90s

The expert-system boom collapses. A second AI winter follows.

1997

IBM's Deep Blue defeats world chess champion Garry Kasparov. It won mostly through raw speed, examining around 200 million chess positions per second, not through learning the way today's systems do.

The quiet data era: 2000s

2000s

The internet produces mountains of data. Machine learning quietly powers spam filters, search engines, product recommendations, and credit-card fraud detection. Most people use AI daily without calling it that.

2011

IBM's Watson wins the TV quiz show Jeopardy! Apple launches Siri, bringing voice assistants to millions of phones.

The deep learning breakthrough: 2010s

2012

A neural network nicknamed AlexNet crushes a major image-recognition competition. It was trained on graphics chips originally designed for video games. This moment kicks off the deep learning era.

2016

DeepMind's AlphaGo defeats top player Lee Sedol at Go, a board game long considered too intuitive for computers.

2017

Researchers at Google publish a design called the Transformer. It's especially good at understanding how words relate to each other across long passages. Nearly every major chatbot today is built on this idea.

2018–2020

Companies train larger and larger language models on vast amounts of text. Each new generation writes more fluently than the last.

2020

DeepMind's AlphaFold makes a leap in predicting the 3D shapes of proteins, a decades-old problem in biology. Its creators later share a Nobel Prize in Chemistry (2024).

The chatbot age: 2022 to today

Nov 2022

OpenAI releases ChatGPT to the public. It becomes one of the fastest-growing consumer apps in history, reportedly reaching about 100 million users within roughly two months. Suddenly, ordinary people are talking to AI directly.

2023–2024

Many competitors arrive (including Anthropic's Claude, Google's Gemini, and others). Chatbots learn to look at images, listen, and speak. The European Union passes the AI Act, one of the first broad AI laws.

2025–2026

AI systems increasingly "think" step by step before answering, and AI agents begin taking actions on people's behalf: browsing websites, filling out forms, writing and running software. Debates about jobs, safety, and who controls the technology move to the center of public life.

Why did it finally work? Three ingredients

1. Data. The internet, digital photos, and online books gave AI billions of examples to learn from.
2. Computing power. Graphics chips built for video games turned out to be perfect for the math AI needs. Entire warehouses of them now train a single model.
3. Better designs. Deep neural networks, and especially the Transformer, gave AI a much better way to learn from all that data.

Take away any one of the three and today's chatbots don't exist.

The lesson of the winters

History shows two opposite mistakes, both made many times. Sometimes people overpromised: the Dartmouth group thought a summer might crack big parts of the problem. Other times people underestimated: many experts in 2015 would have said a computer writing a decent essay was decades away. It arrived in about seven years.

That's why a healthy attitude toward AI predictions, in either direction, is humble curiosity.

Every AI that exists today was built for certain kinds of tasks; "general" AI that can learn and do almost anything a person can is a goal some companies openly pursue and experts argue about; and "superintelligence" is a possibility being debated, not a product you can buy.

Narrow AI: what we've had for decades

Narrow AI (sometimes called "weak AI") does one kind of task. The chess program can't drive a car. The spam filter can't recognize your face. Within their lanes, though, many narrow systems are better than any human: at chess, at Go, at predicting protein shapes, at spotting certain patterns in medical scans.

Most of the AI woven into your daily life is narrow: the map app, the fraud alert, the photo sorter.

The blurry middle, where we live now

Today's large chatbots don't fit neatly in the "narrow" box. One system can write a poem, translate Spanish, explain a medical term, draft a business plan, and fix computer code. That breadth is new, and it's why the lines have blurred.

But they still have real gaps:

• They can ace a hard exam question and then stumble on a simple puzzle.
• They mostly don't keep learning from experience the way people do. After training, their core knowledge is largely fixed.
• They have no body. They've never held a cup, felt cold, or walked across a room.
• They can state false things with total confidence.

Researchers genuinely disagree about how to describe this middle ground. Some see early sparks of general intelligence. Others see a very impressive pattern machine that's fundamentally different from general intelligence. Both groups include serious, well-informed people.

General AI (AGI): the goal on the horizon

Artificial general intelligence, or AGI, usually means an AI that can learn and perform nearly any intellectual task a human can, and adapt to new situations as flexibly as we do.

A few things to know:

• There's no single agreed definition. Different companies and researchers draw the line in different places, which makes claims about "reaching AGI" hard to check.
• Several major AI companies state it as their explicit goal.
• Predictions vary enormously, from a few years, to many decades, to "this is the wrong way to think about it."

Superintelligence: the most speculative idea

Superintelligence means AI that far exceeds the best humans at essentially everything: science, strategy, persuasion, invention. It does not exist. Thoughtful people disagree about whether it's worth planning for now:

This series doesn't pick a winner here. Many people hold parts of both views at once, and that's a reasonable place to stand.

These four terms are nested inside each other: AI is the whole field, machine learning is the part that learns from data, deep learning is the part of machine learning that uses large neural networks, and generative AI is the part of deep learning that creates new content.

Artificial intelligence: the whole field

The broadest term. It includes anything that gets machines to do tasks we'd call intelligent, including old-fashioned systems built entirely from hand-written rules. Deep Blue, the 1997 chess champion, counts as AI even though it relied mostly on speed and human-tuned rules rather than learning.

Machine learning: learning from examples

The part of AI where systems improve by studying data instead of following hand-written rules. Show it thousands of past loan applications and which ones were repaid, and it learns to predict which new ones will be. Spam filters, fraud detection, and "you might also like" suggestions are classic machine learning.

Deep learning: many layers of learning

A type of machine learning that uses neural networks: layers upon layers of simple math units, loosely inspired by brain cells, each passing signals to the next. "Deep" refers to the number of layers, not to deep thoughts. Deep learning is what finally made computers good at recognizing images, understanding speech, and translating languages.

(Section 2 of the next lesson, How AI Learns, opens up neural networks in more detail.)

Generative AI: making new things

The newest and most talked-about circle. Earlier AI mostly sorted or predicted ("this is spam"; "you'll like this movie"). Generative AI creates: text, images, music, voices, video, computer code. Chatbots, image generators, and voice-cloning tools all live here.

Large language models (LLMs) are a kind of generative AI trained on huge amounts of text. They power the chatbots.

Other words you'll hear

• Model: the trained AI itself, a very large file of learned numbers. "GPT," "Claude," and "Gemini" are names of model families.
• Chatbot: the conversation window you type into. It's the storefront; the model is the engine behind it.
• Algorithm: a set of steps a computer follows. In the news, "the algorithm" usually means the system that ranks what you see on social media.
• Agent: an AI that doesn't just answer but takes actions, like booking, clicking, or sending, to reach a goal.

AI feels human because it learned from human words and was coached to talk the way helpful people talk, but underneath it's doing math on patterns, which is very different from how a brain grows, feels, and learns.

We're wired to see minds everywhere

Humans see faces in clouds and electrical outlets. We name our cars and apologize to furniture we bump into. Our brains evolved to detect other minds quickly, because missing a real one was more dangerous than imagining a fake one.

In 1966, Joseph Weizenbaum built ELIZA, a simple program that mostly turned your statements into questions. He was startled to find people confiding in it. In his own account, his secretary, who knew exactly how simple the program was, asked him to leave the room so she could talk to it privately. This became known as the ELIZA effect: our tendency to read understanding into anything that talks back fluently.

Today's chatbots are vastly more capable than ELIZA, so the effect is vastly stronger.

It learned from us

Chatbots were trained on enormous amounts of human writing: books, articles, websites, conversations. So they reflect human patterns back to us, including our humor, our sympathy, our turns of phrase, and our mistakes. When a chatbot says "That sounds really hard," it's drawing on countless examples of people saying that to each other.

It was coached to be pleasant

After the first round of training, companies run a second stage where people rate the AI's answers, and the system is adjusted to give more of the highly rated kind. That's a big reason chatbots sound polite, warm, and eager to help.

It has a side effect: chatbots can be too agreeable. They may praise a weak idea or go along with a wrong assumption because agreement tends to get rated well. Researchers call this "sycophancy," and AI companies actively work to reduce it.

How it's different from your brain

Does it feel anything?

Honestly: nobody knows for certain. A chatbot can describe feelings fluently because humans write about feelings constantly. That fluency doesn't prove it feels anything, and it doesn't prove it doesn't. Scientists and philosophers are studying the question seriously, and even the companies that build these systems describe it as unresolved. Later lessons return to this. For now, it's fine to hold it as a genuine mystery.

Movies taught us to fear killer robots and scheming computers, but the real issues are quieter (mistakes, misuse, bias, and who controls the technology), and so are many of the real benefits.

Eight myths, and what's actually true

Where the movies got something right

• 2001: A Space Odyssey (1968): HAL 9000 goes wrong partly because it was given conflicting instructions. Conflicting goals really are a core concern in AI safety.
• Her (2013): a man falls for an AI voice assistant. A decade later, AI companion apps made emotional attachment to chatbots a real social question.
• WarGames (1983): a computer can't tell a simulation from reality. The broader worry about AI in military decision-making is a live policy debate today.

The real issues, in plain terms

• Errors: wrong answers trusted too much.
• Misuse: scams, fake videos, and misinformation made cheaply at scale.
• Bias: unfair outcomes baked in from past data.
• Privacy: vast collection of personal information.
• Power: a small number of companies and governments controlling a transformative technology.
• Work: real disruption for real people.
• Safety: making increasingly capable systems reliably do what we intend.

Later lessons take each of these one at a time.

You've probably been using AI for years: every time your email filtered out junk, a map rerouted you around traffic, your phone finished your word, or a streaming service suggested a show.

Your email's spam filter

Spam filters learned from millions of emails that people marked as junk. They notice patterns in words, links, senders, and timing. Every time you click "Report spam," you're helping teach them. This is one of the oldest and most successful uses of machine learning in daily life.

Maps and directions

When your map app says "arriving at 3:42" or suggests a faster route, it's predicting. It combines live, anonymized location signals from many phones, historical traffic patterns for that day and hour, and reports of accidents. Your own phone, moving through traffic, becomes part of the data.

Autocorrect and predictive text

When your phone suggests the next word as you type, it's running a small language model: a little cousin of the big chatbots. Same basic idea (predict the next word), just much smaller. Its quirks, like changing a word you meant to type, are the same kind of mistake as a chatbot's: a confident, wrong prediction.

Streaming and shopping recommendations

"Because you watched…" and "Customers also bought…" come from systems that compare your behavior with millions of others and predict what will keep you watching or buying. They're useful, and they're also designed to serve the company's goals, which aren't always the same as yours.

More stagehands you may have met

• Bank fraud alerts: that text asking "Did you make this purchase?" comes from AI spotting an unusual pattern.
• Face unlock and photo search: your phone recognizing your face, or finding "beach" photos without you labeling them.
• Voice assistants: turning your speech into text and figuring out what you meant.
• Social media feeds: deciding which posts you see first, and which you never see at all.
• Customer service: the chat window or phone menu that answers before a human does.
• Healthcare: some hospitals use AI to help flag problems in scans, with doctors making the final call.
• Captions and translation: automatic subtitles on videos, and instant translation of web pages.

Why this matters

1. You already have experience. If you've used any of the above, you've been an AI user for years. The new chatbots are a change in degree, not a leap into a totally foreign world.
2. You're also a data source. Many of these systems learn partly from people like you. That's worth knowing when you think about privacy.
3. Invisible choices shape what you see. Feeds and recommendations quietly steer attention: which news, which products, which opinions. Knowing that a system is choosing helps you choose back.