What Is AI? A Simple Explanation

My aunt asked me over dinner last month what AI actually is. She’d seen it on the news, her bank had started putting “AI-powered” on everything, and her grandson kept telling her he used it for homework. She wasn’t scared of it. She just felt like everyone was talking about something she was supposed to already understand.

I gave her a bad answer. Something about computers learning from data. She nodded politely and changed the subject.

This post is the answer I wish I’d given her. If you’ve been wondering what AI is and feel a bit behind, you’re not. Most people who use the term every day would struggle to explain it. So let’s fix that.

The short answer

AI, or artificial intelligence, is when a computer does a job that used to need a human brain.

That’s really it. Not a human body. A human brain. Reading, recognizing, sorting, deciding, predicting, writing. Things that take judgment or pattern-spotting instead of muscle.

Some examples make it clearer.

Your phone looks at a photo and knows your sister is in it. That’s AI. Your email puts the junk in the junk folder before you see it. AI. Netflix figures out you’d probably like a Norwegian crime show. AI. A program takes a paragraph in Spanish and hands it back in English. AI. A chatbot writes you a birthday poem about your dog. Also AI.

Notice that none of this says the computer thinks like you, feels anything, or knows what it’s doing. It just says the computer produced a result that used to require a person. Hold onto that. It’ll matter later.

One odd thing about the word: once a computer gets good at something, we stop calling it AI. Nobody calls a calculator artificial intelligence, but doing math in your head was once considered a sign of a sharp mind. Spell-check used to be impressive. Now it’s just a thing. So “AI” tends to mean “the stuff computers only recently figured out,” and the definition keeps sliding forward.

A little history (only the useful parts)

AI is not new. That surprises people.

The name was invented in 1956 by a small group of researchers who got together for a summer at Dartmouth College to see whether machines could be taught to reason. They were confident. Some of them thought they’d have it mostly solved within twenty years.

Before that, in 1950, a British mathematician named Alan Turing wrote a paper asking whether a machine could think. He suggested a simple test: if you’re typing back and forth with something and can’t tell whether it’s a person or a program, then the program might as well count as intelligent. People still argue about whether that’s a fair test, but it set the goal.

What followed was a lot of false starts. Researchers would get excited, money would pour in, the programs would work in the lab and fall apart in the real world, and the money would leave again. People in the field call those dry spells “AI winters.” There were a couple of them.

I bring this up because today’s AI boom didn’t come out of nowhere. It’s seventy years of slow progress that suddenly hit a tipping point. Which raises the question: what tipped it?

Two ways to make a computer act smart

This is the part that made it click for my aunt, so I’ll take my time. Say you want a computer to spot spam email.

Option one: you write the rules yourself. If the sender is unknown and the email has five exclamation marks, spam. If it mentions a price and a bank transfer, definitely spam. You keep adding rules until the filter is pretty good. The problem is that you have to be able to write the rules down, and for a lot of tasks, you can’t. Try writing rules for recognizing a cat in a photo. What does a cat look like, in terms a computer could check? Pointy ears? So do foxes. Fur? Lots of things have fur. You’d be at it forever and still miss a cat curled up in a box.

Option two: you show the computer examples and let it work out the rules itself. Ten thousand emails, each one labeled “spam” or “not spam.” A million photos, each one labeled “cat” or “not cat.” The computer finds the patterns that separate one pile from the other. You never write a single rule. That second option is called machine learning. And almost every time someone says “AI” today, this is what they mean. The switch from option one to option two is the biggest thing that ever happened in this field. It turned AI from “clever people writing clever rules” into “collect a mountain of examples and let the machine sort it out.”

So how does a machine actually "learn"?

The word “learn” makes it sound magical. It’s not. It’s closer to tuning a guitar than going to school. Picture a machine covered in millions of tiny dials. At the start, every dial is set randomly, so the machine’s guesses are useless. You show it a photo of a cat. It guesses “dog.” You tell it “wrong, that was a cat,” and the machine turns each dial a hair in the direction that would have made “cat” slightly more likely. Then you show it another photo. And another. Millions of them. By the end, the dials are set so that cat photos reliably come out as “cat.” Nobody sat down and defined what a cat is. The machine just adjusted itself, one tiny nudge at a time, until it stopped being wrong. That’s machine learning. The “learning” is a very large number of small corrections.

It’s a bit like how a toddler learns what a dog is. You don’t hand a two-year-old a definition. You point at dogs and say “dog” a few hundred times, and at some point they start pointing at dogs themselves. They also point at a sheep and say “dog,” which, as it happens, is also what AI does. More on that later. There are a few versions of this you might hear about. Sometimes the examples come with labels, like the spam and cat cases. It’s the workhorse of the industry, and it’s behind things like fraud detection and reading medical scans. Sometimes the computer gets no labels at all and just looks for groups and patterns in the data on its own. And sometimes it learns by trial and error, getting a “reward” when it does well, which is how a program learned to beat one of the world’s best Go players back in 2016. If you take one thing from this section, take this: a machine learning system is a very good guesser. It gives you the answer that best fits the patterns it saw during training. Often that looks exactly like expertise. Sometimes it’s confidently, cheerfully wrong.

Deep learning, in plain words

You’ll hear “deep learning” and “neural networks” a lot. Here’s what they are without the textbook. A neural network is a particular way of building that dial-covered machine. The dials are arranged in layers, and each layer hands its work to the next one. When it looks at a photo, the first layer notices simple stuff like edges and blobs of color. The next layer combines those into shapes. The one after that combines shapes into things like eyes and ears. Keep stacking, and by the top layer the machine can say “cat.” “Deep” just means there are a lot of layers. The idea has been around for decades. It didn’t really work until three things showed up at the same time: huge amounts of data (the internet), cheap computing power (the graphics chips built for video games turned out to be perfect for this), and a few clever tricks for training big networks without them falling apart. Around 2012, all of that came together, and neural networks went from a niche interest to the thing everybody uses.

The Canadian part of the story

If you’re reading this in Canada, you might like to know that a lot of this happened here. Geoffrey Hinton, a professor at the University of Toronto, kept working on neural networks through the years when most of the field had written them off. In 2012, a program from his lab won a big image recognition contest by such a wide margin that the rest of the field changed course within a couple of years. Yoshua Bengio was doing related work in Montreal, and the two of them, along with Yann LeCun, who’d spent time in Hinton’s lab early on, won the Turing Award in 2018, which is about as big as prizes get in computing. Hinton later shared the 2024 Nobel Prize in Physics. Out in Edmonton, Richard Sutton co-wrote the standard textbook on the trial-and-error kind of learning and got his own Turing Award for it. Canada was also the first country to launch a national AI strategy, back in 2017. So when you hear about the Toronto, Montreal and Edmonton AI scenes, that isn’t marketing. A big chunk of this was figured out here.

What ChatGPT and friends actually do

This is the part most people are curious about, because the chatbots are the first time AI has felt like something you can talk to. I’m serious. That’s the whole job. You give the model some text, it guesses what word should come next, adds it, looks at the new text, and guesses again. Over and over, until it has written a paragraph. It sounds too dumb to be useful, and for a long time it was. But something surprising happens when you make the model enormous and train it on a huge chunk of everything humans have written. To predict the next word well across all of that text, the model has to pick up grammar, facts, arguments, jokes, coding styles, how a recipe is laid out, how an apology sounds. All of it, just from playing “guess the next word” billions of times. So when you ask it for a pirate poem about a lighthouse, it isn’t looking one up. It’s generating the most likely next word, then the next, given that request. And the most likely continuation of “write a pirate poem about a lighthouse” is, sure enough, a pirate poem about a lighthouse.

After that basic training, the companies do a second round where humans rate the model’s answers and it gets tuned to give the kind of responses people find helpful and polite. Image generators, the ones that turn “a raccoon in a spacesuit” into a picture, work on a related idea. They learned how descriptions and images go together, and now they can produce a new image to match a description.

Now the catch, and please don’t skip it.

If you ask a language model about something it doesn’t have good information on, it won’t necessarily tell you that. It will often produce the most convincing-sounding answer it can, which may be made up. People in the industry call this “hallucination.” It’s the reason you should never take a fact from a chatbot without checking it, especially names, numbers, dates and quotes.

Three kinds of AI you'll hear about

People throw around a few terms as if they mean the same thing. They don’t.

Narrow AI is a system built for one job or a limited set of jobs. Every AI that exists right now is narrow AI. That includes the chatbots. They’re very broad at language, but they can’t learn to fix your sink by watching you do it.

General AI, or AGI, is the idea of a machine that could learn and do anything a person can, with the same flexibility. This doesn’t exist. Whether we’re close, and whether the current approach will ever get there, are questions the people building these systems argue about constantly.

Superintelligence is the more far-out idea of a machine that outperforms humans at pretty much everything. It’s something serious researchers think about, and also something movies love, and it helps to keep those two separate.

Quick rule of thumb: when a headline says “AI did something,” it means narrow AI. When a headline says AI is going to take over the world, it’s talking about the other two, which are ideas rather than products.

The AI you already use every day

Here’s what surprised my aunt most. She thought AI was arriving. It’s been in her house for years.

Her phone unlocks by looking at her face. Her camera decides how bright to make a photo before she even taps. Her email hides the spam. Her map app guesses the traffic and picks the route. Her bank texted her last winter about a weird purchase, which was a program noticing that the purchase didn’t fit her pattern. Her TV app decides what to show her first. The little word suggestions above her keyboard are a mini version of the same idea behind ChatGPT.

None of it says “AI” on the box. It’s just features. That’s what AI looks like when it works: it blends into the product and you forget it’s there. The chatbots only feel different because you’re talking to the model directly, and talking is the thing we most associate with having a mind.

What AI is not

I’d put this section at the top if I thought people would read it there.

AI is not conscious. Nothing we’ve built shows any sign of experiencing anything. When a chatbot says “I understand how you feel,” it’s producing the words that fit the conversation, the same way it would produce a muffin recipe if you’d asked for one.

It has patterns. That’s why a chatbot can explain something perfectly and then get a slightly reworded version of the same question wrong, or invent a book that doesn’t exist. Smart people disagree about whether this counts as “understanding.” For day-to-day purposes, assume it doesn’t.

AI is not neutral. It learns from data, and data comes from people, with all our habits and blind spots baked in. Train a hiring tool on ten years of past hiring decisions and it’ll learn ten years of past hiring prejudices. This is one of the best-documented problems in the whole field.

AI doesn’t know when it’s wrong. A human expert who isn’t sure usually sounds unsure. A language model sounds exactly as confident when it’s making things up as when it’s right.

And AI is not magic. Everything it can do came from data, computing power and a lot of engineering. It has no secret line to the truth. Think of it as a fast, capable, slightly unreliable co-worker. That mental picture will serve you better than either “robot genius” or “fancy autocomplete.”

Why is everyone talking about it now?

Fair question. If the field is seventy years old and the technology has been in your phone for a decade, why did AI suddenly take over every conversation?

Two reasons.

First, it got a lot better, quickly. A new design for neural networks came out in 2017, companies started training models on enormous amounts of text with enormous amounts of computing power, and the results jumped in a way that normal people could see, not just researchers.

Second, and I think this is the bigger one, someone put a chat box on it. When ChatGPT came out in late 2022, ordinary people could type a question in plain English and get back something that sounded like a person. Before that, AI was buried inside products. Suddenly it was something you could argue with at midnight. Which is what made it feel real, and it kicked off the wave of money, competition and worry we’re still in.

The worries, and which ones I take seriously

Any honest explanation of AI has to deal with the concerns, and some of them are real.

Jobs, first. AI is good at reading, writing, summarizing, sorting and spotting patterns, and that describes a lot of office work. I don’t think the near future is “robots replace everyone.” I think it’s closer to “people who use these tools well end up replacing people who don’t,” which is still a big deal if you’re the one being replaced. Past technology shifts were rougher on workers than the cheerleaders admitted at the time.

Fakes, second. If a machine can generate convincing text, images, voices and video, then lying gets cheap. Scam calls using a cloned voice of a relative are already happening. So are fake videos of politicians. Society hasn’t caught up.

Bias, which I covered above, matters most in high-stakes decisions: loans, policing, healthcare. Privacy matters because these systems are trained on huge amounts of data, and whether “posted online” means “fair to use for training” is being fought over in court right now. Electricity is a real issue too. Training and running big models uses a lot of it.

Then there’s the long-term question: as these systems get more capable, can we keep them doing what we actually want? Researchers are split. Hinton, who helped create the technology, quit Google in 2023 partly so he could talk about the risks more freely. Other respected people in the field think the fear is overdone. Nobody knows for sure, and I’d argue that’s a reason to pay attention rather than shrug.

What I'd tell a friend

Two things at once.

One: this is real. It’s not a fad like 3D TVs. Cheap, fast pattern recognition changes how a lot of work gets done, and learning to use these tools is worth your time, the same way learning to use a search engine was.

Two: it’s a tool, with specific strengths and specific ways it fails. It’s great for drafts, summaries, translations, brainstorming, explaining things, and finding patterns in piles of data. It’s bad at facts it wasn’t given, terrible at taking responsibility, and very good at sounding sure when it shouldn’t be. Use it for the first list. Double-check it on the second. Keep your own judgment in charge.

Wrapping up

So, what is AI? It’s software that does jobs we used to think needed a human mind. These days, that almost always means the software learned to do the job from a huge pile of examples rather than from rules someone wrote by hand. It isn’t conscious, isn’t always right, and isn’t new. It just recently got good enough at words and pictures that everybody noticed at the same time.

You don’t need to be afraid of it and you don’t need to worship it. Treat it like any powerful tool: learn what it’s good for, learn where it breaks, and keep your hands on the wheel.

My aunt, for the record, now uses it to plan her garden. She also checks everything it tells her against her seed catalogue. That’s about the right attitude.

FAQ

Is AI the same thing as machine learning?

Not exactly. AI is the big goal of getting machines to do smart things. Machine learning is the main way we get there now, by having the machine learn from examples instead of following handwritten rules. Nearly all modern AI uses machine learning, so people often use the words interchangeably.

Is ChatGPT AI?

Yes. ChatGPT, Claude, Gemini and similar tools are large language models, which are a type of AI. They’re the most visible examples right now, but far from the only kind.

Can AI think or feel?

There’s no evidence it does. It produces things that look like thinking because it learned from an enormous amount of human writing. But nothing about how it works suggests it experiences anything.

Will AI take my job?

Depends on the job. Work that’s mostly routine reading, writing and sorting is the most exposed. Work that needs you to be physically present, take responsibility, build relationships or make judgment calls with incomplete information is less so. For most people, the tools will change the job before they remove it.

Is AI dangerous?

It can be, in ways that are already here (scams, fakes, biased decisions) and in ways people argue about for the longer term. The sensible response is the same as with any powerful technology: understand it, put rules around the harmful uses, and keep humans responsible for the decisions that matter.

What’s the difference between AI and automation?

Automation follows fixed instructions to do the same thing over and over. AI handles jobs where the instructions can’t be fully written out ahead of time, so the system has to make judgment calls based on patterns. A conveyor belt is automation. A camera that inspects what’s on the belt for defects is AI.