Key points
- Artificial intelligence is commonly described as technology that has machines make the judgment calls people's minds used to make. There's no single agreed definition, and it isn't the name of one program — it's an umbrella term for that whole category of technology.
- Unlike plain automation, which just follows a fixed sequence, AI usually takes on work that calls for human-style judgment. Where exactly to start calling something AI varies from person to person.
- Machine learning sits inside artificial intelligence, and deep learning sits inside machine learning. The three aren't siblings — they're nested inside one another.
- Most AI that works well today wasn't handed rules by a person. It looked at a huge pile of examples and found the rules itself.
- Most of today's AI is built around a defined range of work. Models that handle many kinds of tasks have appeared, but many people hold that no AI yet handles anything a person can, the way a person does.
The analogy
You've probably stood in front of a washing machine, pressed the single "auto" button, and walked away. Older washing machines couldn't do that. A person had to set the water level, the wash time, and the number of rinses, turning dials by hand. The machine only ever ran exactly what it was set to.
A newer washing machine gives the drum a small spin first, to sense how heavy the load is and how much water it's soaking up. Then it decides the water level and time on its own. A tub stuffed with towels gets a different answer than a few thin shirts. The point where a machine starts making a call a person used to make — that's roughly where artificial intelligence stands. Exactly where to draw the line is something people see differently.
Some washing machines go a step further, scanning tens of thousands of past wash records to work out on their own which cycle suited which kind of load. Artificial intelligence is the name that covers this whole range at once, every stage past the dial where judgment enters in place of a fixed setting.
In detail
Following a fixed order is usually called automation
Automation and artificial intelligence look alike on the surface — both take work off a person's hands. What's often named as the dividing line is whether the answer is already decided. An elevator that gets the third-floor button always stops at the third floor. It's automation, following a sequence someone wrote down in advance.
A washing machine that weighs the load and sets the water level is a little different. Press the same button, and the answer still depends on what's inside. Looking at the situation and deciding in the moment is a job that used to belong to a person. Artificial intelligence is the name for the whole family of technology that moves that job onto a machine.
The line isn't cut with a knife, though. Weighing something and looking up a value in a table counts as automation to some people and as the simplest kind of AI to others; drawing its own standard out of a mountain of cases is usually called artificial intelligence. The word "AI" printed on product boxes these days mixes the two pretty loosely.
Artificial intelligence contains machine learning, which contains deep learning
The three words show up together so often they look like three equal things side by side, but really it's a big bowl with a smaller bowl inside it, and a smaller bowl inside that. The outermost bowl is artificial intelligence — in the broad view, every method for having a machine stand in for human judgment belongs here, and in academic circles it's common to count the old approach of a person writing rules in by hand, one line at a time, as part of it too.
The bowl just inside that is machine learning: instead of a person writing the rules, you show the machine a mountain of examples and let it find the rules itself. The innermost bowl is deep learning, the branch of machine learning that stacks its processing into many deep layers.
Almost everything making news right now comes out of that innermost bowl. That's why the three words end up used as if they meant the same thing, even though plenty of what counts as artificial intelligence isn't machine learning at all.
Writing the rules by hand was the mainstream for a long time
For a long stretch, the mainstream approach in AI research was to write down everything an expert already knew, in as much detail as possible. Fever plus a cough points to a particular illness — that kind of expert judgment got copied over into a list of conditions and outcomes.
This worked fine wherever the rules were clean. The trouble was that most of the world isn't that tidy. Try writing rules for recognizing a photo of a cat and you hit a wall fast. Say the ears are pointed, and a cat with floppy ears shows up. Say it has fur, and a hairless cat shows up.
Learning from examples existed early on too, but as data and computing power grew it became the mainstream. Instead of writing rules, show the machine hundreds of thousands of photos and let it find the pattern on its own — that method is called machine learning. The bundle of rules it finds gets saved under the name model.
For now, many systems are strong at one job
An AI that wins at Go only plays Go. You can't ask that program what's for dinner tonight. An AI built only for translation only translates. Many AIs are built around a single job, and outside that range they don't work well. This kind is called narrow AI.
The line started to look blurry once AI that could hold a conversation showed up — it summarizes, it translates, it writes code, and some models handle images or sound too. Even so, whether it has reached the stage of setting its own goal for a genuinely new problem and figuring out its own method, the way a person does, is something people disagree on.
Which is why you hear both "AI has already surpassed people" and "AI still has a long way to go" at the same time. Inside one narrowed-down job, it's already better than a person; widen the range even a little, and it often gets weaker.
More precisely
Artificial intelligence isn't a term with one agreed-upon definition. Even in academic circles, it's used loosely to cover the entire field of research aimed at getting machines to imitate or stand in for intelligent human activity. What counts as artificial intelligence and what doesn't has kept shifting with the times — a capability that once looked remarkable tends to get demoted to "just software" once it becomes common.
The washing machine analogy breaks down in places too. A washing machine measures one number: weight. AI built through training today is closer to tuning millions to hundreds of billions of dials all at once. And a washing machine's decision can be traced by whoever built it, but with AI that's built through training, often even its own builders can't fully explain why it produced a given answer. The two share only the fact that they stand in for a judgment call — how that judgment actually gets made is much more opaque in the one built through training.
Try it yourself
Common misconceptions
It's easy to think artificial intelligence thinks the way people do, but actually most of today's AI produces answers from patterns learned across many examples. Even when it looks similar, it's hard to say it works the way human thinking does, and whether that can be called "thinking" is something people see in several ways.
It's easy to think artificial intelligence, machine learning, and deep learning are the same thing, but actually artificial intelligence is the widest term, with machine learning nested inside it and deep learning nested inside that.
It's easy to call anything that runs automatically "AI", but actually a machine that only repeats a fixed sequence is usually called automation. Where to start counting something as AI is debated, but whether it takes on work that needs human judgment is a commonly used yardstick.
One-line summary
Artificial intelligence is commonly explained as the umbrella name for every technology that lets machines take over judgment calls people used to make, with machine learning and deep learning nested inside it, one after another.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-18