AI AI basics: what it is and how it got here
What is artificial intelligence? A plain English definition
A clear definition of artificial intelligence, how it differs from ordinary software and automation, and what today's systems can and cannot actually do.
The short answer
- Artificial intelligence is software that gets its behavior from patterns found in data, rather than from rules a programmer wrote out by hand.
- The trade is adaptability in exchange for explainability: an AI system handles cases nobody anticipated, but nobody can point to the line of code that caused a given answer.
- Everything shipping today is narrow AI, meaning it does one kind of task, even when that task is as broad seeming as writing text.
- You already use AI in route time estimates, photo search, predictive text and card fraud alerts, none of which are usually labeled as AI.
- There is no understanding behind the output, only very good pattern matching at enormous scale, which is why confident wrong answers are normal rather than surprising.
Artificial intelligence is software that gets its behavior from patterns in data, rather than from rules a programmer typed out by hand. That is the whole definition, and it is worth reading twice, because it is the only reliable way to tell AI apart from ordinary software that has been given a marketing coat of paint. A chatbot, a face unlock system and a fraud alert on your card are the same idea in different clothes: somebody gathered a large pile of examples, a training program measured the regularities in them, and the result now makes predictions about cases it has never seen.
The spam filter test
Suppose you have to stop junk mail reaching an inbox. There are two ways to build it, and the difference between them is the difference between software and AI.
The first way is rules. You sit down and write conditions. If the subject line contains "free money", add 5 points. If the sender's domain was registered in the last 30 days, add 3. If the message has more than four links and almost no text, add 2. Anything scoring over 8 goes to the junk folder. This is ordinary software, and it has one enormous advantage: it is completely inspectable. If an invoice from a supplier gets blocked, you can read the rules, find the exact line responsible, and change it in a minute. It also has one enormous weakness. Spammers write "fr33 m0ney", your rule stops firing, and you are in a manual arms race that never ends.
The second way is learning. You collect a million messages that people have already sorted into spam and not spam. A training program works through them, measuring how strongly each feature of a message (individual words, formatting, sender history, time of day, hundreds of others) is associated with the spam pile. What comes out is not a list of rules you wrote. It is a large set of numerical weights, tuned automatically, that together score a new message.
That second system is AI. Notice what you gained and what you gave up. It catches "fr33 m0ney" without you lifting a finger, because that spelling appeared in the examples too. But when it wrongly bins a client's email, there is no offending line to find, because there is no line. That trade, adaptability in exchange for explainability, runs through every AI system ever built, and it is structural rather than a matter of company secrecy. The inside of a neural network shows why.
The nesting dolls: AI, machine learning, deep learning
Four words get used interchangeably in headlines and they are not interchangeable. They sit inside one another.
| Term | What it covers | Typical example |
|---|---|---|
| Artificial intelligence | Any software doing a task that seems to need intelligence | The whole field, including old rule based expert systems |
| Machine learning | Systems that derive their behavior from examples | Credit scoring, spam filters, demand forecasting |
| Deep learning | Machine learning using neural networks with many layers | Image recognition, speech to text |
| Large language models | Deep learning applied to text at enormous scale | The chatbot you typed a question into |
Nearly everything sold as AI today lives in the innermost two boxes. When a news story says "AI", it almost always means deep learning. How machine learning actually works covers the middle layer, which is the one worth understanding first, because the failure modes there apply to everything built on top.
Three systems you already use and do not call AI
Route timing. When a maps app says 34 minutes, the shortest path search behind it is a classic algorithm from the 1950s and has nothing to do with AI. The estimate of how long that road takes at 5pm on a rainy Tuesday is learned from years of historical traffic. Most products are a mix like this, part plain code and part learned model.
Photo search. Type "beach" into your phone's photo app and pictures of beaches appear, including ones you never tagged. A model was trained on labeled images until it could associate visual patterns with words.
Predictive text and autocorrect. Your keyboard proposes the next word because a model learned which words typically follow which. This is the same basic mechanism as a chatbot, just much smaller and running on the phone in your hand.
Card fraud alerts and streaming recommendations belong on the list too. None of these are marketed as artificial intelligence, and all of them are.
Narrow AI versus the kind in films
Every system that exists today is narrow AI: it does one kind of task, and it cannot transfer that skill sideways. A chess engine that would beat any human alive cannot read a menu. A medical imaging model that spots fractures cannot tell you the date.
Language models feel like the exception because language touches everything, so a single model can draft a letter, explain photosynthesis and write code. That breadth is real and it is useful. But the task underneath is still one task: produce a plausible continuation of text. The model is not switching between skills, it is doing the same statistical operation on different subject matter. That is also why it will produce an equally fluent answer when it has no idea, a habit covered in why AI makes things up.
The general kind, a system that learns any new task the way a person does, does not exist, and experts disagree sharply about how close it is. What AGI means sets out why those predictions vary so much.
How to judge any AI claim in two minutes
- Ask what it predicts. Every AI system has one output: a category, a number, a next word, a box drawn on an image. If nobody can state the output in a sentence, the claim is vague.
- Ask what it learned from. Source and size of the training data tell you what the system will be good at and where its blind spots are.
- Ask what it does when it is wrong. Does it flag low confidence, or does it answer anyway? Most systems answer anyway, and sound just as certain when they are wrong.
- Ask who checks the output. For anything with money, health or legal consequences attached, a person should still be in the loop.
What to do next
The honest summary is that everything shipping today is pattern matching at scale. That is not a put down. Pattern matching at this scale is genuinely useful and has changed what software can do. It just means you should treat output as a well informed draft rather than an answer, and keep the checking step for anything that matters. If a word in an article blocks you, the AI jargon glossary is built for exactly that.
Common questions
Is automation the same as AI?
No. A rule that moves every email from your bank into a folder is automation, and a person wrote that rule. AI is the case where nobody wrote the rule and the system worked out the pattern from examples instead.
Does AI think or understand anything?
Not in the way you do. A language model picks likely continuations of text based on statistical patterns in what it was trained on. That can look like understanding because the output is fluent, but there is no belief or intention behind it.
Is a chatbot the same thing as AI?
A chatbot is one application of AI, not the whole field. Fraud detection, speech recognition, medical imaging triage and recommendation feeds all use the same underlying approach and have nothing to do with chat.
How can I tell whether a product really uses AI or just says so?
Ask what it learned from and what happens when it meets something unusual. If the vendor can describe the training data and the failure behavior, the claim is probably real. If the answer is only about how smart it is, be skeptical.