What is it?
"Artificial intelligence" is an umbrella term, not a single technology. It covers any software built to perform tasks that normally need human-like judgement: recognising a face in a photo, translating a sentence, recommending a film, flagging a fraudulent payment, or holding a conversation.
The terms nest inside each other:
- AI: the whole field.
- Machine learning (ML): the part of AI where a system learns its behaviour from examples rather than being programmed rule by rule.
- Deep learning: ML that uses large layered neural networks. It powers most of today's headline results in vision, speech and language.
- Generative AI: deep-learning models that produce new content — text, images, audio, code. See What is Generative AI?.
An important caveat: every system in everyday use today is "narrow". It is good at the task it was trained for and can fail in surprising ways outside it. "AI" does not imply understanding, intent or common sense.
Why does it matter?
AI has moved from research labs into ordinary software. Spam filters, search ranking, voice typing, photo apps, code editors and customer-support tools all use it. Knowing the basics helps you in three practical ways:
- You can judge what a tool is likely to do well (pattern recognition at scale) and badly (guaranteed facts, rare edge cases).
- You can ask better questions of vendors and colleagues: What data was it trained on? How is it evaluated? What happens when it is wrong?
- You can avoid both extremes — treating AI as magic, or dismissing it as a toy.
How does it work?
The classic way to write software is to hand the computer explicit rules: if the email contains these words, mark it as spam. That breaks down when rules are too numerous or too fuzzy to write down, such as "what does a cat look like?"
Machine learning flips the process:
- Collect examples — for instance thousands of emails labelled spam / not spam.
- Train — an algorithm adjusts millions (or billions) of numeric parameters so the model's outputs match the examples as closely as possible.
- Evaluate — test the model on examples it has not seen, to measure how well it generalises.
- Use it (inference) — feed new inputs to the trained model and read its output.
The result is a model: a large file of numbers (parameters, also called weights) plus the code to run it. It has no stored list of rules; its behaviour is spread across those numbers. That is why ML systems can be powerful and also hard to explain or debug.
Example
Suppose you want to detect support tickets that mention a refund.
- Rule-based approach: search for the word "refund". It misses "I want my money back" and wrongly matches "no refund needed".
- Machine-learning approach: train a classifier on several thousand tickets labelled by a human. It learns that phrases like "money back", "charged twice" and "cancel and reimburse" tend to co-occur with refund requests — including phrasings nobody listed.
The ML version is more flexible, but it will still make mistakes, and you need a held-out set of labelled tickets to measure how often.
When should I use it?
AI/ML is a good fit when:
- the task is about patterns in messy data (text, images, audio, behaviour);
- the rules are too many or too fuzzy to write by hand;
- an occasional error is tolerable, or a human reviews the output.
A plain program is usually better when the rules are clear and exact (tax calculation, permissions, unit conversion), when you need guaranteed repeatable results, or when an error is unacceptable and cannot be checked.
Common mistakes
- Treating "AI" as one thing. A recommender, an image classifier and a chatbot are very different systems with different failure modes.
- Assuming it understands. Models find statistical patterns; they can sound confident while being wrong (see AI Hallucinations: Why Models Make Things Up).
- Skipping evaluation. "It looked good in the demo" is not evidence. See How to Evaluate AI Systems and Outputs.
- Using AI where a rule would do. A deterministic function is cheaper, faster and fully predictable.