What is artificial intelligence

The term is wider than a chat in your browser#

The field with this name appeared in the middle of the last century and for a long time had nothing to do with chat. It included chess programs, route planning systems, recognising handwritten digits, and expert systems built on "if — then" rules. All these tasks shared one thing: it was not obvious in advance how to write down the solution as an ordinary algorithm.

Historically there were two approaches. The first was to write down human knowledge as rules and let the program apply them. The second was not to write rules at all, but to give the program many examples and let it find the pattern itself. The second approach is called machine learning, and it is what produced everything you see today.

Layers that people constantly mix up#

It is easiest to think of nested boxes:

• Artificial intelligence — the whole field, including old rule-based systems.
• Machine learning — a way to get a program from data rather than from a programmer's instructions. It includes simple methods such as linear regression and decision trees.
• Neural networks — one of the tools of machine learning, and the most visible one right now. Deep learning means neural networks with many layers.
• Large language models — a particular kind of neural network trained on text. They are what sits behind the chat apps.

When the news says "AI has learned to…", it almost always means the innermost box — a specific trained model. The inner boxes are explained in more detail in what is an LLM and how neural networks work.

Training and running are different things#

This distinction explains half of all misunderstandings. Training is a long and expensive process: the model looks at examples many times and adjusts its internal numbers so that it makes fewer mistakes. Running a finished model — also called inference — is a single pass of data through numbers that are already fixed.

Some consequences worth remembering follow from this.

A finished model does not learn from your conversation. What you tell it does not become part of its knowledge: in the next chat it will remember nothing, unless the developer has attached separate memory storage. That does not mean your conversation disappears — the service still has it.

A model's knowledge is limited to what was in its training data. About events after that point it either says nothing or makes things up. Connecting search or a document collection to a finished model is a separate technique called RAG (retrieval-augmented generation): ordinary search finds the relevant pieces of text and they are added to the model's request. The model itself stays the same; only what it sees before answering changes.

Fine-tuning is the in-between case: you take a trained model and train it further on your own examples. That changes behaviour and style, but it does not turn the model into a reference book.

What happens when a model answers#

To a model, text is a sequence of tokens — pieces of words. The model takes your request, computes probabilities for the next token, picks one, appends it to the text and repeats the whole thing. It has no prepared plan for its answer — the answer is produced piece by piece.

That leads to the key property: the model produces a plausible continuation, not a checked fact. Plausible and true often coincide, but not always, and the model cannot feel the difference.

Limits and risks#

Models are confidently wrong. Made-up facts are called hallucinations: a paper that does not exist, a link to a page that is not there, an exact quote nobody ever said, a legal reference that looks right but is not. Such an answer has exactly the same tone as a correct one — you cannot tell them apart by how they sound.

A model does not know what it does not know. A phrase like "I'm not sure" appears not because the system measured its own ignorance, but because that continuation seemed fitting. You cannot rely on it as a signal.

Anything you send to someone else's service has left your computer. Passport details, client correspondence, other people's medical records and access keys do not belong there — the owner of the service sets the storage terms, not you.

An answer does not replace a specialist. A diagnosis, a legal position, a financial decision belong to a living person with a licence and obligations. Where a mistake costs health, money or rights, a person who is accountable for the decision makes it. A model is accountable for nothing.

And one practical limit: comparisons like "this model is smarter than that one" go out of date within weeks. Treat product names — ChatGPT, Claude, Gemini and the rest — as examples, not as a ranking.

Step-by-step plan

1. Separate the three ideasCheck yourself: if you can explain the difference between the field of AI, machine learning and a neural network, the rest is easier to read.
2. Understand token predictionRemember one mechanism: the model picks the next piece of text by probability, over and over.
3. Tell training from inferencePut into your own words why a model does not remember your chat and does not know recent events.
4. Catch a model making a mistakeAsk about a narrow topic you know well and find an inaccuracy — that is when hallucination stops being theory.
5. Adopt a checking ruleVerify every fact, number or link from an answer against the original source before you use it.

Check yourself

Sources

• Artificial intelligence — WikipediaThe history of the field and its main approaches
  free
• Machine Learning Crash Course by GoogleA free introductory course
  free
• Elements of AIA free online course on what AI is, for people without a technical background
  free

Was this helpful?

Yes No