AI vocabulary in four words: token, context, prompt, hallucination

The vocabulary of generative AI fits in four words: token, context, prompt, hallucination. These are the four notions that explain why a tool forgets what you told it, why it bills you by volume, why the same question gives two different answers, and why an impeccably written answer can be entirely false. Everything else in the jargon is either a variant of these four words or a brand name that will have changed within a year.

This guide is written for a professional who uses an AI tool without making it their job. It deliberately cites no product name and no version number: what ages quickly has no place in a definition.

Why are four words enough?

Because the questions users actually ask always come down to one of these four mechanisms. “Why did it forget the table I gave it at the start?” is a question of context. “Why does it cost me so much?” is a question of tokens. “Why doesn't it do what I want?” is a question of prompt. “Why does it invent references?” is a question of hallucination.

Learning twenty acronyms does not make you a better user. Understanding these four mechanisms immediately changes the way you work with the tool.

What is a token?

A token is the elementary piece of text that the model handles. It is neither a letter nor exactly a word: it is a fragment, often a syllable or a root. A common word usually fits in a single token; a rare word, a proper name or a technical term is split into several.

Three practical consequences:

• Billing is per token, on input as well as output: a long document is expensive to read.
• Languages other than English, such as French with its accents and verb forms, are split into more tokens than English for the same content.
• The orders of magnitude that circulate depend on each tool's tokeniser. Do not copy any of them: use the counter your supplier provides if you need to budget.

Simply remember that everything you send is counted, including the forty-page document you paste in “just in case”.

What is the context?

The context (the context window) is the amount of text the model can have in front of it at any one moment: your instruction, the attached documents, and the whole history of the current conversation.

The most accurate image is a workbench. You lay an instruction on it, then a document, a correction, a follow-up. When the bench is full, what was laid down first falls to the floor. The model does not warn you: it keeps answering, it simply no longer sees the beginning.

From this follows the most profitable habit in the whole practice: new subject, new conversation. A long conversation that has drifted across three different matters produces worse answers than a fresh conversation in which you paste back only the three elements that are actually useful.

Two useful clarifications. First, the context is not a lasting memory: unless a dedicated feature exists, the tool remembers nothing from one conversation to the next. Second, a very large context is not an invitation to pour everything into it: the more off-topic material you add, the more you dilute what matters.

What is a prompt?

The prompt is the instruction you write. The word has become standard; you can say “instruction” without losing anything.

A useful prompt contains four elements, and there is no magic formula beyond them:

1. The role and the audience: who you are writing for, with what level of detail.
2. The task, stated with a verb: summarise, compare, reword, check, list.
3. The material: the text, the figures, the real constraints.
4. The expected form: length, structure, tone, what not to do.

A frequent mistake is to ask for validation rather than examination. “Confirm that this contract is fine” steers the answer; “what is wrong with this contract?” produces something you can use. It is the same flaw as in an interview: the question closes the answer.

To go further on building job-specific instructions, the course Generative AI at work covers the subject through concrete cases.

What is a hallucination?

A hallucination is a false answer produced with the same aplomb as a correct one: a non-existent reference, an invented date, a legal article that was never written, a plausible but false figure.

It is not a lie, and the nuance is decisive. A language model produces the most probable continuation of the text; it has no mechanism that would make it prefer “I don't know”. It does not lie: it does not know that it does not know.

The risk is not theoretical. In June 2023, a federal court in New York sanctioned lawyers who had filed a brief citing non-existent court decisions, produced by a generative AI tool. The case is public and documented; it illustrates exactly what an unchecked, credible answer produces.

Two signals should trigger a check:

• Precision. The more precise an answer is (a number, a date, a reference), the more confidence it inspires, and the more it needs checking.
• Stability. Ask the same question again in a fresh conversation: what comes back identical is generally solid, what appears only once needs checking.

The full method is detailed in our guide how to check an AI answer.

Key points

• The token is the unit into which text is split, and the unit of billing.
• The context is like a workbench: once it is full, what was laid down first falls off.
• When the subject changes, open a fresh conversation rather than continuing the old one.
• The prompt states the role, the task, the material and the expected form.
• A hallucination is a confident error, not a lie.
• The more precise an answer is, the more it deserves a check.

What about the other words you hear?

Three families come up often, and each can be described in a sentence. A language model is the program that produces text token by token. An agent is an arrangement that chains several calls to a model in order to pursue a goal instead of answering a question. The privacy settings decide whether your exchanges are used to train the service: it is a setting, and it is the subject of the guide what data can you share with an AI.

The rest of the vocabulary (model names, version numbers, menu labels) changes several times a year. Describe the action and the intention, not the name of the button: your internal memo will last longer.

Where does this vocabulary apply in your job?

The same four words play out differently depending on what you handle. A teacher who has an assignment reworded, a tradesperson who has a quote drafted, a healthcare worker who has a report summarised: all of them run into the saturated context, the cost per token and the confident but false answer. The only thing that changes is what you absolutely must not get wrong.

That is why skillrung's artificial intelligence topic treats these mechanisms with cases adapted to your job rather than with generic examples.