Generative AI training: understand it, no jargon

In brief

Generative AI is software that produces text, images, sound, or code from a request written in everyday language. It does not look up a ready-made answer: it computes the most likely continuation, based on what it learned from huge amounts of content. That is why it writes very well, and can also state something false with confidence.

What it is, in plain terms

You write a request, and generative AI produces an answer: an email, a summary, a table, an image. ChatGPT, Claude, Gemini, and Microsoft Copilot are assistants built on this technique. You talk to them in plain language, as you would to a colleague. Other tools apply the same technique to images, voice, or video.

At its core, the tool does not consult a directory of answers. The French data protection authority (CNIL) describes it as probabilistic: the system has learned patterns from a mass of text and images, and it produces the most plausible content for your request. Plausible does not mean true. Remember that sentence: it explains both its strengths and its limits.

What it can do at work

The CNIL sorts its uses into three families: creating content from instructions, reworking existing content, and analyzing data. At the office, that means:

Write and rephrase. A delicate email, a job posting, a shorter or firmer version of a text.
Read for you. Summarize a forty-page contract, compare two versions of a document, pull the decisions out of a set of meeting notes.
Search and cite. A web search with the links so you can check: that is the specialty of Perplexity. To ask questions of your own documents, see Gemini Notebook.
Get answers from a spreadsheet. Find a discrepancy, write a formula, draw a chart from an export.
Produce a file or an image. A presentation, a Word document, a visual. For images, see for example Midjourney.
Carry out a multi-step task. You describe the expected result; the assistant plans, searches, then returns a deliverable that you review.

Its limits

Used well, generative AI saves time. Misunderstood, it produces errors that look right. Five limits to keep in mind:

It gets things wrong without warning. A figure, a date, or a quote can be made up and still sound plausible: this is called a hallucination. Check everything that matters.
It does not know your company. It knows only what you give it about your context: the file, the customer, your rules.
It can reproduce biases. The CNIL recommends checking that answers do not contain discriminatory biases.
It does not decide. It prepares, suggests, formats. Judgment, signature, and responsibility stay human.
Your data is not always safe. Depending on the tool, the plan, and the settings, what you write may be kept and used to train the AI. See AI and GDPR at work.

The words to know

The word	What it means
Prompt	What you write to the AI to ask it for something.
Assistant	The app you talk to: ChatGPT, Claude, Gemini, Copilot.
Agent	An assistant that chains several steps on its own to reach a result.
Hallucination	A false answer, given with confidence.
Context	Everything you provide to help the AI: the document, the reader, the tone, the constraints.
Connector	A link between the assistant and your tools (email, calendar, folders), which you allow or remove.

Writing a good prompt

A good prompt reads like a request made to a colleague who does not know the file. It says four things: the context, the reader, the expected result, the format. Add one limit: what the AI must not make up.

An example: “Here are the notes from a customer meeting. Write a follow-up email of 150 words at most, in a friendly tone. Include only the commitments mentioned in the notes. Do not add any discount or date.” Then review, and ask for a precise correction rather than starting over. If you have an example of the expected result, attach it: it is the most direct way to get the right tone.

Learn with Nova

Nova is subscription software made for people who are not specialists. It does not give a course on AI: it shows what AI makes possible in your job, then guides you to the result, with the tool to use, its price, the prompt, and the precautions. You give your level, and the path adapts.

Use cases are sorted by job on AI training by job. The Solo plan costs €19 per month; team plans start at €129 per month. Try the interactive demo or see pricing.

Frequently asked questions

What is the difference between AI and generative AI?

AI is a broad family: systems that infer, from what they receive, predictions, recommendations, decisions, or content. Generative AI is the part of that family that creates content: text, images, sound, code.

Does generative AI understand what it writes?

It does not reason the way a person does. It produces the most plausible answer based on what it has learned. The result is often accurate and well written, but nothing guarantees that it is true: checking is up to you.

Is it free?

To get to know it, yes: ChatGPT, Claude, Gemini, and Copilot each have a free version, with usage limits. Paid individual plans range from €4.99 per month (Google AI Plus, price in France) to $20 per month (ChatGPT Plus, Claude Pro), prices checked on October 3, 2026.

Can you give it company documents?

Not in a free personal account. The CNIL recommends submitting only information you are authorized to share, and checking what the vendor does with the content you enter. Ask your company which tools are approved.

Which tool should you start with?

With a general-purpose assistant: ChatGPT, Claude, Gemini, or Microsoft Copilot. Choose the one your company approves, or the one that fits your software.

Sources

CNIL’s questions and answers on using a generative AI system (July 18, 2024, in French), CNIL, the French data protection authority
AI Act, Article 3: definitions (AI system), European Commission, AI Act Service Desk
Chat, Work, and Codex in ChatGPT, OpenAI
ChatGPT pricing, OpenAI
Claude pricing, Anthropic
Google AI plans (France, in French), Google
Data controls FAQ, OpenAI

Understand generative AI without a word of jargon