AI glossary

40 short definitions, no jargon, sorted by topic.

Basics

AI agent: An AI assistant that can carry out several actions in a row to reach a goal: look up information, fill in a document, send a message. Its actions have real effects, which is why human control matters.
AI assistant: A chat application based on a language model, such as Vibe (Mistral AI), ChatGPT, Claude, Gemini or Copilot. You ask it a question or give it a task, and it replies in writing.
Bias: The tendency of an AI to reproduce stereotypes or imbalances found in the data it learned from. Stay alert, especially with texts about people.
Generative AI: Artificial intelligence that can produce new content: text, images, sound or video. It responds to a written or spoken request.
Hallucination: A false answer that an AI presents with confidence: an invented figure, quote or source. It often looks perfectly convincing.
Language model: A program trained on very large amounts of text. It predicts the most likely next words, which lets it write, summarise or translate.
Multimodal: Describes an AI tool that understands or produces several types of content: text, images, sound, video. For example, it can analyse a photo of a document.

Everyday use

AI usage policy: A short document that sets the company’s rules: approved tools, forbidden data, required checks, who to contact in case of doubt.
Automation: Having software carry out a repetitive task, without manual work each time. AI can be added to sort, summarise or write.
Context: The information you give AI so that it understands the situation: for whom, why, with what constraints. Without context, AI fills the gaps with guesses.
Digital sobriety: Using digital tools, including AI, only when they are useful. Each request uses energy: a precise request avoids pointless attempts.
Human oversight: The principle that a person checks the AI’s work and keeps the final decision. It is essential when the decision affects people or commits the company.
Image generation: Creating images with AI from a written description. Check the usage rights and label this content when necessary.
Prompt: The written request you give an AI tool. A good prompt states the goal, the context, the material to use and the expected format.
Shadow AI: Employees using AI tools without the company knowing or having approved them. The main risk: company data sent outside any framework.
Source: The document, website or person that a piece of information comes from. An AI answer without a source you can check is a lead, not a fact.
Speech synthesis: Producing an artificial voice from text. Some AI tools can imitate a real voice, which requires the consent of the person concerned.
Tool register: A list of the AI tools used in the company: what they are for, who uses them, what data goes into them, which contract covers them. It helps you document how you use AI.
Transcription: Automatically turning an audio recording, such as a meeting, into written text. Tell participants before recording.
Web search: A feature that lets an AI assistant look things up on the internet before answering. It is useful for current news, as long as you open the sources it cites.

Data

Anonymisation: Transforming data for good so that it is impossible to identify a person. It takes more than removing the name: a job title or a town can be enough to recognise someone.
Data hosting: The place where the data you give a tool is stored and processed. Hosting in the European Union makes it easier to comply with the GDPR.
Data processing agreement (DPA): A contract between your company and the provider of a tool that processes personal data for you. It sets out what the provider may do with that data, and the GDPR requires it in this case.
Personal data: Any information that identifies a person, directly or indirectly: name, email, phone number, address, customer number, photo, voice.
Pseudonymisation: Replacing identifying details with codes, for example “Client A” instead of a name. The person can still be identified with the list that matches codes to names, so the GDPR still applies.
Transfer outside the EU: Sending personal data to a country outside the European Union. The GDPR requires suitable safeguards for such a transfer.

Regulation

AI Act: The European regulation on artificial intelligence. It sets rules according to the level of risk of each use, and also applies to companies that use AI.
AI literacy (Article 4): An AI Act obligation for companies that use AI: take measures to help their teams understand AI and use it well. Since July 2026, the text has stated that no specific level has to be guaranteed. No certification is required.
CNIL: France’s data protection authority (CNIL). It informs, advises and carries out checks.
GDPR: General Data Protection Regulation. This European law governs how personal data is collected and used, including in AI tools.
High-risk system: An AI Act category for uses of AI that can seriously affect people, for example in recruitment. These systems are subject to stricter rules.
Processor (GDPR): A company that processes personal data on behalf of another, following its instructions. The provider of a professional AI tool often plays this role, which calls for a data processing agreement.
Transparency (Article 50): AI Act rules that require people to be informed, applicable since 2 August 2026. A chatbot must say that it is an AI, and some generated content must be labelled.

Technical

API: A technical access point that lets two pieces of software exchange information. For example, it can connect an AI tool to your management software.
Context window: The amount of text an AI can take into account at once. Beyond that, it may lose track of the start of a long conversation or a long document.
Local hosting: Running an AI tool on your own computers or servers rather than with a provider. The data stays with you, but installation requires technical skills.
Model training: The phase in which an AI learns from large amounts of data. Some tools can reuse your conversations to improve their models: check the setting or the contract.
Open source: Software whose code is published and can be reused under a licence. By extension, AI models are called open when they can be downloaded and installed on your own machines.
Prompt injection: A technique that hides instructions in a document, an email or a web page to hijack an AI assistant. For example, invisible text that tells it to ignore your instructions.
Token: A small piece of text, often part of a word, that AI processes one at a time. Usage limits and some prices are calculated in tokens.

Start learning A term is missing? Ask the chat