AI glossary: 20 key artificial intelligence terms explained

Since 2022, when ChatGPT reached the general public, artificial intelligence has been in every conversation. But when it is time to scope a project, the technical vocabulary quickly becomes a wall. Here is a short guide to help you find your way, without having to become a data scientist.

A brief history of AI in a few dates

The idea of a machine imitating human intelligence came long before modern computers. In 1950, Alan Turing devised his famous test: if a machine can hold a conversation that cannot be told apart from a human's, we can consider that it thinks. Six years later, the Dartmouth conference officially coined the term "artificial intelligence" and launched the discipline.

The decades that followed swung between big promises and disappointment. The expert systems of the 1980s, meant to capture human knowledge as logical rules, fell short on a large scale: this was the "AI winter". The revival came in 2012 with AlexNet, which proved the power of deep neural networks for image recognition. Then the pace picked up: AlphaGo beat the world Go champion in 2016, Google invented the Transformer architecture in 2017, and in 2020 GPT-3 showed that very large models develop unexpected capabilities.

At the end of 2022, ChatGPT put it in everyone's hands. What had been a lab topic reached the general public within a few months. Since then, generative AI has become multimodal (text, image, voice, video), has taken the form of agents able to chain several actions, and has gained its first regulatory framework with the EU AI Act, adopted in 2024. The current stage, already well under way, is its quiet integration into everyday business tools.

20 AI terms you need to know

Here is the vocabulary you will come across most often when discussing an AI project, in no particular order.

AI (Artificial Intelligence)

A field of computer science that simulates human cognitive abilities: reasoning, perception, language. It covers a wide range of techniques, from symbolic logic to deep learning.

Machine learning (ML)

Instead of being programmed rule by rule, the machine learns to recognise patterns from examples.

Deep learning

A subset of ML that uses neural networks with many layers. It is behind the AI revolution of the 2010s.

Neural network

A model loosely inspired by the brain, made of layers of artificial neurons that pass on and transform information until they produce an output.

LLM (Large Language Model)

A language model trained on huge bodies of text. It understands and generates natural language. Examples: GPT, Claude, Gemini, Mistral, Llama.

Transformer

A neural network architecture introduced by Google in 2017. Thanks to its attention mechanism, it underpins almost every modern LLM.

Token

The smallest unit of text an LLM processes, often a word or part of a word. AI API pricing is usually quoted per million tokens.

Prompt

The instruction given to the model. The quality of the answer depends heavily on how clear, precise and well contextualised the prompt is.

Prompt engineering

The craft of writing prompts that get the best results. An emerging role since 2022, somewhere between writing, teaching and engineering.

Hallucination

When a model makes up a false answer and presents it with confidence. It is the main reliability challenge for LLMs in professional use.

RAG (Retrieval-Augmented Generation)

A technique that grounds the model's answers in real documents (yours). It sharply reduces hallucinations and lets the AI cite its sources.

Embedding

A numerical representation (a vector of numbers) of a text, an image or a piece of data. Two things with similar meanings have similar embeddings.

Vector database

A database built to store and search embeddings. A cornerstone of RAG. Examples: Pinecone, Qdrant, Weaviate, pgvector.

Fine-tuning

Further training a pre-trained model on specific data to specialise it in a domain, a style or a particular answer format.

Inference

The stage where the model generates an answer from a prompt. Not to be confused with training, which needs far more computing power.

Computer vision

The branch of AI that interprets images and video: object recognition, detection, segmentation, motion tracking, number plate reading.

NLP (Natural Language Processing)

The automatic processing of language: understanding, generation, translation, summarisation, classification. Anything involving unstructured text.

AI agent

An autonomous system that can chain several steps to complete a task: browsing the web, calling tools, writing code, making decisions.

Multimodal

A model that handles several types of data in the same flow (text, image, audio, video). Almost all modern models are multimodal.

AI Act

An EU regulation coming into force in stages between 2024 and 2027. The world's first comprehensive legal framework for AI, structured around the risk level of each use.

What next?

This vocabulary moves fast. If a term you hear in a meeting seems obscure, ask for a concrete explanation: most AI concepts have a simple definition under the technical wrapping. What matters is not knowing everything, but knowing which questions to ask to scope a useful project.