UNWELLCOLLECTIVE / FREE RESOURCE

AI vocabulary

40 words to find your way.

Web text available in Italian, French and English. PDFs are available only in Italian and French.

01Training

The stage in which a model learns by analysing large amounts of data. It is lengthy and costly and takes place before the model is made available to the public.

02AI agent

AI that does more than answer: it uses tools, opens files, browses the web and takes a sequence of steps to complete a task.

03AI Act

The European Union regulation on artificial intelligence, which entered into force in 2024 with obligations taking effect in stages. It classifies AI systems by risk level.

04Alignment

Work to make AI behave according to human intentions and values: useful, honest and not harmful.

05Hallucination

When AI confidently states something false: an invented quotation, an incorrect figure or a source that does not exist. Always check what matters.

06API

The interface through which a program uses AI without going through a chat. Usage may be billed based on the tokens consumed.

07Benchmark

A standardised test used to compare models. It provides an indication but does not tell you how a model will perform on your particular work.

08Bias

Distortion in AI responses inherited from data or training: stereotypes, overrepresented viewpoints or gaps on certain topics.

09Chatbot

A program you interact with through text or speech. Many current chatbots combine a large language model with a chat interface.

10API key

A secret code that identifies an API user and may determine who is billed. Protect it like a password and never publish it.

11Knowledge cutoff

The date beyond which a model has no information from its training. For recent facts, AI needs to search the web or receive the relevant data.

12Training data

The texts, images or sounds a model learned from. Unbalanced or outdated data can affect its responses.

13Deep learning

Machine learning using neural networks with many layers. It is behind many modern AI systems, including chatbots, translators and image generators.

14Deepfake

A false but realistic image, video or voice generated with AI to make it appear that someone said or did something.

15Embedding

A representation of text as a list of numbers that captures aspects of its meaning. It helps find similar texts even when they use different words.

16Fine-tuning

Additional training of an existing model on specific data to specialise it for a task, style or field.

17Context window

The amount of text a model can consider at once, measured in tokens. Beyond that limit, earlier parts of the conversation may be left out or condensed.

18GPU

The processor used in graphics cards. It performs many calculations in parallel, making it useful for model training and running local models.

19Guardrail

Rules and filters designed to prevent AI from producing certain content or taking certain actions.

20Generative AI

AI that creates new content, such as text, images, music, video or code, rather than only classifying or recognising information.

21Inference

The stage when an already trained model processes input and produces a response. This is the activity typically billed when using an AI API.

22Artificial intelligence (AI)

A collection of techniques that enable computers to perform tasks usually associated with human intelligence, such as understanding text, recognising an image or answering a question.

23Jailbreak

An attempt to bypass an AI system's safety rules using specially crafted prompts.

24LLM (Large Language Model)

A model trained on large amounts of text to predict the next token. This capability supports writing, summarising, translation and conversation.

25Machine learning

The computer learns patterns and makes predictions from many examples instead of relying only on hand-written rules.

26Model

The trained system that performs the task. ChatGPT, Claude and Gemini are services built around one or more models.

27Diffusion model

A technique used by many image generators: it starts from noise and progressively removes it to produce an image guided by the input.

28Local model

A model that runs on your own computer rather than external servers. It can keep processing local and avoid per-request provider charges, but requires suitable hardware.

29Open-weight model

A model whose trained parameters, or weights, can be downloaded so it can run on your own computer. It is not necessarily fully open source.

30Multimodal

A model that understands, and sometimes produces, several content types: text, images, audio or video.

31Parameters

The internal numbers a model adjusts during training. They may number in the billions. More parameters can mean greater capacity but also higher costs.

32Prompt

The instruction you give an AI. Clear context, goals, format and examples can help it produce a useful response.

33System prompt

Background instructions, often not visible to the user, that define an assistant's role, tone and boundaries throughout a conversation.

34Quantisation

A technique that reduces the numerical precision used by a model so it requires fewer computing resources. This can affect output quality.

35RAG (Retrieval-Augmented Generation)

A technique in which AI retrieves information from documents before generating an answer grounded in those sources. It can reduce hallucinations about specific data.

36Neural network

A mathematical structure loosely inspired by the brain: many connected calculations that transform an input into an output.

37Text-to-speech (TTS)

Technology that reads written text aloud using a synthetic voice. Speech-to-text performs the reverse task of transcribing speech.

38Temperature

A setting that influences variation in the model's output. Lower values tend to make outputs more consistent; higher values allow more varied choices. It does not guarantee factual accuracy.

39Token

A unit of text that a model reads and generates, such as a short word, part of a word or punctuation. Model usage and limits are often measured in tokens.

40AI watermark

A hidden signal embedded in generated content to help identify later that it was produced by AI.