AI Glossary: Key Terms Explained in Plain Language

This is NodoAI’s artificial intelligence glossary: the terms you’ll actually run into when using these tools, explained in plain language with examples. No textbook definitions: each entry aims to make the concept click on the first read. If you’d rather start with just the essentials, we have our quick glossary with the 20 key terms; and if you want to learn in order, the route to learn AI from scratch.

Fundamentals

Artificial intelligence (AI)

Software able to do tasks we normally associate with human intelligence: understanding language, recognising images, reasoning or creating content. It doesn’t “think” like a person: it finds patterns in enormous amounts of data.

Machine learning

The branch of AI where the system learns from examples instead of following programmed rules. Show it thousands of cat photos and it learns to recognise cats, without anyone writing down “what a cat is”.

Deep learning

Machine learning with many-layered neural networks. It’s the technique behind almost everything we call AI today: ChatGPT, image generators, translators.

Neural network

A mathematical structure inspired (very loosely) by the brain’s neurons. Millions of connections with “weights” that get adjusted during training until the system gets things right.

Model

The result of training a network with data: the trained “brain” you actually use. GPT, Claude or Gemini are models. When an app says “choose model”, it’s letting you pick which brain answers.

Training

The process of teaching the model with huge amounts of data. It’s extremely expensive and slow; that’s why it happens rarely and the model then “freezes” until the next version.

Inference

What happens every time you use the already-trained model: you give it an input and it produces an output. When you chat with an AI, you’re doing inference.

Parameters

The “internal dials” the model learns during training; they’re counted in billions. More parameters usually means more capability, but also more cost. It’s not the only factor that matters.

Training data

The texts, images or audio the model was taught with. They determine what it knows, what it’s ignorant of, and the biases it carries.

Knowledge cutoff

The date the model’s training data reaches. Ask about anything after it and it either doesn’t know or needs to search the internet. It explains a lot of “out of date” answers.

Language models

LLM (Large Language Model)

A large-scale language model: the technology behind ChatGPT, Claude or Gemini. It predicts the next word so accurately that the result feels like genuine understanding. We explain it in depth in what is an LLM.

Token

The smallest unit the model chops text into; usually a chunk of a word. “Intelligence” may be 2-3 tokens. AI pricing and limits are measured in tokens, which is why the term is worth knowing.

Context window

How much text the model can “hold in mind” at once (measured in tokens). If your conversation or document exceeds it, the model starts forgetting the oldest parts.

Prompt

The instruction or question you give the AI. The quality of the answer depends enormously on the quality of the prompt: context, goal and desired format.

Prompt engineering

The craft of writing prompts that get better results. Less magic than it sounds: it’s about giving context, examples and clear criteria. Full guide in what is prompt engineering.

Hallucination

When the AI states something false with total confidence: it makes up facts, quotes or sources. The most important term in this glossary for using AI wisely: always verify what matters.

Temperature

A setting that controls how much the model “risks” when answering. Low = more predictable, conservative answers; high = more creative and varied (and more prone to rambling).

Fine-tuning

Lightly re-training a model with your own data to specialise it: your brand voice, your industry, your formats. Much cheaper than training from scratch.

RAG (retrieval-augmented generation)

A technique that lets the AI consult your documents or a database before answering, instead of relying on memory alone. Like letting it check the notes: fewer hallucinations about YOUR information.

Embeddings

A numerical representation of a text’s meaning. It lets machines “measure” whether two sentences are about the same thing even if they share no words. The basis of semantic search and RAG.

Multimodal model

A model that understands and combines several formats: text, image, audio, video. You show it a photo of your fridge and it suggests dinner: that’s multimodality.

Reasoning model

A model that “thinks before answering”: it spends time chaining internal steps on complex problems. Slower and pricier, but far better at maths, logic and planning.

SLM (small language model)

A compact model that can run on a phone or laptop without a connection. It trades some capability for privacy, speed and near-zero cost.

Open source vs closed

An open model can be downloaded, modified and run on your own machines (Llama, DeepSeek); a closed one is only used through its creator’s service (GPT, Claude). Control and privacy versus convenience and peak performance.

Agents and tools

Chatbot

A conversation interface with an AI. It answers, but doesn’t act: the key difference from an agent.

AI agent

An AI that doesn’t just answer: it executes multi-step tasks using tools (browsing, writing files, calling APIs) until it meets a goal. The full guide: what are AI agents.

MCP (Model Context Protocol)

An open standard that connects AIs to external apps and data (your email, your database, your tools) in an orderly way. The “USB of AI”. We cover it in what is MCP.

API

The door through which one program talks to another. Companies integrate AI into their products through the model’s API, paying per use (per tokens, usually).

Copilot

An AI embedded inside a tool you already use (code editor, Office, email) to assist you in the flow of work, suggesting or completing as you go.

CLI (command-line interface)

Using a tool by typing commands in the terminal. The most powerful coding agents (Claude Code, Codex CLI, Gemini CLI) work this way.

Automation

Chaining actions so they happen on their own (if X happens, do Y). With AI in the middle, automation stops being rigid: it can read, decide and write at every step.

No-code

Tools for building automations or apps without programming, by connecting visual blocks (Make, n8n, Zapier). AI has given them a second life.

Wrapper

A product that internally just calls another company’s model with an interface on top. Not bad in itself, but it’s worth knowing what you’re paying for: the layer or the model.

Image, audio and video

Generative AI

The family of AI that creates new content (text, image, music, video) instead of just classifying or predicting. This whole glossary exists because of it.

Diffusion model

The dominant technique in image generation: it starts from random noise and “cleans” it step by step until it forms the image you asked for.

Text-to-image / text-to-video

Generating images or video from a written description. “A lighthouse at sunset painted in oils” → image. Quality depends as much on the model as on your description.

TTS (text-to-speech)

Converting written text into natural spoken voice. Today’s voices are so good that the question is no longer “does it sound human?” but “do I have the rights to use it?”.

Voice cloning

Recreating a specific person’s voice from audio samples. Useful with consent (dubbing your own content); a crime or a scam without it. Never clone a voice without permission.

Deepfake

Video, image or audio manipulated with AI to impersonate someone. Just knowing they exist already protects you: learn to detect deepfakes.

Upscaling

Increasing the resolution of an image or video with AI, “inventing” the missing pixels with good judgement. It’s how old photos get restored.

Watermarking / content labelling

Signals (visible or hidden in the file) that identify content as AI-generated. Regulation is pushing towards their widespread use.

Ethics and safety

Bias

Unfair tendencies the model inherits from its training data. If the data carries prejudice, the model reproduces it with an appearance of neutrality. That’s why important decisions aren’t delegated.

Data privacy

What the provider does with what you type: does it store it? Train on it? Practical rule: don’t paste into an AI anything you wouldn’t put in an email to a stranger, unless you know its policy.

AI detector

A tool that tries to identify whether a text or image was AI-generated. They fail more than they promise: use them as a hint, never as proof. We analyse it in how AI detectors work.

Synthetic content

Any content substantially created or altered by AI. The term you’ll see in platform policies and labelling laws.

Risk-based regulation

The dominant legal approach: the more impact a use has, the more obligations it carries (healthcare or credit demand more than a photo filter). It’s the logic behind the new AI rules.

AGI (artificial general intelligence)

A hypothetical AI able to match or surpass a human at any intellectual task. Today it’s equal parts research goal and marketing term; nobody has built it.

How to keep learning

Honest advice: don’t try to memorise this glossary. Come back whenever you run into a term; that’s how they actually stick. The five that cause the most confusion in our experience: token, context window, hallucination, RAG and agent. Master those five and you’ll follow 90% of AI conversations. To go from vocabulary to practice, follow the route to learn AI from scratch.