AI glossary in plain language: the terms you actually need

The basics: AI, machine learning, and generative AI

Let's start at the root of the tree. Almost everything else hangs off these four terms, and understanding them gives you the full map: AI is the field, machine learning is the dominant method, deep learning is the powerful version of that method, and generative AI is the branch that exploded in 2022 and is the reason you're reading this.

• Artificial intelligence (AI): the general field — software that does tasks that used to require human intelligence: understanding language, recognizing images, making decisions. It's the umbrella, not one specific technology.
• Machine learning: instead of hand-coding rules, you show a program thousands of examples so it learns the patterns on its own. It's how your email learns what spam looks like without anyone writing the list of forbidden words.
• Deep learning: machine learning using neural networks — mathematical structures (very loosely) inspired by the brain, with many layers. It's what sits underneath almost all modern AI.
• Generative AI: the branch that doesn't just classify or predict, but creates new content: text, images, audio, video, code. ChatGPT, Gemini, and Claude are generative AI; your email's spam filter is not.
• Model: the result of training — the already-trained "brain" that answers you. When people say "a new model came out," they mean a new version of that brain, not a different app.

The everyday words: prompt, token, context, and hallucination

These are the terms you run into as soon as you start using an AI assistant seriously. Mastering them isn't optional: they explain why the tool sometimes answers brilliantly and sometimes makes up facts with a completely straight face.

• Prompt: the instruction you write to the AI. The better the prompt, the better the answer.
• Prompt engineering: the craft of writing prompts that get better results: giving context, examples, and the desired format instead of a loose one-liner. It's not magic; it's being specific.
• Token: the unit the model splits text into — roughly chunks of words. It matters because AI services charge and set limits by token: when a tool talks about usage limits, it's almost always counting tokens.
• Context window: how much text the model can "hold in mind" at once — your conversation, your attached documents, its own replies. If the conversation runs too long, the oldest part falls out: that's why the assistant sometimes "forgets" what you said at the beginning.
• Hallucination: when the model makes up false information with total confidence — citations that don't exist, prices nobody published, imaginary laws. It's not a rare glitch: it's a consequence of how these systems work, and the reason every important fact gets verified before you use it.
• Inference: the moment the already-trained model generates an answer. Training is extremely expensive and done by few companies; inference is what happens every time you type.

The terms you'll be sold: LLM, RAG, fine-tuning, and agents

Now we enter commercial-proposal territory. These terms are legitimate and useful — but they're also the ones most used to inflate budgets. If an agency quotes you something using these words, now you'll know exactly what's being offered.

A nuance worth real money: RAG and fine-tuning solve different problems. RAG gives the model access to your information (your catalog, your policies) so it answers with your own, up-to-date data — Google Cloud explains it as connecting the model to external sources before it answers. Fine-tuning retrains the model to change how it behaves — tone, format, specialty. For most small businesses, RAG (or simply attaching documents to your assistant) is the cheap path; fine-tuning is almost never the first step.

• LLM (large language model): the type of model behind ChatGPT, Gemini, and Claude — trained on enormous amounts of text to understand and generate language. ChatGPT is not an LLM: it's the app; the LLM is the engine inside.
• Parameters: the "internal settings" the model learned during training — counted in billions. More parameters usually means more capability, but also more cost; don't buy the number as a sales argument.
• RAG (retrieval-augmented generation): a technique where the model searches your documents first and answers based on what it found, reducing made-up answers. It's what's usually behind "a chatbot trained on your information" — which, most of the time, involves no training at all.
• Fine-tuning: retraining an existing model with your own examples to specialize it. Useful in very specific cases; expensive and hard to maintain. Be suspicious if it's offered as the first step.
• AI agent: a program that receives a goal and decides on its own what steps to take using tools (email, calendar, the web). It's the buzzword of the moment and the one dragging the most hype — we wrote a full jargon-free guide about it.
• Multimodal: a model that works with several formats at once — text, images, audio, video. It's what lets you send a photo of an invoice and ask what it says.
• API: the "technical door" through which one program uses another service. When a tool says "we connect to the WhatsApp API," it means an official, code-based integration.

The security terms: deepfakes, cloned voices, and training data

This part of the vocabulary isn't about buying smarter: it's about not getting robbed. Scammers adopted AI faster than many businesses did, and knowing these terms is the first line of defense. If you only keep one section of this glossary, make it this one — then read the full guide on AI scams.

• Deepfake: a fake video, image, or audio clip generated with AI that imitates a real person — their face, voice, gestures. Already used in "relative in trouble" scams and fake celebrity investment videos.
• Voice cloning: with seconds of public audio (a video, a voice note), someone can generate a near-identical copy of a person's voice. That's why the golden rule is to verify through another channel before moving money, even if "the voice" sounds like your kid's.
• AI phishing: the same old scam emails and messages, but now well written, personalized, and without the spelling mistakes that used to give them away. Bad writing is no longer a reliable signal of anything.
• Training data: the information the model was built from. It matters two ways: the model only "knows" up to where its data reaches, and what you type into free tools may — depending on the service and its settings — be used for training. Check each tool's privacy options before pasting sensitive client information.
• AI watermarking / labeling: signals (visible or hidden) indicating content was AI-generated. Big platforms are adopting them, but at the time of writing they're neither universal nor foolproof: the absence of a label doesn't prove something is real.

How to use this glossary so nobody sells you hot air

Vocabulary is self-defense. When a vendor drops terms from this list, use them right back: ask whether "the chatbot trained on your information" is actually RAG or real fine-tuning (and why), ask what happens when the system hallucinates and who answers for the mistake, and ask for the total monthly cost in tokens or usage, not just the implementation fee.

And a final word about the noisiest term of all: AGI (artificial general intelligence) — a hypothetical AI capable of any human intellectual task. It doesn't exist today, there's no consensus on when it will, and no product you can buy has it. If a commercial proposal uses it as a selling point, you know who you're dealing with.

If you want to go from vocabulary to practice, the natural next step is our guide to free AI courses in Spanish and the free tools that work well in Spanish. With this glossary in hand, you'll understand twice as much in half the time.