The Plain-English AI Glossary

You don't need a computer-science degree to understand AI — you need someone to skip the jargon. That's the whole idea behind Ailly, and it's the rule for every definition below: if your neighbor wouldn't get it, we rewrote it.

The basics

AI (artificial intelligence)

Software that can do things we used to think required a human — writing, answering questions, spotting patterns, making decisions. Not a robot brain; a very good pattern-matcher.

Machine learning

The way most AI gets good at something: instead of being programmed with rules, it studies millions of examples and figures out the patterns itself. Like learning to cook by tasting ten thousand dishes rather than reading one recipe.

Model

The actual "brain" you're talking to — the result of all that learning, packed into one program. ChatGPT, Claude, and Gemini are all models (or apps wrapped around one).

LLM (large language model)

A model trained on enormous amounts of text so it can read and write like a person. The "large" refers to how much it studied. When people say "AI" today, they usually mean an LLM.

Chatbot

Any app that lets you talk to an AI in a back-and-forth conversation. The chat window is the chatbot; the model is what's answering.

Generative AI

AI that makes new things — text, images, music, video — rather than just sorting or labeling what already exists. If it created something that didn't exist before, that's generative.

Training

The months-long, wildly expensive process of teaching a model by showing it examples. By the time you use it, training is finished — you're talking to the graduate, not the student.

Training data

The examples a model learned from — books, websites, code, images. A model can only be as good (or as biased) as what it studied.

Neural network

The layered structure inside a model, loosely inspired by how brain cells connect. You'll hear it in headlines; you'll never need to touch one.

Parameters

The millions or billions of tiny dials inside a model that got tuned during training. More parameters usually means a more capable (and more expensive) model — it's the "engine size" number companies brag about.

Talking to AI

Prompt

Whatever you type to the AI. The single biggest thing you control — a clear prompt gets a clear answer, a vague one gets a guess. (It's a skill, and it's learnable — see our beginner's guide.)

Prompt engineering

The fancy name for writing prompts well: giving the AI your goal, the context, and the shape of answer you want. Less "engineering," more "asking good questions."

Context

Everything the AI can currently see: your conversation so far, any files you shared, any instructions you gave. It can't read your mind — if it's not in the context, it doesn't exist.

Context window

The AI's short-term memory limit — how much conversation it can hold in its head at once. Go past it and the earliest parts fall out, which is why very long chats sometimes "forget" the beginning.

Token

The bite-sized chunks AI reads and writes text in — roughly three-quarters of a word each. Pricing and memory limits are measured in tokens, which is why the word shows up on every bill.

System prompt

Standing instructions given to the AI before you ever say hello — its job description. It's why the same model can act like a customer-service agent in one app and a tutor in another.

Role (or persona)

Telling the AI who to be: "act as a patient math tutor," "you're an experienced small-business accountant." A one-line role often improves every answer that follows.

Few-shot (giving examples)

Showing the AI one or two examples of what you want before asking for yours. "Here's a subject line I like — write five more in this style" beats any description of the style.

Temperature

A creativity dial some tools expose. Low temperature: careful, predictable answers. High: more surprising, more risk of nonsense. Most apps set it for you.

What AI gets wrong

Hallucination

When AI states something false with total confidence — a made-up statistic, a court case that doesn't exist, a fake citation. It isn't lying; it's completing a pattern. Rule of thumb: the more checkable a fact needs to be, the more you should check it.

Bias

When a model's answers tilt unfairly because its training data did. It learned from us, and we're not neutral. Good AI teams measure and correct for it; good users stay aware of it.

Knowledge cutoff

The date a model's training data ends — ask about anything after it and the model either doesn't know or guesses. Many tools now bolt on live web search to patch this.

Behind the scenes

GPT

A family of models from OpenAI (the "GPT" in ChatGPT). The letters stand for something technical you'll never need; treat it as a brand name, like Kleenex.

Fine-tuning

Taking a finished model and giving it extra training on specific examples — a company's support replies, a law firm's documents — so it gets unusually good at that one thing.

RAG (retrieval-augmented generation)

A technique where the AI looks things up in a trusted pile of documents before answering, instead of relying on memory. It's how "chat with your company's files" products work — and it cuts down hallucinations.

API

The plug that lets one program use another. When an app "has AI," it usually means it's calling a model's API behind the scenes — the AI lives elsewhere.

Open-source model

A model whose files anyone can download and run on their own computers — free, private, and customizable, though usually needing more technical skill than a polished app.

Multimodal

AI that handles more than text — it can look at photos, hear audio, read documents, or generate images. "Show it a picture of your broken faucet" is multimodal in action.

Agent

AI that doesn't just answer, but does — takes a goal, breaks it into steps, uses tools, and works until it's finished (booking, researching, filling in forms). The word of the moment; the capability is real but young.

Reasoning model

A newer kind of model that "thinks before it speaks" — spending extra time working through a problem step by step before answering. Slower, pricier, and noticeably better at hard problems like math and planning.

Compute

Raw processing power — the expensive specialized chips (GPUs) that training and running AI eats up. When headlines mention billion-dollar data centers, they mean compute.

Safety & trust

Guardrails

The safety rules built into a model — what it will decline to do, what it won't say. Why a chatbot politely refuses some requests.

Deepfake

A fake photo, video, or voice recording made with AI that looks or sounds real. The reason "seeing is believing" now needs a second look — and why unusual money requests deserve a callback on a number you trust.

Data privacy (and training opt-outs)

The question of what happens to what you type. Some tools may use your conversations to improve their models unless you opt out; business plans usually promise they won't. Worth checking once per tool — then you can stop worrying.

Responsible AI

The umbrella term for building and using AI carefully: testing for bias, protecting privacy, keeping a human in charge of decisions that matter. Less a product, more a habit.

AI fluency

Knowing when to use AI, how to ask it well, and how much to trust the answer. It's not coding — it's judgment. It's also the exact skill Ailly is built to grow, a few minutes a day.

Put the words to work

Vocabulary is step one. Next: learn to talk to AI (the four-part prompt recipe), or jump straight to your field — teachers, real estate agents, small business owners.