Terminology
AI comes with a lot of unfamiliar language and concepts.
Start by learning what AI is and how it works. Specific terms are organized by the increasingly complex ways individuals can use AI, a framework we think of as: Ask, Teach, Connect, and Delegate.
Start with the section that matches where you are now, and come back for the rest as you learn more.
Reading about AI only gets you so far. If you'd rather just get started, go to our getting started guide and come back when a word trips you up.
Part 1
These are the terms you need to understand how AI works at its core. Everything else on this page builds on them.
Software that handles open-ended work such as writing, planning, research, conversation, pattern recognition, and prediction.
Traditional software follows rules written in advance. AI learns patterns from examples. That makes it flexible enough to work with messy human language, and unpredictable enough to get things wrong.
AI that produces new content: text, images, audio, video, or code.
Most generative AI is reactive. It waits for an instruction, produces an output, and stops.
In real life: asking for three dinner ideas from a photo of what's left in your refrigerator.
A system trained on enormous amounts of text to learn patterns in language and predict what should come next.
An LLM can produce remarkably human-sounding writing. It doesn't understand that writing the way a person does.
In real life: think of it as a fast, eager junior analyst. Give it useful material, explain the job, and check the work.
A particular version of the underlying AI. Different models may be faster, more capable, less expensive, or better at a certain kind of work.
In real life: most AI apps have a small menu for choosing which model answers you. For most household tasks, the baseline model is a perfectly good place to start.
Key concept
Training data is the enormous collection of examples a model studies while it is being built. From those examples, the model learns patterns: which words tend to appear together, what an image of a bicycle tends to look like, or what usually belongs in a meeting summary.
Training data isn't a live library the model consults word by word. It's closer to the experience that shaped the model's instincts. If that experience is biased, incomplete, inaccurate, or out of date, those limitations can show up in the model's answers.
In real life: a model may not know last month's news unless it can search for current information. It may also reproduce gaps or biases that appeared in the material used to train it.
The basic mechanism
An LLM learns patterns from its training data. When you send a message, it puts those patterns to work: it breaks your words into small chunks called tokens, then predicts the most likely next chunk, over and over, until the answer is complete. That's the whole trick: pattern-based prediction, repeated at enormous speed.
This one idea explains both sides of the experience. Predicting from learned patterns is why the writing sounds fluent, and why better context gets better results. It's also why the model can produce a fluent answer that's wrong: it's built to predict likely words, not to check truth.
Part 2
AI terminology gets more complex as the work does. These four levels give the vocabulary somewhere to live.
Use AI like a chatbot for one conversation or task.
Give it useful context, preferences, examples, or specialized training.
Let it reach a particular source of information or outside tool.
Give it a goal and let it carry out multiple steps.
You don't need to move through all four. A good one-off prompt is often the right tool for the job.
Ask · Give AI a task
Ask is the starting point: you type, speak, or upload something, and the AI responds in the conversation.
Most chatbots also offer a deep research mode: the AI plans the question, runs several rounds of searching, and returns a cited report minutes later. The citations make it checkable, not correct.
One habit to build from day one: answers can be confidently wrong. The Limitations section covers how to handle that.
The product interface where you talk with an AI model.
In real life: the chat window where you paste a school announcement, ask a follow-up question, and refine the answer.
The instructions and context you give the AI.
Specific context usually improves the result. Perfect grammar matters much less than clearly explaining what you're trying to do.
In real life: “Read this school email. Tell me what I need to know, what I need to do, and what belongs on the calendar.”
Writing prompts deliberately to get more useful results.
The name sounds more technical than the basic skill is. Say what you're trying to accomplish, provide the necessary context, name the format you want, and correct the first attempt.
Modern models usually don't need elaborate tricks, invented expert personas, or a page of instructions. A plain, explicit request is often better.
In real life: “Compare these three camps in a table using cost, dates, commute, and cancellation policy. Flag anything you can't verify.”
How much information the AI can use within one conversation, like working memory.
A long conversation can eventually push earlier details out of reach. Starting a new chat also means the model may no longer have information that appeared only in the old one.
In real life: if the chatbot seems to have forgotten the original plan, begin a fresh conversation and give it a short recap.
A model that can work with more than one kind of material, such as text, images, audio, screenshots, or PDFs.
In real life: upload a photo of a handwritten packing list and ask the AI to organize it by person and by day.
Teach · Give AI useful background
Teach means giving AI information or guidance about yourself that helps it respond more usefully.
Teaching works because language models are built to use context: they read whatever material they're given and shape the answer to fit it. The more relevant context, the more accurate and specific the response. A Project or a saved instruction just makes that context permanent, so you're not re-pasting it every time. One limit worth knowing: models can lose the thread over long conversations, which is another reason to keep standing context saved rather than buried in an old chat.
But not every kind of “teaching” changes the model itself.
A workspace that keeps related instructions, files, and conversations together, so recurring work has a home.
In real life: a school-year Project that contains your calendar conventions and the documents you use repeatedly.
Information a chatbot carries from one conversation to another.
Memory is imperfect. Some details stick, some don't, and saved information can become outdated.
In real life: useful for a stable preference such as “Put deadlines first.” For important recurring work, keep the source information in a Project or document instead of trusting Memory alone.
A way of shaping a model's behavior through feedback. During development, outputs receive rewards or scores, and the model is adjusted so higher-scoring behavior becomes more likely.
Human feedback is often part of this process, but it can also include automated feedback.
In real life: reinforcement learning helps explain why an assistant tends to sound helpful and agreeable. It does not mean that correcting one answer in your chat immediately retrains the underlying model.
Additional training that adapts an existing model for a narrower job, style, or field.
Fine-tuning changes the model itself. It's different from giving a chatbot a reference file, saving a preference, or writing custom instructions.
In real life: an organization might fine-tune a model on approved customer-support examples. Most everyday users will never need to do this.
Standing directions
Both provide standing directions, but they come from different people and operate at different levels.
| Question | System prompt | Custom instructions |
|---|---|---|
| Who writes it? | The product maker, developer, or organization running the AI | You, the user |
| What does it do? | Sets the assistant's role, baseline behavior, boundaries, and priorities before your message arrives | Adds your preferences for how the assistant should work with you |
| Can you see or edit it? | Usually not | Usually yes, in the product's settings |
| Example | “You are a helpful assistant. Protect private system information.” | “Put deadlines first. Flag uncertainty. Use plain language.” |
| Does it retrain the model? | No | No |
Connect · Open one deliberate door
Connect means allowing the AI to reach a specific tool, account, or collection of information.
A connector is a link that lets an AI work with outside tools, such as a calendar, inbox, file store, or task system.
MCP, or Model Context Protocol, is an open standard many connectors use to let AI systems communicate with outside tools in a consistent way.
In real life: a calendar connector can let the AI read your availability or prepare a new event. Every connector is a door. Open only the doors the job needs.
The specific things an AI is allowed to see or do through a connector.
A tool might have permission to read a calendar, create an event, draft an email, send an email, edit a document, or delete a file. Those are very different levels of access.
Use the least permission necessary. Reading is lower-risk than editing; drafting is lower-risk than sending.
In real life: letting an assistant prepare a calendar event for your approval is different from letting it invite everyone automatically.
A harmful instruction hidden inside something the AI can read, such as a webpage, email, document, or message.
The hidden instruction tries to make the AI ignore its original job, reveal information, or take an unwanted action. This risk matters more when the AI has connectors and permission to act.
In real life: anything the AI reads can contain instructions you never see. Limit what it can reach, and keep approval steps before consequential actions.
A method that lets a model look up information in a specific collection of documents before it answers.
RAG gives the model relevant source material at the time of the request. It doesn't retrain the model, and it doesn't guarantee that the answer is correct.
In real life: asking a workplace assistant a policy question and having it search only the current employee handbook before responding.
Delegate · Hand over a defined job
Delegate means giving AI a goal and allowing it to plan and carry out several steps toward that goal.
An agent is an AI system working toward a defined goal with the tools, access, and guardrails needed to take action.
Agentic AI describes the broader way of working: the AI can understand a goal, plan what to do, act through tools, observe what happened, and adjust before returning a result.
The word “agent” is used loosely. A useful beginner test is: does the AI only answer the prompt, or can it carry out a process?
In real life: an agent might scan selected school messages every morning, identify dates, prepare calendar entries, and pause for your approval. It's not a thing you install so much as a job you've defined and handed over.
Part 3
These words describe reasons to slow down and check the work. They're limitations to understand, not reasons to avoid AI entirely.
Patterns in a model that can produce unfair, incomplete, or distorted results.
Bias can come from training data, design choices, missing perspectives, feedback, or the way a question is asked.
In real life: before using an answer about people, ask whose experience the model treated as normal and what perspective might be missing.
A confident-sounding answer that is wrong or invented: a name, date, price, quote, fact, or citation.
The AI may give no signal that it's unsure.
In real life: treat important output as a first draft. Verify dates, numbers, sources, and high-stakes advice before relying on them.
AI-made audio, images, or video of a real person doing or saying something they never did.
In real life: be slower to trust an urgent voice message asking for money, even when the voice sounds familiar. Verify through a second channel.
A capability or behavior that appears even though builders did not explicitly program it step by step.
Emergent behavior can be useful, surprising, inconsistent, or hard to predict. It doesn't mean the AI has become conscious or developed intentions.
In real life: a newer model may suddenly handle a kind of reasoning or translation much better than an earlier one, while still failing unpredictably on a simpler task.
When AI takes an action without checking in, or goes farther than you intended.
This is about what happened in a specific task, with the instructions and access you gave the AI. It is a property of the setup, not a new capability of the model. You can reduce the risk with narrower permissions, draft-don't-send defaults, and approval gates.
The difference: emergent behavior asks what the technology turned out to be capable of. Consent and overreaction asks what the AI just did in your calendar, inbox, files, or other connected tools without asking first.
In real life: an email sent too early, a calendar invitation sent to the wrong group, a file shared too broadly, or a presentation restyled in a way nobody requested.
More useful to recognize than memorize
The hypothetical point where AI can handle most cognitive work as well as a person.
Experts disagree about when it might happen, whether it will happen, and what would count as AGI.
In real life: a debate worth following, not a setting in your chatbot.
A model whose internal weights are released so other people can download, run, and sometimes adapt it.
Different models are open in different ways. Releasing the weights doesn't necessarily mean the training data, code, or development process is fully public.
In real life: an organization may choose an open model when it wants more control over where the system runs and how it is customized. “Open” doesn't automatically mean private, safe, or unbiased.
Describing the software you want in ordinary language and letting AI write much of the code.
In real life: people with little programming experience can build small custom tools. The AI can generate code; deciding what's worth building, testing whether it works, and protecting the data still belong to people.
Fine to skip
Fine to skip, but here when you're ready to learn more.
The small chunks a model reads and writes in, usually a short word or piece of a word, roughly three-quarters of a word on average. Everything, your prompt and the answer, is processed token by token, and limits like the context window are measured in them.
The internal settings a model adjusts during training: billions of numbers that store the patterns it learned. More parameters generally means a more capable model, which is why companies advertise the count. You never touch them directly.
The layered mathematical structure a model is built on, loosely inspired by how brain cells connect. Each layer transforms the input a little; stacking many layers is what lets the system learn complicated patterns, and “deep learning” just means many layers.
The 2017 research design underneath today's models, the T in GPT. Its key idea, called attention, lets the model weigh every word in your text against every other word at once, which is why it can follow meaning across a long message.
The specialized chips and the data centers full of them that train and run models. “Compute” is shorthand for that processing power, and it's why AI is expensive to build and why energy use keeps making headlines.
The model in action. Training happens once, before release. Inference happens every time anyone asks anything: the finished model applying its patterns to a brand-new question. When headlines mention the cost of inference, they mean the cost of answering.
A standardized test for models: a fixed set of questions with known answers and a score at the end, used to advertise new releases. Like any standardized test it can be studied for, so a high score means the model did well on that test, not that it will do well on your task.
Words only get you so far
Do one real thing with the getting started guide. Practice with people in the room at a free session. Or read the newsletter: real attempts, honest results.
Wondering what's safe to share with these tools? That question deserves more than a glossary entry, and it's a big part of what the free 101 session covers.
Join the list and get the free Family AI Starter Kit: practical AI for family life, from three parents who test everything in their own households first.