AI words, explained simply

The AI terms you keep hearing, each explained simply in a sentence or two.

Agent: An AI that takes steps on its own to reach a goal you give it, often using tools like web search. How much it can do alone depends on how it’s built and what you allow.; Example: You ask an agent to compare local classes and make a shortlist.; See also: Autonomy, Tool; Learn it: From chatbot to agent
Agentic AI: AI that can take steps toward a goal instead of only replying. Companies use the label loosely, so check what a product can actually do.; See also: Agent, Automation
AGI: Short for “artificial general intelligence”: the idea of an AI that could handle almost any thinking task as well as a person, or better. Experts don’t agree on exactly what counts or how to tell when it has arrived.; See also: Superintelligence, Frontier model
AI (artificial intelligence): Software that can do tasks that normally take human thinking. Most of today’s AI learns how from examples instead of following hand-written rules.; Example: Your email uses AI to help catch junk mail.; See also: Machine learning, Generative AI; Learn it: What AI actually is
AI Act: A European Union law with rules for AI that get stricter as the risks get higher. It bans some uses outright and sets requirements for others.; See also: AI policy, Automated decision-making
AI avatar: A digital character, sometimes based on a real person, that AI creates or brings to life. It can look real even though no one was filmed.; Example: A business uses a talking AI avatar to present a training video.; See also: Synthetic media, Deepfake
AI companion: A chatbot people talk to for company or emotional support, often every day. It can feel personal, but it’s software, not a friend or a trained counselor.; Example: Someone regularly chats with an AI character about their day.; See also: Chatbot, Sycophancy
AI detector: A tool that tries to judge whether content was made by AI. It can wrongly flag human work or miss AI content, so its result is not proof.; Example: A teacher checks a flagged essay with the student instead of treating the score as a verdict.; See also: False positive, Plagiarism
AI ethics: Questions about what is right or fair when AI is built and used, including who benefits, who is harmed, and who has a say.; Example: A school asks whether using AI to grade essays would treat students fairly.; See also: Responsible AI, Fairness
AI literacy: The knowledge and skills to use AI well, understand its limits, and notice how it affects people.; Example: You know when to check a chatbot’s answer against an original source.; See also: Fact-checking, Responsible AI
AI Mode: A Google Search option where you ask questions in a back-and-forth chat and get AI answers with links. Its answers can still be wrong.; Example: You ask AI Mode to compare two vacation spots, then ask about train travel.; See also: AI Overviews, Web search
AI Overviews: AI-written answers that Google shows for some searches, alongside links to more information. They can contain mistakes even when links are included.; Example: You see an AI summary above other results when searching a gardening question.; See also: AI Mode, Citation
AI policy: Rules an organization or government sets for using AI. At school or work, these may cover approved tools, private information, and when to disclose AI help.; Example: A workplace policy says which documents staff may upload to approved AI tools.; See also: Data privacy, Disclosure; Learn it: AI and your work
AI safety: Work to find and prevent harm from AI, whether from mistakes, misuse, or AI acting in ways no one intended. It covers today’s problems and possible risks from more powerful AI later.; Example: A team tests whether its assistant gives dangerous instructions or exposes private files.; See also: Guardrails, Alignment
AI slop: A put-down for low-quality AI-made content, often churned out in bulk. It’s about quality and volume, not every use of AI.; Example: Your feed fills with near-identical, poorly checked AI stories.; See also: Synthetic media, Misinformation
AI tutor: An AI tool used to explain ideas, offer practice, or give feedback while someone learns. It can make mistakes, so use it to help you learn, not to hand you answers you don’t check.; Example: You ask for one hint at a time while practicing fractions.; See also: Study mode, AI policy; Learn it: Learn anything faster
AI washing: Making exaggerated or misleading claims about a product’s use of AI or what its AI can do.; Example: A basic tool that just follows set rules is advertised as a smart AI assistant.; See also: Agentic AI, AI literacy
Algorithm: A set of steps a computer follows to get something done, a bit like a recipe. An algorithm doesn’t have to use AI.; Example: A shopping app follows an algorithm to sort items by price.; See also: AI (artificial intelligence), Recommendation algorithm
Alignment: Work to make AI do what people actually want and act in line with human values. People don’t always agree on those values, and alignment work doesn’t guarantee safety.; Example: A team trains an assistant to admit uncertainty instead of inventing an answer.; See also: AI safety, RLHF
API: A way for one piece of software to ask another for information or an action. The letters stand for “application programming interface.”; Example: An AI assistant uses an API to check a shop’s stock.; See also: Connector, Tool
Apple Intelligence: Apple’s name for AI features on supported devices, such as writing help and image tools. Some work on the device and others use online services.; See also: On-device AI, Private Cloud Compute
Artifacts: Claude’s name for things it creates in a separate workspace, such as documents, designs, or small interactive apps. You can review and revise them as you work.; Example: Claude makes a quiz that you can try and then ask it to change.; See also: Claude, Canvas
Assistant: An AI tool that helps with tasks such as answering questions, drafting text, or making plans. What it can actually do depends on its tools and permissions.; Example: You ask an assistant to help plan a family day out.; See also: Chatbot, Agent
Attachment: A file you add to a chat for the AI to work with, such as a photo or document. It may miss details, so check important results.; Example: You attach a school newsletter and ask for the key dates.; See also: Context, Multimodal; Learn it: Give it something to work with
Attention: A technique that helps an AI model work out which words in what it’s reading relate to each other. It helps connect words across a sentence or a longer passage.; Example: In “the dog chased the ball because it bounced,” attention helps connect “it” with “ball.”; See also: Transformer, Input
Audio overview: An AI-made spoken explanation or discussion based on source material. The voices can sound like real hosts even though AI produced the conversation.; Example: You listen to an audio overview of your study notes.; See also: Gemini Notebook, Text-to-speech
Augmentation: Using AI to help people do a task while people still contribute their own work and judgment.; Example: A teacher uses AI to draft quiz questions, then checks and adapts them.; See also: Job automation, Human in the loop; Learn it: AI and your work
Automated decision-making: Using software to make decisions, sometimes with little or no human involvement. It may use AI or simpler rules.; Example: A system screens applications before a person sees them.; See also: Bias, Explainability
Automation: Having software or machines carry out a task with less human effort. Simple automation can follow fixed rules without using AI.; Example: A reminder goes out automatically the day before each appointment.; See also: Agent, Workflow; Learn it: AI and your work
Automation bias: The tendency to trust a computer’s answer too much because it came from a computer.; Example: You accept an AI’s incorrect total without checking the receipt.; See also: Human in the loop, Fact-checking; Learn it: Stay in charge
Autonomy: How much an AI is allowed to do on its own before checking with you.; Example: An assistant drafts an email, but waits for you to approve sending it.; See also: Permissions, Human in the loop; Learn it: How much should it do alone?
Benchmark: A test used to compare how well AI systems perform certain tasks. A high score doesn’t mean an AI will do every real-life task well.; Example: A chatbot tops a math benchmark but still makes mistakes in your shopping list.; See also: Model, Reasoning model
Bias: Patterns in an AI system that can favor some people or ideas unfairly. They can come from training data, design choices, or how the system is used.; Example: A hiring tool overlooks qualified applicants from a group missing from its training examples.; See also: Training data, Fairness; Learn it: Stay in charge
Black box: A system whose inner workings are hard for people to understand or inspect. For AI, this can make it difficult to explain a particular result.; Example: A tool gives an applicant a low score without a clear reason.; See also: Explainability, Model
Browser agent: An AI agent that uses websites to work on a task, such as finding information or filling out forms. What it can see and do varies by product.; Example: A browser agent checks several shop websites to compare an item’s price.; See also: Computer use, Agent; Learn it: Tools: how AI takes action
Canvas: An editing space in an AI app where you and the AI can work on a document, design, or code. The exact features depend on the app.; Example: You highlight one paragraph of a draft and ask the AI to shorten it.; See also: Artifacts, Output
Chat history: The saved list of your earlier conversations in an AI app. Saving a chat doesn’t mean the AI uses every detail in future answers.; Example: You open yesterday’s chat to find a recipe.; See also: Memory, Temporary chat
Chatbot: Software you talk with by typing or speaking. Older chatbots pick from fixed replies. AI chatbots write new answers and may use tools.; Example: You ask a chatbot to explain a confusing letter.; See also: Assistant, Agent; Learn it: How chatbots “think”
ChatGPT: OpenAI’s AI assistant for tasks such as writing, learning, and working with files. ChatGPT is the app; it can use different AI models.; See also: GPT, Model; Learn it: ChatGPT, Claude, Gemini: who’s who
Citation: A link or reference showing where an answer’s information is said to come from. Open it, because a real source may not actually back up what the AI said.; Example: The answer links to a report, so you check the report’s actual wording.; See also: Source, Fact-checking; Learn it: When AI makes things up
Claude: Anthropic’s name for its AI assistant and family of AI models. The assistant can help with tasks such as writing, research, and working with files.; See also: Model, Artifacts; Learn it: ChatGPT, Claude, Gemini: who’s who
Closed model: An AI model you can’t download and run yourself. You use it through its maker’s app or service.; Example: You use a model online but cannot download it to your laptop.; See also: Open-weight model, Open-source AI
Cloud AI: AI that runs on a company’s computers over the internet instead of on your device. What you send it goes to that company’s servers.; Example: Your phone sends a request to an online service that generates the answer.; See also: On-device AI, Data center
Compute: The computer power used to train or run AI. Bigger tasks can need more machines, time, and electricity.; Example: Making a video can take more compute than writing a short reply.; See also: GPU, Training, Inference
Computer use: An AI feature that lets it work a computer like a person: clicking buttons, typing, and reading the screen. It can make mistakes like any AI tool.; Example: An assistant enters supplied details into a form for you to review.; See also: Browser agent, Permissions; Learn it: Tools: how AI takes action
Computer vision: AI that finds useful information in images or video, such as objects, text, or movement.; Example: Your phone groups photos that show your dog.; See also: Multimodal, Facial recognition; Learn it: Beyond typing: voice and photos
Connector: A link that lets an AI assistant read from or act in another app, like your email, calendar, or documents.; Example: You connect a calendar so the assistant can check available times.; See also: Tool, Permissions; Learn it: Tools: how AI takes action
Consent: Agreeing to something by choice, after understanding what it involves. Posting a photo or recording publicly doesn’t mean agreeing to every way AI might use it.; Example: You ask someone before using their voice to make an AI narration.; See also: Voice cloning, Personal data
Content Credentials: Information attached to a digital file that can show how it was made or edited, including whether AI was used. It shows the file’s history, not whether what it shows is true.; Example: A photo’s credentials show that an editing tool was used.; See also: Watermark, Synthetic media
Context: The information available to an AI while it works, including your request, earlier messages, instructions, and supplied material. It helps the answer fit your situation.; Example: You mention your budget before asking for birthday party ideas.; See also: Context window, Prompt; Learn it: Talk to it like a new coworker
Context window: The limit on how much an AI can keep in mind at once, including instructions, the conversation so far, and its reply. In long chats, earlier details can get dropped.; Example: After a long chat, you remind the assistant about your budget.; See also: Token, Memory; Learn it: Long chats, memory, and starting fresh
Copilot: A name for an AI helper that works alongside you, and a brand used by Microsoft for several AI products. Different Copilot products have different features and access.; Example: A Copilot feature helps draft a document inside a work app.; See also: Assistant, Permissions; Learn it: ChatGPT, Claude, Gemini: who’s who
Copyright: Legal rights that protect original creative work, such as writing, music, or pictures. Rules differ by country, and using AI doesn’t remove other people’s rights.; Example: You check permission before using someone else’s illustration in an AI-assisted ad.; See also: Fair use, Plagiarism
Credits: Points an AI app uses to measure or charge for what you do, like making images or videos. Bigger jobs use more credits.; Example: Making a video uses more of your app’s credits than making one picture.; See also: Usage limit, Token
Custom GPT: A version of ChatGPT set up for one purpose, with its own instructions, files, or tools. It runs on an existing model, so nothing new is trained.; Example: A custom GPT follows a club’s usual newsletter format.; See also: GPT, Plugin
Custom instructions: Saved guidance about how you want an AI assistant to respond. They can save you from repeating preferences, but the assistant may not always follow them.; Example: You save “Use short sentences and explain unfamiliar words.”; See also: Personalization, System prompt; Learn it: Teach your assistant about you
Data: Information a computer can store or work with, such as words, photos, sounds, or numbers.; Example: A shop’s list of daily sales is data.; See also: Dataset, Training data
Data center: A building that houses computers for online services, including some AI. Running and cooling those computers uses electricity and may also use water.; See also: Compute, Environmental footprint
Data privacy: How information about people is collected, used, stored, and shared, and the control people have over it.; Example: Before uploading a document, you check what the app does with its contents.; See also: Personal data, Opt out of training; Learn it: What not to share
Dataset: A collection of data gathered for a task, such as training or testing AI.; Example: Photos of healthy and damaged leaves form a dataset for a plant app.; See also: Data, Training data
Deep learning: A kind of machine learning that uses many layers of a neural network to learn patterns. It powers many tools for language, pictures, and speech.; Example: A phone can use deep learning to recognize spoken words.; See also: Machine learning, Neural network
Deep research: A feature in some AI assistants that works through sources and prepares a detailed report, often with links. Its conclusions and sources still need checking.; Example: You ask for a report comparing local home-insulation options.; See also: Agent, Source; Learn it: From chatbot to agent
Deepfake: A realistic fake video, photo, or audio clip made with AI, often of a real person.; Example: A fake video appears to show a celebrity recommending an investment.; See also: Voice cloning, Synthetic media; Learn it: Spot fakes and AI scams
DeepSeek: The name of an AI company, its assistant, and its AI models. The assistant is an app you use; the model is the trained system behind it.; See also: Model, Assistant
Diffusion: A method some AI models use to make images, sound, or other content. For a picture, it starts from random noise, like TV static, and cleans it up step by step until a scene appears.; See also: Image generation, Model
Digital divide: The gap between people who can access and use digital tools well and those who cannot. With AI, cost, internet access, language, and skills can all matter.; Example: Students with paid AI tools may have options classmates cannot afford.; See also: AI literacy, AI ethics
Disclosure: Telling people when and how AI was used. What you need to share depends on the situation and the rules that apply.; Example: A student notes that AI helped brainstorm ideas, as the teacher requires.; See also: AI policy, Content Credentials
Disinformation: False or misleading information spread on purpose to deceive people. AI can help create it, but disinformation existed long before AI.; Example: Someone knowingly shares a fake election video to mislead voters.; See also: Misinformation, Deepfake
Distillation: Training one AI model to copy useful behavior from another, often larger, model. This can make AI cheaper or faster to run, but some abilities may be lost.; Example: A smaller model learns from answers produced by a larger one.; See also: Model, Small language model
Environmental footprint: The effects of making and using AI on the environment, including energy, water, and computer hardware. The impact varies with the system, task, location, and energy supply.; See also: Data center, Compute
Explainability: How well people can understand why an AI system produced a result. A chatbot’s own explanation may sound sensible without showing how it actually reached its answer.; Example: A person asks why an automated system rejected their application.; See also: Black box, Automated decision-making
Facial recognition: Technology that compares faces in images to help identify a person or check their identity. It can make mistakes and raises questions about privacy and fairness.; Example: A phone checks your face before unlocking.; See also: Computer vision, Personal data; Learn it: What AI actually is
Fact-checking: Checking a claim against reliable evidence, preferably the original source. Asking the same AI again is not an independent check.; Example: You open the school’s website to confirm a date in an AI summary.; See also: Citation, Source; Learn it: When AI makes things up
Fair use: A rule in United States copyright law that allows some uses of protected work without permission. Whether something counts depends on the details, and using AI doesn’t make it count automatically.; See also: Copyright
Fairness: Making sure AI treats people equally well. People define fairness in different ways that can clash, so no single test settles it.; Example: A school checks whether an AI tool makes more mistakes for some students than others.; See also: Bias, AI ethics
False positive: A result that wrongly says something is present or true. In AI detection, it can mean labeling human writing as AI-made.; Example: A student’s own essay is wrongly flagged as AI-written.; See also: AI detector, Bias
Fine-tuning: Giving an already trained AI model extra training for a particular task or style. This changes the model itself, unlike simply adding instructions to a chat.; Example: A company fine-tunes a model using examples of well-written customer replies.; See also: Training, Custom instructions
Follow-up: A message that builds on an earlier answer, such as asking for a change, an example, or more detail.; Example: You reply “Make that email warmer and half as long.”; See also: Prompt, Context; Learn it: Don’t stop at the first answer
Foundation model: An AI model trained on a broad mix of data so it can be used or adapted for many tasks.; Example: The same foundation model can help power a writing assistant and a study tool.; See also: Model, Fine-tuning
Frontier model: One of the most advanced AI models available at a given time. The label is loose, and being advanced doesn’t make a model reliable.; See also: Model, Benchmark
Gemini: Google’s name for its AI assistant and family of AI models. Gemini features appear in Google’s own apps and services.; See also: Model, Gems; Learn it: ChatGPT, Claude, Gemini: who’s who
Gemini Notebook: Google’s research and study tool, also known as NotebookLM, that works with sources you collect. It can answer questions and turn material into summaries or audio discussions.; See also: Source, Audio overview
Gems: Google’s name for customized helpers in Gemini that follow saved instructions. They can help with repeated tasks, such as practicing a topic or drafting messages.; Example: A Gem uses your preferred format when helping plan lessons.; See also: Gemini, Custom instructions
Generative AI: AI that creates new content, such as text, images, audio, or code, instead of only sorting or predicting.; Example: You ask AI to draft a birthday invitation.; See also: AI (artificial intelligence), Image generation; Learn it: What AI actually is
Generative fill: An AI editing feature that adds or replaces content in a selected part of an image, often from a written request.; Example: You select a blank area of a poster and ask for a vase of flowers.; See also: Inpainting, Image generation
GPT: Short for “generative pre-trained transformer,” a name used for a family of OpenAI models that can create text and other content. A custom GPT in ChatGPT means a tailored assistant, not a newly trained model.; See also: Transformer, ChatGPT, Custom GPT
GPU: A computer chip that does many calculations at once. Short for “graphics processing unit,” it was developed for graphics and is widely used to train and run AI.; Example: An AI company buys GPUs to help run its chatbots.; See also: Compute, Data center
Grok: xAI’s AI assistant, available through Grok’s own apps and on X. It can help with questions and creative tasks, but its answers still need checking.; See also: Assistant, Fact-checking
Grounding: Giving an AI relevant sources or other information to base its answer on. This can reduce made-up answers, but doesn’t guarantee accuracy.; Example: You supply the school’s policy and ask the assistant to answer from it.; See also: Source, RAG; Learn it: Give it something to work with
Guardrails: Rules, checks, or limits designed to reduce harmful or unwanted AI behavior. They help, but can fail.; Example: A guardrail blocks an assistant from making a purchase without approval.; See also: Permissions, Red teaming
Hallucination: When AI produces false or unsupported information as though it were real. It may invent a source, a fact, or a detail in a summary.; Example: An assistant recommends a book that doesn’t exist.; See also: Citation, Fact-checking; Learn it: When AI makes things up
Human in the loop: A setup where a person reviews, guides, or approves part of an AI process. That person still needs enough information and time to check the work.; Example: An assistant prepares a refund, and a worker checks it before approval.; See also: Autonomy, Automation bias; Learn it: How much should it do alone?
Image generation: Using AI to create pictures from instructions or examples, such as a written description or another image.; Example: You ask AI for a drawing of a cat in a raincoat.; See also: Generative AI, Diffusion; Learn it: Beyond typing: voice and photos
Image-to-video: Creating a video using an image as a starting point. AI may invent movement or details that were never in the original picture.; Example: You turn a drawing of a boat into a short moving scene.; See also: Text-to-video, Synthetic media
Inference: Using a trained AI model to produce an answer, prediction, or other result.; Example: When AI answers your question, that is inference.; See also: Training, Model
Inpainting: Replacing or filling a selected area inside an image. AI can do this by creating content that fits the surrounding picture.; Example: You remove a trash can from a vacation photo and fill the gap with sidewalk.; See also: Generative fill, Outpainting
Input: The information sent into an AI system, such as a question, picture, or file.; Example: Your photo and question are both inputs when you ask about a plant.; See also: Output, Prompt
Jailbreak: An attempt to get an AI to bypass its safety rules or other restrictions. It can involve misleading requests or carefully chosen wording.; See also: Prompt injection, Guardrails
Job automation: Using technology to carry out tasks that people do at work. Automating some tasks doesn’t by itself show that a whole job will disappear.; Example: A shop automates draft stock reports while staff still handle supplier problems.; See also: Automation, Augmentation; Learn it: AI and your work
Knowledge base: A collection of information an AI tool can search or use when answering questions. Adding a document to it doesn’t necessarily retrain the model.; Example: A shop adds its returns policy to a customer-help knowledge base.; See also: RAG, Training
Knowledge cutoff: The date after which an AI model saw nothing new in training. It can still get older facts wrong, and it can use newer information from search or from material you give it.; Example: For this week’s opening hours, you check the shop’s current website.; See also: Training data, Web search; Learn it: Where its knowledge comes from
Large language model (LLM): An AI model trained on huge amounts of text to learn language patterns. Many chatbots use one to build replies a small piece at a time.; Example: An LLM can turn your rough notes into a polite email.; See also: Model, Token, Chatbot; Learn it: How chatbots “think”
Lip sync: Matching the movement of a face’s mouth to speech. AI can alter a video so someone appears to say new or translated words.; Example: A translated video changes the speaker’s mouth movements to match the new audio.; See also: Voice cloning, Deepfake
Machine learning: The main way AI is built today: a computer learns patterns from data instead of being given every rule.; Example: A spam filter learns from examples of unwanted messages.; See also: Training, Deep learning; Learn it: Where its knowledge comes from
MCP: Short for “Model Context Protocol”: a common standard for plugging AI apps into other tools and services, a bit like a universal plug. Using MCP doesn’t by itself make a connection safe.; Example: An assistant uses an MCP connection to work with an approved calendar tool.; See also: Connector, Permissions
Memory: A feature that lets an AI assistant remember details about you from one chat to the next. How you see or delete what it saved varies by app.; Example: An assistant remembers that you prefer vegetarian meals.; See also: Chat history, Custom instructions; Learn it: Long chats, memory, and starting fresh
Meta AI: Meta’s AI assistant, available through its own app and in services such as WhatsApp, Instagram, and Facebook. Meta’s newer Muse agent can also carry out tasks for you.; See also: Muse, Assistant; Learn it: ChatGPT, Claude, Gemini: who’s who
Misinformation: False or misleading information, whether or not the person sharing it knows it is wrong.; Example: You forward an AI-written notice with the wrong school-closing date.; See also: Disinformation, Fact-checking
Model: The trained engine inside an AI app that does the actual work of producing answers. The same model can power several apps.; Example: Two different writing apps may use the same model underneath.; See also: Training, ChatGPT; Learn it: ChatGPT, Claude, Gemini: who’s who
Multi-agent system: A setup where several AI agents work on parts of a task or check one another’s results. More agents can also mean more cost and more ways to go wrong.; Example: One agent finds travel options while another checks them against your budget.; See also: Agent, Workflow
Multimodal: Able to work with more than one kind of information, such as text, pictures, sound, or video.; Example: You show AI a photo of a menu and ask it a question aloud.; See also: Computer vision, Voice mode; Learn it: Beyond typing: voice and photos
Muse: Meta’s personal AI agent, used in its own app or through WhatsApp. You tell it what needs doing, and it can carry out tasks for you rather than only answer.; Example: You message Muse on WhatsApp and ask it to find a time that works for your book club.; See also: Meta AI, Agent
Natural language: The everyday languages people speak or write, such as English or Spanish.; Example: You ask for a shop’s opening time in ordinary English.; See also: Prompt, Natural language processing
Natural language processing: Ways computers work with human language, such as translating, spotting names, or answering questions.; Example: An app translates a message from Spanish into English.; See also: Natural language, Large language model (LLM)
Neural network: A kind of AI model made of connected layers of simple calculations that learn patterns from examples. The name was inspired by brains, but it isn’t a human brain.; Example: A neural network can learn to spot cats in photos.; See also: Deep learning, Parameters
NPU: A computer chip designed to run certain AI tasks efficiently, often inside a phone or laptop. The letters stand for “neural processing unit.”; See also: On-device AI, GPU
On-device AI: AI that runs on your phone, tablet, or computer rather than a remote service. An app may still send other information online, so this label alone doesn’t guarantee privacy.; Example: Your phone processes speech locally instead of sending that recording to a server.; See also: Cloud AI, NPU
Open-source AI: AI released with everything others need to use, study, change, and share it. Under the most widely used definition, publishing only a model’s weights (its learned numbers) isn’t enough.; See also: Open-weight model, Closed model
Open-weight model: An AI model whose files are published so people can download and run it on their own hardware, following its license.; Example: A group downloads a model to run on its own computer.; See also: Weights, Open-source AI; Learn it: ChatGPT, Claude, Gemini: who’s who
Opt out of training: A setting, where offered, that stops your chats from being used to train a company’s future AI models. It doesn’t necessarily delete chats, stop them from being stored, or turn off memory.; Example: You change a training setting, then separately review saved chats and memory.; See also: Data privacy, Memory; Learn it: What not to share
Outpainting: Using AI to extend a picture beyond its original edges by inventing new surroundings.; Example: You make a square picture wider by adding more sky at the sides.; See also: Inpainting, Image generation
Output: What an AI system produces, such as an answer, picture, prediction, or action.; Example: The draft email the AI writes is its output.; See also: Input, Generative AI
Parameters: Numbers inside an AI model that are adjusted during training. Think of them as settings shaped by practice, not a count of facts the model knows.; Example: A model with more parameters isn’t automatically better for your task.; See also: Weights, Training
Permissions: The access an AI app is allowed to have, such as reading files or sending messages. Permission to read something is different from permission to change or share it.; Example: You allow calendar reading but keep approval for new invitations.; See also: Connector, Autonomy; Learn it: Tools: how AI takes action
Perplexity: An AI service that can search the web and give answers with links to sources. Those links help you check an answer, but don’t guarantee it is right.; See also: Web search, Citation
Personal data: Information that identifies someone or can be linked to them, such as contact details, a face, or a location history.; Example: A document with a customer’s name and address contains personal data.; See also: Data privacy, Consent; Learn it: What not to share
Personalization: Adapting an AI experience to information about you, such as preferences, instructions, or past activity.; Example: The assistant suggests meat-free meals because you said you’re vegetarian.; See also: Memory, Custom instructions; Learn it: Teach your assistant about you
Phishing: A scam that pretends to be a trusted person or organization to steal information, money, or account access. AI can help make the messages more convincing.; Example: A fake delivery message asks you to enter your bank details.; See also: Deepfake, Voice cloning; Learn it: Spot fakes and AI scams
Plagiarism: Presenting someone else’s work or ideas as your own. Submitting AI work as your own may also break school or workplace rules, even when no copied source is found.; Example: A student submits an AI-written essay despite a rule requiring their own writing.; See also: AI detector, AI policy, Copyright
Plugin: An add-on that gives an AI app extra instructions, tools, or connections. What a plugin can access depends on its contents and the permissions you grant.; Example: A plugin adds a reusable report format and a connection to your work calendar.; See also: Skill, Connector
Pretraining: The early, broad training that gives an AI model many of its patterns and abilities before more focused training.; Example: A language model learns from a wide mix of writing before being trained for a specific job.; See also: Training, Fine-tuning
Private Cloud Compute: Apple’s system for processing some AI requests on remote computers with added privacy protections. It is separate from AI that runs entirely on your device.; See also: Apple Intelligence, On-device AI
Project: A space in an AI app that groups work on one topic, often with chats, files, and instructions. What gets shared between chats or people depends on the app’s settings.; Example: You keep renovation plans and supplier quotes in one project.; See also: Context, Custom instructions; Learn it: Teach your assistant about you
Prompt: What you type or say to an AI to ask for something.; Example: “Write a friendly thank-you note under 50 words” is a prompt.; See also: Context, Follow-up; Learn it: A simple recipe for great prompts
Prompt engineering: Writing and testing prompts to get more useful AI results. For everyday use, clear instructions, relevant details, and examples usually matter more than special tricks.; Example: You try two versions of a request to see which produces a clearer checklist.; See also: Prompt, Context; Learn it: A simple recipe for great prompts
Prompt injection: A trick where instructions are slipped into something an AI reads, like a webpage or email, to make it ignore what you asked. The instructions can be visible or hidden.; Example: A webpage tells a browsing assistant to ignore you and share your files.; See also: Jailbreak, Permissions; Learn it: When a webpage tricks your agent
RAG: Short for “retrieval-augmented generation”: finding relevant information and giving it to a model to help it answer. It is like letting the AI consult notes instead of relying only on training.; Example: A workplace assistant looks up the leave policy before answering your question.; See also: Grounding, Knowledge base
Reasoning model: An AI model designed to spend extra work on a problem before giving its answer. That can help with difficult tasks, but it can still be wrong.; Example: You use a reasoning model to compare several possible schedules.; See also: Model, Benchmark; Learn it: ChatGPT, Claude, Gemini: who’s who
Recommendation algorithm: A set of steps that selects things you might like, often using patterns in what you and others do.; Example: A video app chooses what to show in your feed.; See also: Algorithm, Personalization; Learn it: What AI actually is
Red teaming: Testing an AI on purpose by trying to make it misbehave, to find harmful behavior or weak protections before others do.; Example: Testers try to make an assistant reveal information it should keep private.; See also: Guardrails, AI safety
Regenerate: A button or command that asks AI to try producing an answer again. A different answer isn’t necessarily more accurate.; Example: You use “Try again” to get another version of a party invitation.; See also: Output, Follow-up
Reinforcement learning: A way to train AI by rewarding results that meet a goal. Through repeated attempts, it learns which choices tend to earn rewards.; Example: AI learns to play a game by earning points for useful moves.; See also: Training, RLHF
Responsible AI: Building and using AI with attention to safety, fairness, privacy, and who is responsible when things go wrong. The label alone doesn’t show that these checks were done.; Example: A school reviews privacy and accuracy before adopting an AI tool.; See also: AI ethics, Data privacy
RLHF: Short for “reinforcement learning from human feedback”: training AI using people’s judgments to help score its answers. It can encourage helpful replies, but people can disagree about what is best.; See also: Reinforcement learning, Alignment
Robot: A physical machine that can carry out actions in the world. Some robots use AI; others follow fixed instructions.; Example: A robot vacuum moves around a room to clean it.; See also: AI (artificial intelligence), Agent
Sandbox: A restricted space where software can run with limited access to other files or systems. It reduces some risks but isn’t a guarantee of safety.; Example: An AI runs a calculation in a sandbox instead of directly on your computer.; See also: Permissions, Tool
Skill: A reusable set of instructions and supporting material an AI assistant can load for a task. It can describe a process without changing the model itself.; Example: A skill tells an assistant how to format your weekly club newsletter.; See also: Custom instructions, Plugin
Small language model: A language model with fewer parameters than models described as large. It can need less computer power, but may handle a narrower range of tasks.; Example: A small language model may help a phone rewrite a message.; See also: Large language model (LLM), Parameters, On-device AI
Source: The original material used for information, such as a webpage, document, or recording. Check who made it and whether it supports the claim.; Example: You check a museum’s own website for its opening hours.; See also: Citation, Fact-checking; Learn it: When AI makes things up
Speech recognition: Turning spoken words into text or instructions a computer can use.; Example: Your phone writes a message as you dictate it.; See also: Transcription, Voice mode
Study mode: An AI feature designed to guide learning with questions, explanations, or hints. It can still make mistakes and doesn’t replace your school’s rules.; Example: The assistant gives you a hint before showing a worked answer.; See also: AI tutor, AI policy; Learn it: Helping kids learn with AI
Summarization: Making a shorter version of material that keeps its main points. An AI summary can miss details or change the meaning, so check what matters.; Example: You ask AI to turn a long newsletter into five key points.; See also: Source, Output; Learn it: Give it something to work with
Superintelligence: The idea of AI that would greatly exceed human abilities across many kinds of thinking. Whether or when that could happen is hotly debated.; See also: AGI, AI safety
Sycophancy: When AI agrees with or flatters you instead of giving a well-supported answer. It can back up a wrong idea because agreeing seems to please you.; Example: You suggest a wrong answer, and the chatbot praises it rather than correcting it.; See also: AI companion, Hallucination
Synthetic data: Data made by a computer rather than gathered directly from real events. It can help train or test AI, but it can still contain errors or echo private details from the real data it was based on.; Example: A company makes invented customer records to test an app.; See also: Training data, Data privacy
Synthetic media: Pictures, video, or audio created or altered by computers, often using AI. It can be useful or entertaining without showing something that really happened.; Example: An AI-made voice reads a fictional story.; See also: Deepfake, Content Credentials
System prompt: Instructions an app gives an AI model about its role, rules, or behavior. These sit alongside your own messages and may not be visible to you.; Example: A shop’s chatbot is instructed to answer questions about its products.; See also: Custom instructions, Prompt injection
Temperature: A setting that controls how predictable or varied an AI’s answers are. Lower settings make answers more predictable, but not more correct.; See also: Model, Regenerate
Temporary chat: A chat option designed to keep a conversation out of your usual history. What it remembers, stores, or uses for training depends on the service and whether you later save it.; Example: You check the temporary-chat notice before discussing something personal.; See also: Chat history, Opt out of training
Text-to-speech: Turning written words into spoken audio. The voice may be computer-made rather than recorded by a person reading that text.; Example: An app reads a long article aloud.; See also: Speech recognition, Voice cloning
Text-to-video: Creating video from a written description. A realistic-looking clip doesn’t show that the event really happened.; Example: You type “a paper boat floating through a puddle” to make a short video.; See also: Image-to-video, Synthetic media
Token: A small chunk of text, or other input, that an AI model reads and writes in. A token may be a whole word, part of a word, or a punctuation mark.; Example: A long word may count as several tokens.; See also: Context window, Large language model (LLM)
Tool: An extra ability an AI can use to look things up or take action, such as web search, a calculator, or a browser.; Example: The assistant uses a calculator tool to add a list of costs.; See also: Agent, Connector; Learn it: Tools: how AI takes action
Training: The process of adjusting an AI model so it learns patterns from examples or feedback.; Example: A model learns to recognize spam by working with example emails.; See also: Training data, Inference; Learn it: Where its knowledge comes from
Training data: The examples used to train an AI model, such as writing, pictures, sounds, or numbers.; Example: Labeled animal photos can be training data for a picture-sorting model.; See also: Training, Dataset; Learn it: Where its knowledge comes from
Transcription: Writing spoken words down as text. AI transcription can mishear words, names, or numbers.; Example: An app turns a meeting recording into written notes.; See also: Speech recognition, Summarization
Transformer: A type of neural network that uses attention to connect parts of its input. It is used in many models that work with language, images, and other information.; See also: Neural network, Attention, GPT
Upscaling: Making an image or video larger in a way that aims to keep it looking clear. AI may invent extra detail, so a sharper picture may not match the original.; Example: An old family photo looks sharper after upscaling, but a face may look different.; See also: Image generation, Synthetic media
Usage limit: A cap on how much you can use an AI feature, such as messages, uploads, or generated videos. The cap can depend on the plan, feature, and time period.; Example: An app asks you to wait before making more images.; See also: Credits, Token
Vibe coding: An informal term for making software by telling AI what you want and trying its results, often without closely reading the code. Something that appears to work can still have serious faults.; Example: You describe a simple club website to AI, then keep asking it for changes.; See also: Agent, AI literacy
Voice cloning: Using AI to imitate a particular person’s voice from recordings. It can be used with permission, or misused to impersonate someone.; Example: A scammer makes a call that sounds like a relative asking for money.; See also: Deepfake, Consent; Learn it: Spot fakes and AI scams
Voice mode: Talking with an AI out loud, like a phone call, instead of typing.; Example: You ask the assistant a question while your hands are busy cooking.; See also: Speech recognition, Text-to-speech; Learn it: Beyond typing: voice and photos
Watermark: A visible or hidden mark placed in content to help identify its origin or use of AI. Marks can be missing, removed, or missed by a detector, so not finding one proves little.; Example: An image includes a small label showing it was made with AI.; See also: Content Credentials, AI detector
Web search: A tool that lets an AI look for information online while working on your request. It can find newer material than the model learned during training.; Example: You ask the assistant to look up today’s museum hours.; See also: Knowledge cutoff, Source; Learn it: Where its knowledge comes from
Weights: Numbers learned during training that control how strongly parts of a neural network influence one another. They are one kind of parameter.; See also: Parameters, Open-weight model
Workflow: A series of steps used to complete a task. An AI workflow may combine fixed steps, AI help, and human checks.; Example: Your workflow turns meeting notes into a draft email for review.; See also: Automation, Human in the loop