AI is a broad name for computer software that can recognize patterns and use them to make a prediction or create something new. Chatbots, image generators, recommendation systems, and voice assistants are all examples.
A model is the part of an AI system that has learned from examples. Training changes millions or billions of numbers inside the model until it becomes good at finding particular patterns. The finished model uses those learned numbers to turn something you provide into a result.
What you provideYour question, image, audio, or data
->
The AI modelPatterns learned from many examples
->
What it returnsAn answer, prediction, image, or action
The technical name for using a trained model to produce a result is "inference."
First, the company trains the model using many examples. Later, you use the trained model by asking a question or giving it a file. A company may keep what you provide and use it as a future training example, but that is a separate decision.
02
A chatbot builds its reply one small piece at a time#
The main AI system inside a chatbot is called a large language model. It repeatedly predicts which small piece of text should come next. Because it has learned many patterns in language, the finished reply can sound thoughtful even when a fact inside it is wrong.
A chatbot can explain ideas, copy styles, combine information, and work through a problem. It does not have human experience. Unless the product clearly says it searched reliable sources, do not assume it checked the claims in its answer.
!Facts that matter: names, dates, quotes, statistics, citations, and current information.
!Advice where a mistake could cause harm:medical, legal, financial, safety, or mental-health decisions.
!Decisions about people: hiring, grading, discipline, counseling, or access to services.
!Private material: names, contact details, account numbers, medical information, private messages, or anything an employer or organization requires you to protect.
!Personal messages and decisions: an apology, condolence, recommendation, evaluation, or other task where your own judgment and care are the point.
What it needs to knowGive only the background needed for this task.
LimitsName the audience, tone, length, and limits.
FormatDescribe the form you want back.
"Turn these notes into a one-page checklist for new volunteers. Use plain language, preserve every date, flag anything unclear, and do not add facts that are not in the notes."
Then ask it to identify weak points: "Which parts of this answer are you least certain about?" and "What should I check before I use this?"
An AI feature can only work with information it can access. That information may include a question, email thread, meeting recording, document, photo, location, or account history. Some work happens on your device. Other work happens on computers run by the company.
The work happens on your phone or computer. This can keep information from leaving the device, but some features send harder requests to the company's computers.
2
Sent to a server to answer you
A server is a computer run by the company, usually in a data center somewhere else. Your content leaves your device and may be stored for chat history, security reviews, or legal requirements.
3
Used to improve future models
Some consumer services use chats or feedback as training examples. A setting may let you stop this for future conversations.
4
Connected to a profile or other apps
The service may remember details about you or read from connected email, files, photos, or other accounts. It may use this information to change later answers, recommendations, or advertising.
Can the company use this conversation as an example when it builds a later version? Turning this off usually affects future conversations. It cannot remove information from a model that has already been built.
Is the conversation saved? Who can see it? When is it deleted? A mode called "temporary" or "private" may still keep a copy for a short time for security checks.
Does the assistant use details from one conversation in later conversations? Turning off this feature and deleting details already saved may require two separate actions.
Menu names and available controls can change. These links go to instructions or privacy information published by the company that makes each product. Work and school accounts may be controlled by an administrator.
AI can help a scammer write a believable message, copy a familiar voice, or create a convincing video. Hearing or seeing someone on a call is no longer enough to prove who they are.
Stop
If a request is urgent, secret, or financially unusual, end the conversation and contact the person or organization another way.
End the conversation. Hang up or stop replying. Do not use a number or link they gave you.
2
Contact them another way. Call a number already in your contacts or printed on an official statement or card. Do not use a number supplied by the person who contacted you.
3
Ask for information a stranger would not know.Use a family verification phrase or a question whose answer is not available online.
4
Ask someone else to check. Tell a family member, coworker, bank employee, or other trusted person what happened before you send money or information.
Start with the bank, card company, payment service, or account that was affected. Contact it through its official app, website, statement, or the number printed on your card. Do not use a phone number or link from the suspicious message.
An AI system that can carry out a series of steps instead of giving one answer. It may use other software to search the web, read files, run code, or send messages. Because it can take actions, a person should control what it can access and check what it does.
A general name for computer systems that recognize patterns and use them to perform tasks. Examples include recognizing a face in a photo, turning speech into text, recommending a song, or answering a question.
The model's short-term working space. It limits how much of a conversation, document, or set of instructions the model can consider at one time. If the limit is exceeded, some earlier material may be left out or summarized.
Fake or altered audio, video, or images made to imitate a real person or event. AI can now create these during a live call, so seeing or hearing someone is not enough to confirm who they are.
The common term for an AI response that is false or made up. The answer may still be detailed and confident because the AI is trying to produce likely language, not prove that each statement is true.
The technical word for the moment when an AI uses what it learned during training to produce a result. When you ask a chatbot a question and it writes an answer, that is inference.
A type of AI model trained to work with language. It writes a response by repeatedly predicting which small piece of text should come next. Chatbots such as ChatGPT, Claude, Gemini, and Copilot use large language models.
The part of an AI system that has learned patterns from examples. Training adjusts a large set of numbers inside the model. The finished model uses those numbers to turn a question, image, sound, or other material into a result.
Able to work with more than one kind of material. For example, the same AI may be able to read text, discuss a photo, listen to speech, and reply with text or audio.
Anything you give an AI system so it can respond. A prompt may include a question, instructions, background information, a document, an image, or several of these together.
A small piece of text used by a language model. One token may be a whole word, part of a word, a punctuation mark, or a space. The model reads and writes text as tokens rather than as complete sentences.
The process used to build a model. The AI works through many examples, makes a guess, compares the guess with the expected result, and adjusts its internal numbers. Repeating this process helps it become better at a task.
About this guide
This material is intended as a general reference.#
AI products, settings, and policies change frequently. Check the current documentation for a product before making a decision that depends on its features, privacy terms, or security practices.