The AI words you keep seeing in the news, explained in one line and then in plain English.
An AI that takes actions for you, not just answers.
A chatbot replies with text. An agent can also click, type, browse, send emails or run code to finish a task. That makes it more useful and more risky, because a mistake becomes an action instead of a wrong sentence.
Go deeper →Software that guesses whether text was written by AI.
Detectors look for statistical patterns, such as how predictable each word is. They produce a probability, not proof, and they regularly flag human writing as AI, especially formal or simple writing.
Go deeper →A set of step-by-step rules a computer follows.
Every app runs on algorithms. In AI coverage the word usually means the rules that decide what you see, like a feed ranking, or how a model was trained.
A hypothetical AI as capable as a person across most tasks.
There is no agreed test for AGI and no agreed date. When a company says it is "close to AGI", treat it as a goal or a marketing claim, not a measurement.
Go deeper →A program you talk to in plain language.
ChatGPT, Gemini, Claude, Copilot and Meta AI are chatbots built on large language models. What you type may be stored and, depending on your settings, used to improve the product.
Go deeper →How much text an AI can consider at once.
Think of it as the AI’s short-term memory for one conversation. Once a chat gets longer than the window, the oldest parts drop out and the AI can lose track of earlier details.
A building full of computers that run AI and the internet.
Training and running AI models takes enormous computing power, which uses electricity and, for cooling, water. That is why AI growth shows up in energy and water debates.
Go deeper →Fake video, image or audio made to look real with AI.
Deepfakes can put a real person’s face or voice into something they never did or said. They are used in scams, harassment and misinformation, and they are getting cheaper to make.
Full encyclopedia entry →Extra training that specializes a general AI model.
A company takes a general model and trains it further on its own examples, for instance customer support chats, so it behaves in a specific way.
AI that creates new text, images, audio or video.
Older AI mostly sorted and predicted, like spam filters. Generative AI produces new content in response to a prompt. Chatbots and image generators are both generative AI.
The chip that does most AI computing.
Graphics processing units were built for video games but are very good at the math AI needs. Shortages and prices of GPUs shape which companies can build the biggest models.
When an AI states something false with confidence.
Language models predict likely words, not verified facts, so they can invent quotes, sources, laws or numbers. Always check anything important, such as legal, medical or money questions, against a primary source.
Full encyclopedia entry →The type of AI behind chatbots.
An LLM is trained on huge amounts of text to predict the next word. That simple trick, at very large scale, lets it write, summarize, translate and answer questions.
Full encyclopedia entry →A proposed file that tells AI tools what is on a website.
It works a bit like robots.txt but for AI. It is a proposal, not a standard, and major AI companies have not confirmed using it.
Go deeper →Computers learning patterns from examples instead of fixed rules.
Instead of programming every rule, you show the system lots of examples and it finds the patterns itself. Most modern AI, including chatbots, is built with machine learning.
The trained AI system itself.
The model is the "brain" that a product like ChatGPT runs on. One company can offer several models with different speed, cost and ability, for example GPT, Gemini or Claude versions.
AI that handles more than one kind of input, like text and images.
A multimodal assistant can read a photo of a document, listen to your voice or describe a picture, not just read typed text.
An AI model whose files anyone can download and run.
Some companies publish their model weights so others can run and modify them. That helps research and privacy, since it can run on your own device, but it also means safety controls can be removed.
What you type to an AI.
A prompt is your question or instruction. Clear, specific prompts with context usually get better answers than short vague ones.
Hidden instructions that trick an AI into doing something else.
An attacker hides text in a web page, email or document. When an AI tool reads it, it may follow those instructions instead of yours, for example leaking data. It is one of the main security risks of AI agents.
Full encyclopedia entry →A locked-down space where software runs without touching the rest of your system.
AI tools that run code are often placed in a sandbox so mistakes or attacks stay contained. A "sandbox escape" means the software got out of that box.
Go deeper →The text, images and other material an AI learned from.
What a model knows, and its blind spots and biases, come from its training data. Whether your own chats become training data depends on the product and your settings.
Go deeper →Copying a real person’s voice with AI from a short sample.
A few seconds of audio from social media can be enough to make a convincing copy. Scammers use cloned voices to pose as relatives asking for urgent money.
Full encyclopedia entry →A hidden signal that marks content as AI-generated.
Some AI companies add invisible markers to text, images or audio so it can be identified later. Watermarks can often be weakened or removed, so they help but do not settle the question alone.
Go deeper →