AI Glossary, Explained Simply.

Computer programs that can do work that usually needs a human brain: understanding language, recognising pictures, making decisions, and creating…

AI that doesn't just answer, but can take action by itself to reach a goal, such as opening websites, writing files, running code or sending emails…

The question of who owns content made by AI, and whether its training data violates other people's rights. The law is still developing; in practice…

The principles of using AI responsibly: not deceiving people, not using someone's face or voice without permission, not spreading false information…

AI that creates a human voice from text, or copies a specific voice (voice cloning). The quality is now almost impossible to tell apart from a real…

A way for other programs to "talk" to an AI automatically, without opening a chat. Businesses use APIs to connect AI to their website, WhatsApp or…

Making repetitive work run by itself, with no human stepping in. With AI, automation can now handle work that needs "understanding", not just…

When AI keeps leaning one way or getting things wrong in the same pattern, because its training data is unbalanced. Common examples: a model is more…

Asking AI to show its thinking steps before it answers, or using a "reasoning" model that thinks first. This improves accuracy on maths, logic and…

The ability to keep a character looking the same across many pictures or clips, and the biggest problem in AI video. It's solved with reference…

The maximum amount of text (in tokens) an AI can "remember" in one conversation, including your instructions, the documents you upload and its own…

A collection of data used to train or test AI: text, pictures, video or records. The quality of the dataset decides the quality of the model; garbage…

A branch of machine learning that uses "neural networks" with many layers, inspired by the brain. This is the technology behind today's image…

A fake video or audio clip made with AI to show someone doing or saying something that never happened. It's the same technology as creative AI video…

The technique behind most image and video generators. The model learns to "clean up" a picture full of random specks (noise), little by little, until…

A way of turning text (or pictures) into a list of numbers that represents its meaning, so a computer can find content that is "similar in meaning"…

Giving AI a few examples of the result you want inside the prompt, so it follows the same pattern. It works better than a long explanation.

Training an existing model again with specific data so it gets better at one particular task or style. It's cheaper than training a model from…

AI that creates brand-new pictures from a text description. You write what you want, and it draws it. This is the base of all AI video: usually the…

AI that creates video clips from text or pictures. It can animate a still picture (image-to-video) or create a clip straight from a description…

A computer chip originally made for game graphics, but very good at the parallel calculations AI needs. A shortage of GPUs is why AI capacity is…

Making sure an AI's answer "stands on" real sources (documents, web search or up-to-date data) instead of guesses. It greatly reduces hallucination.

When AI gives an answer that sounds confident but is wrong or made up: names, numbers, sources or facts that don't exist. This is the main reason you…

A picture you give the AI as a guide for look, style or composition. Different tools use different names for it, but the purpose is the same: "make…

A way to make video by giving AI one picture as the starting point, then directing the movement. It's more controlled than text-to-video because the…

The moment a model is "run" to produce an answer, as opposed to training, which happens only once. Every time you press Enter in ChatGPT, that's…

The waiting time between sending an instruction and the AI starting to reply. Bigger models or reasoning models usually have higher latency.

AI matching a character's lip movements to a voice recording so it looks like they're really speaking. Important for talking videos, dubbing and…

A type of AI trained on a huge amount of text so it can understand and write language. ChatGPT, Claude and Gemini are all LLMs. It doesn't "know"…

Running AI directly on your own computer or phone, not on another company's servers. Pros: privacy and no subscription cost. Cons: it needs powerful…

A way of building AI by letting the computer learn patterns from data, instead of programming every rule one by one. Almost all modern AI is built…

An open standard that lets AI connect to all kinds of tools and data (Gmail, Notion, databases) in one consistent way, like "USB for AI". It's…

The actual "brain" of an AI: a giant maths file that has already been trained. One company can have many models of different sizes and strengths…

The number or name that tells model releases apart. Each new version is usually smarter, faster or cheaper. Your tools often use a specific version…

AI that can understand and create more than one type of media (text, pictures, audio, video) in a single model. You can upload a picture and ask…

A mathematical structure made of many artificial "neurons" connected in layers. Each connection has a weight that gets adjusted during training. It…

An AI model whose weight files are released publicly, so anyone can download it, run it themselves and change it. It's free to use but needs a…

The values inside a model that get adjusted during training; this is where the model's "knowledge" is stored. The number of parameters (for example…

Giving AI a specific role ("You are a TikTok marketing expert") so its answers are more focused and fitting. One of the simplest and most effective…

The instruction or question you give an AI. The quality of the answer depends heavily on the quality of the prompt: the clearer the context, goal and…

The skill of writing prompts so AI gives you the result you want, every time. It involves giving a role, context, examples, an output format and…

A technique where AI "reads" your documents first before answering, so its answer is based on your facts, not general memory. It's the main way…

A model designed to "think" longer before answering: planning, checking and correcting itself. Slower and more expensive, but much better for complex…

An AI video technique where you give a starting picture AND an ending picture, and the AI fills in the movement in between. Very useful for…

Any text, picture, audio or video made by AI, rather than fully recorded or written by a person. Platforms increasingly require labels on synthetic…

Background instructions set before the conversation starts: the AI's role, tone, rules and limits. Regular users don't see it, but it shapes the AI's…

A small unit of text that AI reads and writes, roughly one short word or part of a longer word. API prices and conversation length limits are counted…

The ability of AI to call outside tools (a calculator, web search, a calendar, business systems) when needed, and use the results in its answer. It's…

The process of teaching a model with lots of data so it learns patterns. It's expensive, takes weeks and uses thousands of GPUs. Once trained, the…

Increasing the resolution of a picture or video with AI while adding detail, not just stretching the pixels. A common last step before uploading, so…

A special database for storing and searching embeddings quickly. It's the engine behind RAG systems that can search millions of documents in the…

Copying someone's voice from a short recording. Useful for your own content, but dangerous if misused. Only use your own voice, or someone else's…

The sequence of steps to get a job done. In the AI world, a workflow is a recipe: which tools, in what order, with which prompts, so you can get the…

Asking AI to do something without giving any examples, relying fully on its general ability. Fast, but the results are less consistent than few-shot.