AI · AI basics
AI glossary: generative AI terms explained
Short answer
This glossary explains about 60 terms you meet when using generative AI, grouped by topic: models and training, text and LLMs, image and video generation, audio, and safety and law. Each entry gives a plain meaning and a short example, so you can decode a model card, a settings panel or a news story without a technical background.
Models and training
These terms describe what a model is and how it gets built. They apply to text, image, video and audio models alike.
| Term | Meaning | Example |
|---|---|---|
| Artificial intelligence (AI) | Software that performs tasks that normally need human intelligence | Translation, face unlock, a chat assistant |
| Machine learning | AI that learns patterns from examples instead of following hand-written rules | A spam filter trained on labeled emails |
| Neural network | A model made of many layers of simple connected units whose strengths are learned | The core of every modern image or language model |
| Model | The trained network itself, a file of learned numbers plus the code to run it | An app such as a chatbot runs one or more models |
| Parameters (weights) | The numbers inside a model that training adjusts | A 7B model has about 7 billion parameters |
| Training data | The examples a model learns from | Web text, licensed books, image and caption pairs |
| Pretraining | The first, largest training stage on broad data | Learning to predict the next token on web-scale text |
| Fine-tuning | Further training of an existing model on a smaller, focused dataset | Training a base model on customer-support answers |
| LoRA | Low-rank adaptation: a small add-on file that fine-tunes a model cheaply without changing the original weights | A style or character LoRA for an image model |
| Foundation model | A large general model that many products and fine-tunes are built on | One LLM powering chat, search and coding tools |
| Open-weight model | A model whose weights can be downloaded and run yourself; the license may still limit use | Running an open model on your own computer |
| Inference | Running a trained model to get an output | Each time you press Generate |
| Quantization | Storing weights with fewer bits to make a model smaller and faster, with a small quality loss | A 4-bit version of a model that fits on a laptop |
| Multimodal model | A model that handles more than one type of data, such as text and images | Asking a question about a photo |
Text and large language models
Terms you see in chat apps, API settings and articles about LLMs.
| Term | Meaning | Example |
|---|---|---|
| Large language model (LLM) | A large neural network trained to predict the next token of text | The model behind ChatGPT, Claude, Gemini, DeepSeek, Qwen |
| Token | The unit of text a model reads and writes: a word, part of a word or a symbol | unbelievable can be 3 tokens: un, bel, ievable |
| Context window | The maximum number of tokens a model can consider at once, input plus output | A long chat that starts forgetting its beginning |
| Transformer | The neural network design, based on attention, used by almost all LLMs | GPT means Generative Pre-trained Transformer |
| Attention | The step that lets each token weigh which earlier tokens matter to it | Linking it to trophy in a sentence about a suitcase |
| Embedding | A list of numbers that represents meaning, so similar texts get similar numbers | Finding documents similar to a question |
| Prompt | The input you give a model | Summarize this email in two sentences |
| System prompt | Hidden instructions set by the app or developer before the conversation | You are a polite support agent for a bike shop |
| Temperature | A setting that controls randomness in the output | Low for data extraction, higher for brainstorming |
| Hallucination | A confident but false or invented statement from a model | A citation to a paper that does not exist |
| RAG (retrieval-augmented generation) | Searching documents first and putting the best passages in the prompt | A help bot that answers from your manuals |
| Knowledge cutoff | The date after which the model has no training data | It does not know about last month's events |
| Reasoning mode | A mode where the model writes intermediate steps before the final answer | Shown as thinking in some chat apps |
| Tool use (function calling) | The model asks the app to run a tool and reads the result | Calling a calculator or web search |
| AI agent | A system where a model plans and takes several tool actions toward a goal | Researching and booking options across websites |
Image and video generation
Settings and techniques that appear in image and video tools such as Midjourney, Stable Diffusion, FLUX, Veo, Kling, Runway and Seedance. Names differ slightly between tools.
| Term | Meaning | Example |
|---|---|---|
| Diffusion model | A model that creates an image by removing noise step by step from random noise | Most text-to-image tools |
| Latent space | A compressed internal representation where the model works instead of on raw pixels | Latent diffusion generates a small latent, then decodes it to an image |
| Text-to-image | Generating an image from a text prompt | A red bicycle against a blue wall at sunset |
| Image-to-image | Generating a new image guided by an input image | Turning a sketch into a finished illustration |
| Inpainting | Regenerating only a masked area of an image | Replacing a coffee cup on a table with a plant |
| Negative prompt | Text describing what to avoid | blurry, extra fingers, watermark |
| Seed | The starting random number; the same seed and settings give a similar result | Reusing a seed to make small prompt changes |
| Guidance scale (CFG) | How strongly the output follows the prompt versus looking natural | Too high gives harsh colors and artifacts |
| ControlNet | An add-on that guides a diffusion model with a pose, edge or depth image | Keeping a character's exact pose |
| Depth map | A grayscale image where brightness shows distance from the camera | Used for parallax video and ControlNet |
| Upscaling | Enlarging an image with a model that adds plausible detail | A 512-pixel image enlarged 4 times |
| Text-to-video | Generating a video clip from a text prompt | A paper boat drifting down a rainy street |
| Image-to-video | Animating a still image into a clip | Making a product photo slowly rotate |
| First and last frame | Giving a start image and an end image and letting the model fill the motion between | A door closed in the first frame and open in the last |
Audio and speech
| Term | Meaning | Example |
|---|---|---|
| Speech-to-text (ASR) | Automatic speech recognition: turning spoken audio into text | Transcribing a meeting or a video |
| Whisper | A widely used open speech recognition model family from OpenAI | Making subtitles in many languages |
| Text-to-speech (TTS) | Turning text into spoken audio | Reading an article aloud |
| Voice cloning | A TTS model that imitates a specific person's voice from a sample | A voiceover in the creator's own voice, with consent |
| Speaker diarization | Labeling who spoke when in a recording | Speaker 1 and Speaker 2 in an interview transcript |
| Word error rate (WER) | The share of words a transcript gets wrong, used to measure ASR quality | A WER of 5 percent means about 1 wrong word in 20 |
| SRT and VTT | Plain-text subtitle formats with timestamps for each line | Uploading captions to a video platform |
Safety, labeling and law
Rules differ by country and change often. This section is general information, not legal advice.
| Term | Meaning | Example |
|---|---|---|
| Deepfake | Realistic AI-generated or AI-altered media that shows people or events that are not real | A video of a politician saying words they never said |
| Content credentials (C2PA) | An open standard that attaches signed information about how a file was made and edited | A photo showing it was edited with an AI tool |
| Invisible watermark | A hidden signal embedded in pixels or audio that software can detect later | A generator marking every image it makes |
| Explicit and implicit labels | In China's labeling rules: a visible label people can see, and a hidden label in metadata or a watermark | An AI-generated tag on a video plus a metadata field |
| EU AI Act transparency duties | EU rules that people must be told when they talk to an AI, and that AI content and deepfakes must be identifiable | A chatbot disclosing it is a bot |
| Training data copyright | The legal question of whether training on copyrighted works is allowed, which differs by country | Lawsuits and policy reports in several countries |
| Guardrails | Rules and filters that block harmful requests or outputs | Refusing to give instructions for weapons |
| Jailbreak | A prompt crafted to make a model ignore its safety rules | A role-play trick to bypass refusals |
| Prompt injection | Hidden instructions in a web page or file that hijack an AI that reads it | A page that tells an AI agent to send your data elsewhere |
| Bias | Systematic unfairness in outputs, often learned from skewed training data | Images of a profession that show only one gender |
Terms people often mix up
Several pairs of terms are confused in news stories and product pages. The difference usually changes what you should do.
| Pair | Difference | Why it matters |
|---|---|---|
| Model vs app | The model generates; the app adds the interface, system prompt, tools and storage | Two apps on the same model can behave differently |
| Token vs word | A word can be one token or several; counts differ by language | Prices and limits are counted in tokens |
| Parameters vs context window | Parameters are what the model learned; the context window is what it can see right now | A huge model can still forget the start of a long chat |
| Fine-tuning vs RAG | Fine-tuning changes the model; RAG gives it documents at question time | For facts that change often, RAG is usually easier |
| Open-weight vs open-source | Open weights can be downloaded; open source also means the training code and data are open under an open license | Read the license before commercial use |
| Watermark vs label | A watermark is a technical signal; a label is a disclosure people can read | Some laws require both |
For a deeper explanation of any term, see the related articles in this section, such as how LLMs work, tokens and the context window, and how diffusion models work.
FAQ
What is the difference between AI, machine learning and deep learning?
AI is the whole field. Machine learning is the part of AI that learns from examples. Deep learning is the part of machine learning that uses many-layered neural networks, which is what modern generative models are.
What is the difference between an LLM and a chatbot?
An LLM is the model that generates text. A chatbot is an app built around one or more models, adding a chat interface, instructions, memory features and tools such as web search.
What does open source mean for AI models?
It is used loosely. Often it means only that the weights can be downloaded. A fully open model also shares training code and data details under an open license. Check the license for commercial use and attribution rules.
What does 7B or 70B mean in a model name?
It is the approximate number of parameters, in billions. A larger number usually means more capability but also more memory and computing power to run it.
Is prompt engineering a real skill?
Yes, in the practical sense of writing clear, specific instructions, giving examples and supplying the right context. Models keep improving at understanding plain requests, but clear input still gives more reliable output.