AI Compare Ninja AI Compare Ninja

Compare AI tools. Use prompts and skills.
Pricing market
Automatic
Language
EN

AI Glossary

The key AI terms explained in one sentence, with an example. Want more depth? Open “For pros”.

151 terms · Last checked: October 2026

Basics

  • Artificial intelligence (AI)

    Artificial intelligence is software that does tasks that usually need human thinking: understanding language, recognising images, writing text or preparing decisions.

  • Generative AI

    Generative AI creates new content: text, images, music, video or code. You describe what you want and the AI produces a draft.

  • Machine learning

    In machine learning, a computer learns from examples instead of someone programming every rule by hand.

  • Neural network

    A neural network is a computing model made of many connected nodes. It is loosely inspired by the brain and underpins almost all modern AI.

  • LLM (large language model)

    An LLM is an AI model trained on huge amounts of text. It can understand, write and summarise text and answer questions.

  • AI model

    An AI model is the trained “brain” behind an AI tool. One app can offer several models that differ in speed, quality and price.

  • Training

    Training is the learning phase of an AI. The model sees a huge number of examples and keeps adjusting until its answers are good enough.

  • Inference

    Inference means using a model that has already been trained. Every answer you get from an AI is inference.

  • Multimodal

    A multimodal AI understands or creates several kinds of content, for example text, images, speech and video.

  • Hallucination

    A hallucination is a false statement that an AI phrases convincingly, such as made-up numbers, quotes or sources.

  • Transformer

    A transformer is the blueprint behind almost all of today's language models. It looks at all the words in a text at once and works out which ones belong together.

  • Foundation model

    A foundation model is a large AI model trained on huge amounts of general data that can then be adapted to many different tasks.

  • Frontier model

    Frontier models are the most capable AI models of their time, usually built by only a few large providers.

  • AGI (artificial general intelligence)

    AGI means an AI that can do most intellectual tasks at least as well as a person. No such AI exists so far.

  • World model

    A world model is an AI that learns how an environment behaves and can predict what happens next.

  • Pretraining

    Pretraining is the first and most extensive learning phase of an AI model, in which it learns how language works from huge amounts of text.

  • Attention mechanism

    Attention is the calculation a language model uses to decide which other words matter for the current word.

  • Deep learning

    Deep learning is machine learning with neural networks made of many layers. It powers language models, image recognition and speech recognition.

  • Natural language processing (NLP)

    Natural language processing is the field in which computers read, understand and produce human language.

  • Computer vision

    Computer vision is the part of AI that analyses images and video, for example to recognise objects, text or faces.

  • Reinforcement learning

    In reinforcement learning an AI tries out actions and learns from reward or penalty what leads to the goal.

  • Supervised learning

    In supervised learning an AI learns from examples that already include the correct answer.

  • Scaling laws

    Scaling laws describe how AI models get predictably better when model size, data and compute increase.

Working with AI

  • Prompt

    A prompt is your instruction to the AI: what you type into the input box.

  • System prompt

    A system prompt is a background instruction that applies before every conversation. It sets how the AI should behave.

  • Prompt engineering

    Prompt engineering means writing and testing instructions so the AI gives reliably good results.

  • Few-shot prompting

    In few-shot prompting you give the AI a few examples so it copies their style and format.

  • Token

    A token is a small piece of text an AI computes with, often part of a word. AI vendors set prices and limits in tokens.

  • Context window

    The context window is how much text an AI can keep in view at once: the conversation plus any documents you added.

  • Temperature

    Temperature is a setting that controls how creative or predictable an AI’s answers are.

  • Reasoning model

    A reasoning model “thinks” in intermediate steps before it answers. This helps with maths, logic, planning and code.

  • Chatbot

    A chatbot is a program you talk to in a chat window. Modern chatbots run on large language models.

  • Memory

    With memory, an AI assistant remembers things about you across conversations, such as your job or writing style.

  • Context engineering

    Context engineering means giving the AI exactly the information it needs for a task, no more and no less.

  • Chain of thought

    Chain of thought means an AI solves a task in intermediate steps instead of answering straight away. This often makes complex answers more accurate.

  • Structured output

    With structured output, the AI answers in a fixed format, such as a table or JSON, so a program can process the answer directly.

  • Vibe coding

    Vibe coding means building software by telling the AI in plain language what you want and barely reading the generated code yourself.

  • Zero-shot prompting

    Zero-shot means giving the AI a task without a single example and relying on it to understand the task anyway.

  • Prompt chaining

    With prompt chaining you split a big task into several prompts; the answer of one becomes the input of the next.

  • Meta-prompting

    Meta-prompting means letting the AI write or improve a prompt instead of writing it yourself.

  • Role prompting

    A role prompt gives the AI a role, such as “You are an experienced editor”, so that tone and perspective fit.

  • AI assistant

    An AI assistant is a program that answers questions and completes tasks, usually through chat or voice.

  • Sampling (top-p, top-k)

    Sampling is how a language model picks one of several possible next words. It decides how predictable or varied texts become.

Skills and agents

  • Skill (SKILL.md)

    A skill is a fixed set of work instructions for one task. It lives in a folder with a text file called SKILL.md and, if needed, extra files such as templates or scripts. The AI follows it step by step so results become more consistent and reliable.

  • AI agent

    An AI agent is an assistant with a goal. It plans the steps itself and uses tools and often several skills to get there. Well-configured agents ask you before important decisions.

  • Tool calling

    With tool calling, the AI may use tools such as web search, a calendar or a calculator instead of only writing text.

  • MCP (Model Context Protocol)

    MCP is an open standard for connecting AI assistants to apps and data, a bit like a USB port for AI.

  • Workflow automation

    Workflow automation connects apps so that recurring tasks run by themselves, today often with AI steps in between.

  • Multi-agent system

    In a multi-agent system, several specialised AI agents work together: one researches, one writes, one checks.

  • Custom GPT

    A Custom GPT is your own version of ChatGPT with fixed instructions and knowledge files. According to OpenAI, Custom GPTs are scheduled to retire on 11 December 2026.

  • Gem (Gemini)

    A Gem is a custom assistant in Google’s Gemini app, with its own instructions for a specific task.

  • Human in the loop

    Human in the loop means a person checks and approves before the AI does something important, such as sending an email or paying money.

  • Agent builder

    An agent builder is a tool for putting AI agents together with little or no coding.

  • Agentic AI

    Agentic AI describes systems that pursue a goal on their own in several steps, use tools and check intermediate results, instead of just answering one question.

  • Agent loop

    The agent loop is how an AI agent works: plan, act, look at the result, plan again, until the task is done.

  • Subagent

    A subagent is a helper agent that a main agent hands a subtask to. It works with its own context and returns only the result.

  • Agent graph

    An agent graph describes an AI workflow as a network of steps and connections, defining which step comes next and when a loop or approval is needed.

  • A2A (Agent2Agent protocol)

    A2A is an open standard that lets AI agents from different providers talk to each other and hand over tasks.

  • Computer use

    Computer use means an AI operates a computer like a person: it sees the screen, moves the mouse and types.

  • AI browser

    An AI browser is a web browser with a built-in AI assistant that summarises pages and can carry out tasks on the web for you.

  • Deep research

    Deep research is a feature where an AI agent reads many web sources and writes a detailed report with citations.

  • Coding agent

    A coding agent is an AI that reads, changes and tests code on its own and proposes changes for review.

  • AGENTS.md

    AGENTS.md is a text file in a project folder that tells AI coding agents how to work in that project.

  • Permission modes (agents)

    A permission mode sets how much an AI agent may do without asking: read only, ask first, or act on its own.

  • Agent orchestration

    Orchestration means steering several AI steps or agents so that they solve a task together.

  • Sandbox

    A sandbox is an isolated environment in which an AI agent can run code or operate programs without putting your real system at risk.

  • Background agent

    A background agent works on a task on its own while you do something else and comes back with a result.

  • Agentic commerce

    Agentic commerce means an AI agent searches, compares and, with your approval, buys products on your behalf.

Data and knowledge

  • RAG (retrieval-augmented generation)

    With RAG, the AI first looks up matching passages in your documents and answers based on them. It relies on your facts rather than only its training.

  • Embedding

    An embedding turns a text into a long list of numbers. Texts with similar meaning get similar numbers.

  • Vector database

    A vector database stores embeddings and quickly finds the texts most similar in meaning to a question.

  • Fine-tuning

    Fine-tuning retrains a finished model with your own examples so it handles a particular style or task better.

  • Training data

    Training data are the examples an AI learns from, such as texts, images or code.

  • Knowledge cutoff

    The knowledge cutoff is the date up to which a model has seen data. It does not know what happened afterwards unless it searches the web.

  • Grounding

    Grounding means an AI bases its answer on verifiable sources, such as web pages or your documents.

  • Benchmark

    A benchmark is a standardised test for comparing AI models, for example in maths, coding or general knowledge.

  • Knowledge graph

    A knowledge graph stores knowledge as a network of things and their relationships, for example “Berlin – is capital of – Germany”.

  • GraphRAG

    GraphRAG is a form of RAG in which the AI retrieves knowledge not only as text snippets but as a knowledge graph with relationships.

  • Chunking

    Chunking means splitting long documents into small sections so an AI can later find exactly the right passages.

  • Semantic search

    Semantic search finds texts by their meaning, not just by matching words.

  • Reranking

    Reranking sorts search results a second time so the most relevant matches end up on top.

  • Synthetic data

    Synthetic data is artificially generated data, often made by an AI, that replaces or supplements real data.

  • LoRA

    LoRA is an economical way to adapt an AI model: instead of changing all weights, you train only small add-on components.

  • Hybrid search

    Hybrid search combines classic keyword search with meaning-based search to find both exact terms and related matches.

  • Knowledge base (for AI)

    A knowledge base is a collection of checked documents an AI uses to answer questions with your own knowledge.

  • Data labelling

    Data labelling means adding the correct answer to data so that an AI can learn from it.

  • OCR (optical character recognition)

    OCR turns text in images or scans into editable text.

  • Model collapse

    Model collapse describes how AI models can get worse when they are trained again and again on text or images made by other AIs.

  • Long context

    Long context means a model can process a lot of text at once, such as whole books or large codebases.

Image, video, audio

  • Diffusion model

    A diffusion model creates images by shaping random noise, step by step, into a clear picture that matches your description.

  • Text-to-image

    Text-to-image AI creates a picture from a written description.

  • Text-to-video

    Text-to-video AI creates short video clips from a description or a starting image.

  • Text-to-speech

    Text-to-speech AI reads written text aloud in a natural-sounding voice.

  • Speech-to-text

    Speech-to-text turns spoken language into written text.

  • Deepfake

    A deepfake is an AI-made or AI-altered image, video or audio that looks real and makes a person say or do something that never happened.

  • Content Credentials (C2PA)

    Content Credentials are signed provenance details embedded in an image, video or audio file. They can show which program or AI created or edited the content.

  • AI watermark

    An AI watermark is an invisible mark in images, text, audio or video that shows the content was generated with AI.

  • Voice cloning

    Voice cloning means using AI to recreate a specific person's voice, often from just a few seconds of recording.

  • Inpainting

    Inpainting means using AI to refill or change a marked area of an image.

  • AI upscaling

    AI upscaling enlarges images or videos and adds plausible details so that they look sharper.

  • Image-to-image

    With image-to-image, the AI takes an existing image as a template and creates a new version, for example in a different style.

  • AI music

    AI music means songs, melodies or background music that an AI generates from a text description.

  • AI avatar

    An AI avatar is an artificial or recreated person in a video who speaks a text with matching lip movements.

Tech and costs

  • API

    An API is an interface that lets programs talk to each other. With an AI API, developers build AI into their own apps.

  • API key

    An API key is like a password for an API. Anyone who has it can use the service at your expense.

  • Rate limit

    A rate limit caps how many requests or messages you may send in a given time.

  • Open-weight model

    With an open-weight model, you can download the trained weights and run the model yourself.

  • Local AI

    Local AI runs on your own computer or server instead of in the cloud. Your data stays on the device.

  • Quantization

    Quantization makes an AI model smaller by storing its numbers less precisely, so it runs on weaker devices too.

  • Parameters

    Parameters are a model’s learned settings. Their number is a rough measure of a model’s size.

  • GPU (graphics processing unit)

    A GPU is a chip that runs a huge number of calculations at the same time. That is why AI training and serving mostly run on GPUs or on similar specialised chips such as Google's TPUs.

  • Test-time compute

    Test-time compute means giving an AI more computing time when it answers, so it can think longer. This often improves hard answers.

  • Mixture of experts (MoE)

    Mixture of experts is a model design in which only part of the model works on each request, like calling in the right specialists.

  • Distillation

    Distillation means letting a small AI model learn from a large one so it becomes almost as good but faster and cheaper.

  • Small language model (SLM)

    A small language model is a language model with relatively few parameters that can also run on a laptop or phone.

  • NPU (neural processing unit)

    An NPU is a chip component that runs AI calculations efficiently, for example in laptops and phones.

  • Prompt caching

    Prompt caching means an AI provider stores an unchanging start of your request, so repeated requests become faster and cheaper.

  • RLHF

    RLHF is a training step in which people rate an AI's answers and the model learns to give the preferred ones.

  • Evals (AI evaluations)

    Evals are repeatable tests that check how well an AI application does its job.

  • LLM-as-a-judge

    With LLM-as-a-judge, one language model rates another model's answers, for example for accuracy or tone.

  • Latency

    Latency is the waiting time until an AI starts answering or finishes.

  • Batch processing (batch API)

    With a batch API you send many AI requests at once and get the results later, usually at a lower price.

  • KV cache

    The KV cache is a temporary store inside a language model that keeps parts it has already computed, so each new word is generated faster.

  • Speculative decoding

    In speculative decoding a small, fast model guesses several words ahead and the large model checks them in one step. That makes answers faster.

  • TPU (tensor processing unit)

    A TPU is a special chip for AI computing developed by Google and used mainly in data centres.

  • FLOPs (compute)

    FLOPs count the calculation steps needed to train or run an AI model. They are a measure of compute.

  • Model card

    A model card is a fact sheet for an AI model: what it is for, how it was tested and where its limits are.

  • Open source AI

    Open source AI is an AI system whose components may be freely used, studied, modified and shared.

  • Overfitting

    Overfitting means a model memorises its training examples and performs poorly on new cases.

Safety and law

  • Prompt injection

    In a prompt injection, attackers hide instructions in text, web pages or files so the AI does something unwanted.

  • Jailbreak

    A jailbreak is an attempt to trick an AI into ignoring its safety rules.

  • Guardrails

    Guardrails are safety rules around an AI. They prevent unwanted answers or actions.

  • Bias

    Bias means an AI systematically favours or disadvantages certain groups or views, usually because its training data is unbalanced.

  • AI labeling

    AI labeling means openly saying when content was made with AI or when you are talking to an AI.

  • EU AI Act

    The EU AI Act is the European law on artificial intelligence. It sorts AI by risk and sets rules for each level.

  • AI literacy

    AI literacy is the knowledge needed to use AI sensibly and safely: what it can do, where it makes mistakes and what to watch out for.

  • Red teaming

    In red teaming, experts deliberately try to make an AI fail or give harmful answers, to find weaknesses before launch.

  • Data poisoning

    Data poisoning is an attack that sneaks manipulated data into an AI's training or knowledge source so it later reacts wrongly.

  • GPAI model (general-purpose AI model)

    Under the EU AI Act, a GPAI model is an AI model that can be used for many different tasks, such as a large language model.

  • Systemic risk (AI Act)

    In the EU AI Act, systemic risk means the particular danger very capable AI models can pose to society, safety or fundamental rights.

  • High-risk AI

    Under the EU AI Act, high-risk AI systems can pose a significant risk to health, safety or fundamental rights, for example because they help decide on job applications or loans or are a safety component of a product. Strict rules apply to them.

  • Provider and deployer (AI Act)

    The EU AI Act distinguishes providers, who develop an AI system (or have it developed) and place it on the market under their own name, from deployers, who use it under their own authority in a professional context.

  • AI alignment

    Alignment is the work of making AI systems do what people actually intend while respecting rules and values.

  • Explainability (XAI)

    Explainability means people can understand why an AI reached a result.

  • Sycophancy

    Sycophancy means an AI tells you what you want to hear: it agrees, praises or changes a correct answer just because you push back.

  • Prohibited AI practices

    The EU AI Act bans some AI uses outright because they are considered an unacceptable risk.

  • Transparency obligations (AI Act)

    Transparency obligations require that people can tell when they are talking to an AI or seeing AI-generated content.

  • TDM opt-out

    With a TDM opt-out, rights holders can state that their publicly available works may not be used for text and data mining, including AI training.

  • Shadow AI

    Shadow AI means employees use AI tools for work without the company knowing or having approved them.

  • Sensitive information disclosure

    Sensitive information disclosure means confidential information reaches unauthorised people through an AI, for example via inputs, answers or connected tools.

  • Conformity assessment (AI Act)

    Conformity assessment is the check of whether a high-risk AI system meets all AI Act requirements before it goes on the market.

Frequently asked questions

Who is the AI Glossary for?

For anyone who uses AI or wants to understand it. The first sentence of every entry is written for beginners. “For pros” adds technical detail.

Where do the explanations come from?

We write every text ourselves. Where facts can change, such as dates or product features, we name the vendor or official source.

What is the difference between a prompt, a skill and an agent?

A prompt is a single instruction. A skill is a fixed set of work instructions for one task. An agent is an assistant with a goal that uses several skills and carries out steps on its own. We recommend setting it up so that it asks you before important steps.

Is the glossary legal advice?

No. Terms such as the EU AI Act or AI labeling are explained in general terms only. For a specific case, ask a qualified lawyer.

Does the glossary cost anything?

No. The glossary is free and needs no sign-up.

Written by AI Compare Ninja with AI assistance and edited by us. Every explanation is our own text.