AI Glossary: 60 AI Terms Explained in Plain English

The Maitz AI glossary explains 60 artificial intelligence terms you keep running into in news, guides and reviews, from LLM and token to AI agent, RAG and MCP. Each term has a short definition checked against official documentation or the original research, plus a link to a Maitz article that goes deeper. Browse A to Z or by topic.

AI glossary index: all terms A to Z

A: AI agent · AI image editing · AI watermarking · API · Artificial general intelligence (AGI) · Artificial intelligence (AI)
B: Benchmark
C: Chain-of-thought prompting · Chatbot memory · Computer use · Content Credentials (C2PA) · Context window · Custom instructions
D: Deep learning · Deep Research · Deepfake · Diffusion model
E: Embeddings
F: Few-shot prompting · Fine-tuning · Foundation model
G: Generative AI · Grounding
H: Hallucination
I: Image-to-video · Inference
J: Jailbreak
K: Knowledge cutoff
L: Large language model (LLM)
M: Machine learning · Model Context Protocol (MCP) · Model distillation · Multimodal
N: Neural network · No-code automation
O: Open-weight model · Optical character recognition (OCR)
P: Parameters · Pretraining · Prompt · Prompt injection
R: Reasoning model · Reinforcement learning from human feedback (RLHF) · Retrieval-augmented generation (RAG)
S: Semantic search · Speech-to-text · System prompt
T: Temperature · Temporary chat · Text-to-image · Text-to-speech (TTS) · Token · Tool use (function calling) · Training on user data · Transformer
U: Usage limit
V: Vector database · Vibe coding · Voice cloning
W: Webhook

Basics

Artificial general intelligence (AGI)

Artificial general intelligence (AGI) is an AI system that could do most intellectual tasks at or above human level. There is no single agreed definition. OpenAI's Charter describes it as "highly autonomous systems that outperform humans at most economically valuable work," and Google DeepMind researchers proposed rating progress toward it in levels.

Artificial intelligence (AI)

Artificial intelligence (AI) is the field of computer science focused on building systems that perform tasks that normally require human intelligence, such as understanding language, recognizing images, making decisions and making predictions. Most AI systems today learn from examples instead of following hand-written rules. Chatbots like ChatGPT are just one kind of AI.

Deep learning

Deep learning is a branch of machine learning that uses neural networks with many layers. Each layer learns a more abstract representation of the data, so the model discovers on its own which features matter. It is the technology behind modern speech recognition, image recognition and language models.

Foundation model

A foundation model is a large model trained on a huge and varied dataset that can be adapted to many different tasks. Stanford researchers coined the term in 2021. Models such as GPT, Claude and Gemini are foundation models that chatbots, tools and apps are built on.

Generative AI

Generative AI is AI that creates new content, such as text, images, code, music or video, from a request written in plain language. The models learn patterns from large amounts of data and use them to produce output that did not exist before. ChatGPT, Gemini and image generators are familiar examples.

Large language model (LLM)

A large language model (LLM) is an AI model with a huge number of parameters, trained on very large amounts of text. It predicts which token is likely to come next in a sequence, which lets it write, summarize, translate and answer questions. ChatGPT, Claude and Gemini are built on models like this, with extra training to act as assistants.

Machine learning

Machine learning is a way of building AI systems in which the computer learns patterns from data instead of a programmer writing every rule. A model is trained on examples and then makes predictions or decisions about new data, such as flagging spam or recommending a movie.

Neural network

A neural network is a machine learning model made of layers of simple connected computing units, called neurons. Each connection has a weight, and during training the weights are adjusted so the network gives more accurate answers. Large language models are very large neural networks.

Back to the index

How models work

Benchmark

A benchmark is a fixed set of questions or tasks used to compare models, such as MMLU, which tests knowledge across 57 subjects. A high score shows ability on the benchmark's tasks, not necessarily on your work, and some questions may leak into training data. That is why it pays to test a model on real tasks too.

Context window

The context window is the amount of text, measured in tokens, that a model can hold and consider at once: the instructions, the conversation, any files and the answer itself. Anthropic describes it as the model's "working memory," separate from what it learned in training. A larger window handles longer documents, but Anthropic notes accuracy can drop as context grows.

Fine-tuning

Fine-tuning is extra training of an already trained model on your own set of examples, so it learns to follow their patterns: a style, a format or a type of task. OpenAI offers this to developers through its API. It is different from custom instructions, Gems or skills, which change the instructions the model receives, not the model itself.

Inference

Inference is the stage where a trained model takes new input and produces an output, for example when you send a question to a chatbot. Unlike training, which happens once at great cost, inference happens on every request. API prices and response speed depend mainly on the cost of inference.

Knowledge cutoff

The knowledge cutoff is the date after which a model has no information from its training data. It does not know about later events, prices or product versions unless it is connected to web search or you give it the information. AI companies list this date in their model documentation.

Model distillation

Model distillation is training a small "student" model to imitate the outputs of a large "teacher" model. The result is a faster, cheaper model that keeps much of the larger model's ability. The idea was described in a 2015 paper by Geoffrey Hinton and colleagues at Google. AI companies usually ban using their models to distill competing models in their terms of use.

Multimodal

A multimodal model understands, and sometimes generates, several types of data: text, images, audio and video. When Google introduced the Gemini family in 2023, it described the models as able to understand images, audio, video and text. In practice, this lets you photograph a document, ask about a chart or talk to a chatbot through your camera.

Open-weight model

An open-weight model is a model whose weights (parameters) are published for download, so you can run it on your own server or computer and adapt it. Examples include OpenAI's gpt-oss, Google's Gemma and Meta's Llama. Open weights are not always fully open source: the training data is usually not released, and the license sets what is allowed.

Parameters

Parameters are the numerical values inside a model, mainly the weights of its neural network, that are set during training. They are what the model has "learned." Model size is usually described by its parameter count, such as 7 billion or hundreds of billions. More parameters do not guarantee a better model, but they usually make it more expensive to run.

Pretraining

Pretraining is the first and largest stage of training a language model: the model reads enormous amounts of text and learns to predict the next token. This is where it picks up general knowledge about language and the world. Later stages, such as fine-tuning and RLHF, turn it into an assistant that follows instructions.

Reasoning model

A reasoning model is a language model that generates internal reasoning steps before giving its final answer. According to OpenAI, these hidden reasoning tokens help it plan, use tools, explore alternatives and solve complex multistep tasks in code, math and science. Answers take longer, and in the API reasoning tokens are billed as output tokens.

Reinforcement learning from human feedback (RLHF)

Reinforcement learning from human feedback (RLHF) is a training method in which people rank a model's answers, a reward model is built from those rankings, and the language model is then trained to prefer the answers people rated highly. OpenAI described it in its 2022 InstructGPT research. It helped turn language models into assistants that follow instructions more helpfully.

Token

A token is the small unit of text a language model reads and writes: a word, part of a word, a character or even a byte. Models count input and output in tokens, so API prices, usage limits and context window size are all measured in tokens. According to OpenAI, one token is about four characters of English text.

Transformer

The transformer is a neural network architecture introduced by Google researchers in the 2017 paper "Attention Is All You Need." Its key idea is the attention mechanism, which lets the model weigh the relationships between all the words in a sequence at once. Most large language models today are based on it, and the T in GPT stands for transformer.

Back to the index

Using chatbots and prompts

Chain-of-thought prompting

Chain-of-thought prompting is a technique where you ask the model to work through the steps of a problem before giving its final answer. Google research from 2022 showed it improves results on math and logic problems. Reasoning models build this process in, so you rarely need to ask for it explicitly.

Chatbot memory

Chatbot memory is the ability to save details from your conversations, such as your name, preferences or projects, and use them in later chats. In ChatGPT, for example, you can see what was saved, delete items or turn memory off in settings. Memory is convenient, but it is best not to share sensitive information with it.

Custom instructions

Custom instructions are standing preferences you write once in a chatbot's settings and it applies them to every new chat: who you are, what tone to use, how long answers should be and in what language. ChatGPT calls the feature Custom Instructions, and other chatbots have similar settings. It saves repeating the same guidance each time.

Deep Research

Deep Research is a mode in chatbots such as Gemini and ChatGPT where an agent plans the research, reads dozens of web sources over several minutes and writes a structured report with citations. It suits broad questions like a market overview, but you should still check the key claims against the sources.

Few-shot prompting

Few-shot prompting means including a few examples of the desired input and output before the actual task, so the model picks up the format and style. OpenAI's 2020 GPT-3 research showed that large models can learn a new task from a handful of such examples without extra training. A prompt with no examples is called zero-shot.

Grounding

Grounding means connecting a model's answer to a source you can check, such as current search results or your own documents. The model then answers from that source rather than only from its training, and usually shows where each detail came from. Google, for example, offers grounding with Google Search in Vertex AI.

Hallucination

A hallucination is an answer from a language model that sounds plausible and confident but is wrong: a fact, quote, number or source that doesn't exist. According to OpenAI, part of the cause is that training and evaluation reward guessing over admitting uncertainty. Newer models hallucinate less, but not never, so check every important detail against a source.

Prompt

A prompt is the instruction or question you give an AI model: the text, files and context that define what you want back. A good prompt states the goal, audience, format and constraints, and sometimes includes an example of the desired result. Anthropic recommends first defining what a good answer looks like, then improving the prompt by testing real results.

System prompt

A system prompt is background instructions a model receives before the conversation that apply throughout it: role, tone, rules and limits. It is usually written by the app's developers, and users typically don't see it. Anthropic recommends using it to give the model a role, such as "You are a careful legal editor."

Temperature

Temperature is a setting that controls how much randomness goes into choosing the next token. A low temperature gives more consistent, predictable answers that suit factual tasks, while a high temperature gives more varied, creative ones. It is usually adjusted in the API or developer tools, not in regular chat apps.

Temporary chat

Temporary Chat is a ChatGPT mode where the conversation is not saved to your history and does not create new memories, and according to OpenAI it is not used to train its models. OpenAI may keep a copy for up to 30 days for safety purposes. It suits a one-off question you don't want kept in your account.

Usage limit

A usage limit is a cap on how many messages or how much work you can do with a chatbot within a set period, such as a few hours or a week. It depends on your plan, the model and the length of the conversation: long messages and large files use more. When you hit the cap, you wait for the reset, switch to a lighter model or upgrade.

Back to the index

Agents, tools and automation

AI agent

An AI agent is a system in which a language model decides on its own which steps to take and which tools to use to complete a task, rather than answering a single message. Anthropic distinguishes workflows, where the path is set in code, from agents, which direct their own process. Because of that independence, agents need clear permissions and approval points.

API

An API (application programming interface) is a defined way for one piece of software to request a service from another. In AI, an API lets developers send a prompt to a model from their own code and get a response without opening the chat app. API usage is usually billed separately from a chat subscription and priced per token.

Computer use

Computer use is an AI model's ability to operate a computer the way a person does: viewing screenshots, moving the mouse, clicking and typing. Anthropic introduced this capability in 2024, and OpenAI introduced a similar agent in 2025. It can handle tasks on websites that have no API, but it needs supervision because the agent can make mistakes or click the wrong thing.

Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open standard for connecting AI apps to external systems: files, databases, tools and workflows. Anthropic introduced it in November 2024, and the official docs compare it to a USB-C port: one standard way to connect assistants like Claude or ChatGPT to a calendar, Notion or an internal tool, instead of a custom integration for each.

No-code automation

No-code automation is building automated processes between apps through a visual interface, without writing code. In tools like Make or Zapier, you connect a trigger, such as a new form submission, to actions like adding a spreadsheet row and sending an email. Make calls such a process a scenario, and it can include AI models too.

Tool use (function calling)

Tool use, also called function calling, is a model's ability to request an external function, such as search, a calculator, a calendar or a database, instead of guessing. The developer defines which tools are available, the model decides when to call them and with what inputs, and the app runs the tool and returns the result. It is the foundation of AI agents.

Vibe coding

Vibe coding is writing software with AI by describing what you want in plain language, taking the model's code and running it, while barely reading the code itself. Andrej Karpathy coined the term in February 2025 in a post saying you can "forget that the code even exists." It is fast for prototypes but risky for code that ships without review.

Webhook

A webhook is a web address that one system automatically sends a message to the moment an event happens, such as a form being submitted or a payment going through. Instead of checking again and again for something new, the receiving system gets the data right away and starts a process. In automation tools like Make, a webhook often serves as a scenario's trigger.

Back to the index

Retrieval and data

Embeddings

An embedding is a representation of text, and in newer models also images, audio or video, as a list of numbers called a vector. According to OpenAI, the distance between two vectors reflects how related the content is. That lets a computer find content by meaning, not just matching words. It is the basis of semantic search, classification and RAG.

Retrieval-augmented generation (RAG)

Retrieval-augmented generation (RAG) is a method where the system first retrieves relevant passages from a document collection, adds them to the context window, and only then has the model write an answer. The answer can draw on current or internal information that was not in the training data, and it can be checked against the source. The method was introduced in a 2020 research paper by Lewis and colleagues.

Semantic search

Semantic search is searching by meaning and intent rather than by matching words alone. The system converts the query and the documents into embeddings and returns the closest matches in meaning, even when they share no words with the query. For example, a search for "refund" can also find "cancel the order and get your money back."

Vector database

A vector database stores embeddings (vectors) and can quickly find the vectors closest to a query. That makes it possible to search by meaning across millions of documents, images or products. It is a common building block in RAG systems and semantic search.

Back to the index

Images, video and voice

AI image editing

AI image editing is changing an existing image with a plain-language instruction, such as "replace the background with a beach" or "remove the person on the left," without manual editing tools. Models like Gemini and ChatGPT Images take an image and a prompt and return an edited version. Check that details like faces, text and logos didn't change by accident.

Diffusion model

A diffusion model is a type of generative model that learns to turn random noise into an image step by step. During training, noise is added to real images and the model learns to remove it one step at a time. To generate, it starts from noise and cleans it up according to the prompt. Many image and video generators use this approach, which took off with the 2020 DDPM paper.

Image-to-video

Image-to-video is generating a short clip from a starting image, guided by a prompt that describes motion, camera and mood. The model keeps the original image as its starting point and animates it. Models like Google's Veo and tools like Higgsfield offer this, and clips are usually limited to a few seconds.

Optical character recognition (OCR)

Optical character recognition (OCR) is technology that detects text in an image or scanned document and turns it into text you can copy, search and edit. Services like Google Cloud Vision recognize text in many languages. Multimodal chatbots can now read text from images without a separate OCR tool, but check the results on tables and handwriting.

Speech-to-text

Speech-to-text converts recorded speech into written text. Models like OpenAI's Whisper, introduced in 2022, were trained on large amounts of audio in many languages. It is used for captions, meeting notes and dictation. Names, technical terms and background noise still cause errors, so review the transcript.

Text-to-image

Text-to-image is generating a new image from a written description. You describe what you want to see, such as the subject, style, lighting and composition, and the model produces a matching image. ChatGPT Images, Gemini and Midjourney are well-known tools. Text inside the image can still come out garbled in some models.

Text-to-speech (TTS)

Text-to-speech (TTS) converts written text into spoken audio. Newer models produce natural-sounding voices with emphasis and emotion in many languages. It is used for narration, audiobooks, accessibility and voice agents. Services like ElevenLabs offer a wide range of voices and languages.

Voice cloning

Voice cloning is creating a synthetic voice that sounds like a specific person, based on recordings of their voice. Services like ElevenLabs can clone a voice from a short sample, or from longer recordings for higher quality. Reputable services require you to confirm you have the right to use the voice, and cloned voices are also used in scams, so verify suspicious calls.

Back to the index

Safety, privacy and trust

AI watermarking

AI watermarking is embedding a marker in AI-generated content so it can be identified later. Google DeepMind's SynthID, for example, embeds an invisible watermark in images, audio, video and text that is designed to survive common changes like cropping or compression. Watermarks help with detection, but not every tool uses them, so a missing watermark doesn't prove content is real.

Content Credentials (C2PA)

Content Credentials are digitally signed information attached to a file that records its origin: which tool created it, and whether and when it was edited. The standard was developed by the C2PA, a coalition whose members include Adobe, Microsoft and Google. OpenAI, for example, adds this information to images created in ChatGPT. It can be lost through screenshots or uploads to some platforms.

Deepfake

A deepfake is an image, audio or video generated or altered with AI that resembles real people, objects, places or events and could falsely appear authentic. That is how the EU AI Act defines it, and the law generally requires disclosing that such content was artificially generated or manipulated. Deepfakes are also used for fraud and impersonation, so check the source before sharing.

Jailbreak

A jailbreak is an attempt to get a model to bypass its safety rules, for example through role-play, indirect wording or a clever sequence of instructions, so it produces content it should refuse. In prompt injection the instructions are hidden in outside content, while in a jailbreak the user is the one trying to get around the limits. AI companies respond with filters and additional training.

Prompt injection

Prompt injection is an attack that hides malicious instructions in content a model reads, such as a web page, email or document, to make it act against the user: reveal information, change its answer or take an action. The risk grows with agents that browse the web and use tools. OpenAI recommends limiting permissions and approving sensitive actions yourself.

Training on user data

Training on user data is when an AI company uses the content you send a chatbot to improve future models. In consumer services it is often on by default and can be turned off in settings, while business accounts and the API are generally not used for training by default. Check the setting in each service you use.

Back to the index

How we built this AI glossary

Every definition was written by Maitz and checked on October 11, 2026 against official documentation from the company behind the technology, or against the original paper that introduced the term. We update the glossary when terms change or new ones appear. Read more about our process on How We Test AI Tools, and if you spot an error, see our editorial standards and corrections policy.

Sources

OpenAI Charter
Morris et al. (Google DeepMind): Levels of AGI (2023)
Google Cloud: What is artificial intelligence (AI)?
LeCun, Bengio, Hinton: Deep learning (Nature, 2015)
Bommasani et al.: On the Opportunities and Risks of Foundation Models (Stanford CRFM, 2021)
Google Cloud: What is generative AI?
Google Machine Learning Crash Course: Introduction to Large Language Models
Anthropic docs: Glossary
Google for Developers: Introduction to Machine Learning, What is ML?
Google Machine Learning Crash Course: Neural networks
Hendrycks et al.: Measuring Massive Multitask Language Understanding (MMLU, 2020)
Anthropic docs (Claude): Context windows
OpenAI API docs: Model optimization (fine-tuning)
Anthropic docs (Claude): Glossary
Google Machine Learning Glossary: inference
Anthropic docs (Claude): Models overview (training data cutoff column)
Hinton, Vinyals, Dean (Google): Distilling the Knowledge in a Neural Network (2015)
Gemini Team (Google): Gemini, A Family of Highly Capable Multimodal Models (2023)
OpenAI: Introducing gpt-oss (2025)
Google AI for Developers: Gemma docs
Google Machine Learning Glossary: parameter
Radford et al. (OpenAI): Improving Language Understanding by Generative Pre-Training (2018)
OpenAI API docs: Reasoning models
OpenAI: Learning to reason with LLMs (2024)
Ouyang et al. (OpenAI): Training language models to follow instructions with human feedback (2022)
OpenAI Help Center: What are tokens and how to count them?
Vaswani et al. (Google): Attention Is All You Need (2017)
Wei et al. (Google): Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022)
OpenAI Help Center: Memory FAQ
OpenAI Help Center: ChatGPT Custom Instructions
Gemini Apps Help: Use Deep Research in Gemini Apps
OpenAI Help Center: Deep research FAQ
Brown et al. (OpenAI): Language Models are Few-Shot Learners (2020)
Google Cloud Vertex AI docs: Grounding overview
OpenAI: Why language models hallucinate (2025)
Anthropic docs: Reduce hallucinations
Anthropic docs (Claude): Prompt engineering overview
OpenAI API docs: Prompt engineering
Anthropic docs (Claude): Giving Claude a role with a system prompt
Anthropic docs (Claude): Glossary (temperature)
OpenAI Help Center: Temporary Chat FAQ
Claude Help Center: Usage limit best practices
OpenAI Help Center: ChatGPT Free Tier FAQ
Anthropic Engineering: Building effective agents (2024)
Anthropic docs (Claude): API overview
Anthropic docs (Claude): Computer use tool
OpenAI: Computer-Using Agent (2025)
Model Context Protocol docs: Introduction
Anthropic: Introducing the Model Context Protocol (2024)
Make Help Center: Scenarios
Anthropic docs (Claude): Tool use overview
Andrej Karpathy, post on X that coined the term (February 2, 2025)
Make Help Center: Webhooks
Gemini API docs: Embeddings
OpenAI API docs: Vector embeddings
Lewis et al. (Meta AI): Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020)
Google Cloud: What is semantic search?
Google Cloud: What is a vector database?
Gemini API docs: Image generation and editing
Ho, Jain, Abbeel: Denoising Diffusion Probabilistic Models (2020)
Gemini API docs: Generate videos with Veo
Google Cloud Vision docs: Detect text in images (OCR)
Radford et al. (OpenAI): Robust Speech Recognition via Large-Scale Weak Supervision (Whisper, 2022)
OpenAI Help Center: ChatGPT Images
Gemini API docs: Image generation
ElevenLabs docs: Text to Speech
ElevenLabs docs: Voices (voice cloning)
Google DeepMind: SynthID
C2PA: Coalition for Content Provenance and Authenticity
OpenAI Help Center: C2PA in ChatGPT Images
EU AI Act, Regulation (EU) 2024/1689, Article 3(60) and Article 50
Anthropic docs (Claude): Mitigate jailbreaks and prompt injections
OpenAI: Understanding prompt injections (2025)
OpenAI Help Center: How your data is used to improve model performance

AI glossary index: all terms A to Z

Basics

Artificial general intelligence (AGI)

Artificial intelligence (AI)

Deep learning

Foundation model

Generative AI

Large language model (LLM)

Machine learning

Neural network

How models work

Benchmark

Context window

Fine-tuning

Inference

Knowledge cutoff

Model distillation

Multimodal

Open-weight model

Parameters

Pretraining

Reasoning model

Reinforcement learning from human feedback (RLHF)

Token

Transformer

Using chatbots and prompts

Chain-of-thought prompting

Chatbot memory

Custom instructions

Deep Research

Few-shot prompting

Grounding

Hallucination

Prompt

System prompt

Temperature

Temporary chat

Usage limit

Agents, tools and automation

AI agent

API

Computer use

Model Context Protocol (MCP)

No-code automation

Tool use (function calling)

Vibe coding

Webhook

Retrieval and data

Embeddings

Retrieval-augmented generation (RAG)

Semantic search

Vector database

Images, video and voice

AI image editing

Diffusion model

Image-to-video

Optical character recognition (OCR)

Speech-to-text

Text-to-image

Text-to-speech (TTS)

Voice cloning

Safety, privacy and trust

AI watermarking

Content Credentials (C2PA)

Deepfake

Jailbreak

Prompt injection

Training on user data

How we built this AI glossary

Sources

Top 5 AI stories this week, in your inbox