AI moves fast and the vocabulary moves faster. If you have been trying to follow along and feel like everyone else got a dictionary you missed, this is that dictionary.
This is not a computer science textbook. It is a plain-language guide to the terms you are most likely to hear in conversations, news articles, podcasts, and sales pitches about artificial intelligence in 2026. We skip the deep math and focus on what the words actually mean and why they matter.
What Is AGI (Artificial General Intelligence)?
AGI refers to a hypothetical AI system that can perform any intellectual task a human can. Current AI systems are narrow: they are very good at specific things (writing, coding, image generation) but cannot generalize across all domains the way a person can. AGI is the goalpost that companies like OpenAI and Anthropic are working toward, though there is serious debate about how close we are, how we would know when we get there, and whether the concept is even well-defined.
What Is an AI Agent?
An AI agent is a system that can take actions on your behalf, not just generate text. Instead of answering “here’s how you could book that flight,” an agent would actually go book the flight. Agents typically combine a large language model with the ability to use tools: browsing the web, calling APIs, clicking through software, reading and writing files. As of early 2026, agents are real but still early. They work well for structured tasks and remain hit-or-miss for anything that requires complex judgment.
What Is an AI Wrapper?
An AI wrapper is a product built on top of someone else’s AI model, usually through an API. The wrapper adds a user interface, a specific workflow, or a niche focus, but the underlying intelligence comes from a model like Claude or GPT. “Wrapper” is sometimes used dismissively (implying the product adds little value), but many genuinely useful products are wrappers. The term is worth knowing because it helps you evaluate what you are actually paying for.
What Is Alignment?
Alignment is the field of research focused on making AI systems do what humans actually want them to do. This sounds obvious, but it turns out to be extremely hard. A system optimizing for a goal can find unexpected and undesirable shortcuts. Alignment work ranges from practical concerns (making chatbots refuse harmful requests) to existential ones (ensuring a superintelligent system does not pursue goals that conflict with human survival). It is one of the central research priorities at Anthropic.
What Is Anthropic?
Anthropic is an AI safety company founded in 2021 by Dario Amodei, Daniela Amodei, and other former OpenAI researchers. Anthropic builds the Claude family of AI models. The company’s founding thesis is that AI development and AI safety research should happen together, not separately. Anthropic is headquartered in San Francisco and is one of the three leading frontier AI labs alongside OpenAI and Google DeepMind.
What Is an API?
API stands for Application Programming Interface. In the AI context, an API is how software talks to an AI model. When you use Claude in the chat interface, you are using it through a consumer product. When a developer connects Claude to their own app, they use the API. The distinction matters because API access is how most business automation and custom AI tools get built. If someone tells you they are “building on the Claude API,” they mean they are writing code that sends requests to Claude and gets responses back.
What Is a Benchmark?
A benchmark is a standardized test used to measure how well an AI model performs. Common examples include MMLU (general knowledge), HumanEval (coding), and GPQA (graduate-level science questions). Benchmarks are useful for rough comparisons between models, but they have limits. Models can be optimized to score well on specific benchmarks without actually being better at real-world tasks, a problem the AI community calls “teaching to the test.” Take benchmark scores as directional, not definitive.
What Is a Chatbot?
A chatbot is any software that conducts a conversation with a human through text or voice. The term predates modern AI by decades. Today it usually refers to AI-powered conversational interfaces like ChatGPT, Claude, or Gemini. The word can feel reductive when applied to current systems, which are far more capable than the rule-based chatbots of the early internet, but it persists because the interaction pattern (you type, it responds) remains the same.
What Is ChatGPT?
ChatGPT is OpenAI’s consumer chatbot product, launched in November 2022. It was the application that brought large language models into mainstream awareness and triggered the current AI boom. ChatGPT is built on OpenAI’s GPT family of models. The product is free at a basic tier, with a paid “Plus” subscription for access to more powerful models and features. ChatGPT is often used as a generic term for AI assistants, much like “Kleenex” for tissues, but it refers specifically to OpenAI’s product.
What Is Claude?
Claude is Anthropic’s family of AI models and the consumer product built around them. The current generation is the Claude 4.5 family, which includes Opus (the most capable), Sonnet (balanced performance and speed), and Haiku (fast and lightweight). Claude is known for strong writing ability, careful reasoning, and a design philosophy that prioritizes safety and honesty. You can use Claude through the web interface at claude.ai, through mobile apps, or through the API for custom development.
What Is ClaudeBot?
ClaudeBot is Anthropic’s web crawler, the automated program that visits websites to gather information for training and improving Claude’s models. If you run a website and check your server logs, you may see visits from ClaudeBot, just as you would see visits from Googlebot (Google’s crawler). Website owners can allow or block ClaudeBot through their robots.txt file. ClaudeBot matters for businesses thinking about AI visibility because allowing it (and other AI crawlers like GPTBot and PerplexityBot) is one of the first steps toward having your content appear in AI-generated answers.
What Is a Context Window?
The context window is the amount of text an AI model can “see” at once during a conversation. Think of it as the model’s working memory. If you paste a 100-page document into a chat, the model can only use it if the document fits within its context window. Context windows are measured in tokens. Current frontier models have context windows ranging from 128,000 to over 1 million tokens. Larger context windows let models work with longer documents and maintain coherence over longer conversations.
What Is a Copilot?
Copilot is Microsoft’s brand name for AI assistants integrated across its product suite, including Windows, Office, and GitHub. More broadly, “copilot” has become a generic term for AI that assists you while you work, as opposed to doing the work entirely on its own. The metaphor is deliberate: a copilot helps the pilot but does not replace them. You will hear the term used both for Microsoft’s specific products and as a general description of AI-assisted workflows.
What Is a Deepfake?
A deepfake is AI-generated media (typically video or audio) that convincingly depicts someone saying or doing something they never actually said or did. The technology uses deep learning to swap faces, clone voices, or fabricate entire scenes. Deepfakes raise serious concerns around misinformation, fraud, and consent. The term is also used loosely to describe any AI-generated media that mimics real people, even when the intent is not deceptive.
What Is a Diffusion Model?
A diffusion model is the type of AI architecture behind most modern image generators, including Midjourney, DALL-E, and Stable Diffusion. The process works by training a model to remove noise from images. At generation time, it starts with pure noise and gradually refines it into a coherent image based on your text prompt. You do not need to understand the math. Just know that when someone says “diffusion model,” they mean the engine behind AI image and video generation.
What Is an Embedding?
An embedding is a way of representing text (or images, or other data) as a list of numbers that captures its meaning. Words or sentences with similar meanings end up with similar numbers. This is how AI systems understand that “dog” and “puppy” are related even though the letters are completely different. Embeddings power search, recommendation systems, and the retrieval step in RAG (see below). They are a foundational building block that most people never interact with directly.
What Is Fine-Tuning?
Fine-tuning is the process of taking a pre-trained AI model and training it further on a specific dataset to specialize its behavior. A company might fine-tune a model on their internal documents so it better understands their terminology and processes. Fine-tuning is more involved than prompt engineering (where you just write better instructions) but less expensive than training a model from scratch. Think of it as the difference between hiring a generalist and giving them industry-specific training.
What Is a Foundation Model?
A foundation model is a large AI model trained on broad data that can be adapted to many different tasks. GPT-4, Claude, Gemini, and Llama are all foundation models. The “foundation” metaphor is apt: these models serve as the base layer on which specific applications are built. Before foundation models, AI development usually meant training a separate model for each task. Now, a single foundation model can write code, analyze documents, generate images, and carry on a conversation.
What Is a Frontier Model?
A frontier model is the most capable AI model available at any given time, pushing the boundary of what AI can do. As of early 2026, the frontier models include Claude Opus 4.5 (Anthropic), GPT-4.5 and o3 (OpenAI), and Gemini 2.5 Pro (Google). The term distinguishes these cutting-edge systems from smaller or older models. Frontier models are relevant to policy discussions because their capabilities raise the most urgent questions about safety and regulation.
What Is Gemini?
Gemini is Google DeepMind’s family of AI models and the brand name for Google’s consumer AI products. Gemini replaced the earlier “Bard” chatbot and is integrated across Google’s product suite, including Search, Workspace, and Android. Google’s AI research division, DeepMind, is widely regarded as one of the top AI research organizations in the world. Gemini models compete directly with Claude and GPT at the frontier.
What Is Generative AI?
Generative AI refers to AI systems that create new content: text, images, audio, video, or code. This is in contrast to AI systems that classify, predict, or analyze existing data. When people say “AI” casually in 2026, they almost always mean generative AI. The entire current wave of AI products, from ChatGPT to Midjourney to AI coding assistants, falls under this umbrella.
What Is GPT?
GPT stands for Generative Pre-trained Transformer. It is the model architecture developed by OpenAI, and the name behind ChatGPT. The acronym tells you three things: it generates content (Generative), it was trained on a massive corpus before being fine-tuned (Pre-trained), and it uses the Transformer architecture (see below). GPT-4, GPT-4.5, and the “o” series reasoning models are all part of OpenAI’s GPT lineage.
What Are Guardrails?
Guardrails are the rules, filters, and constraints built into AI systems to prevent harmful, misleading, or unwanted outputs. When Claude declines to help you build a weapon or when ChatGPT adds a disclaimer to medical advice, those are guardrails at work. The term is also used more broadly in business contexts to describe any policy or technical measure that keeps AI outputs within acceptable bounds for a specific use case.
What Is a Hallucination?
A hallucination is when an AI model generates information that sounds confident and plausible but is factually wrong. The model might cite a study that does not exist, invent a historical event, or confidently state an incorrect statistic. Hallucinations happen because language models are predicting likely text, not looking up verified facts. This is one of the most important limitations to understand when using any AI system. Always verify claims that matter.
What Is Inference?
Inference is the process of running a trained AI model to generate an output. When you send a message to Claude and get a response, that is inference. Training is when the model learns; inference is when it performs. The distinction matters because training is extremely expensive (millions of dollars for frontier models) while inference is comparatively cheap per query. When companies talk about “inference costs,” they mean the cost of actually running the model for users.
What Is a Knowledge Cutoff?
A knowledge cutoff is the date beyond which an AI model has no training data. The model simply does not know about events, publications, or developments that occurred after that date. This is why AI assistants sometimes give outdated answers about recent events. Many current systems compensate by using web search to supplement their training knowledge. When a model tells you its knowledge cutoff, it is being transparent about the boundary of what it learned during training versus what it can look up in real time.
What Is a Large Language Model (LLM)?
A large language model is an AI system trained on massive amounts of text to predict and generate language. “Large” refers to the number of parameters (typically billions to trillions) and the scale of training data. LLMs are the technology behind Claude, ChatGPT, Gemini, and most other AI assistants. They work by learning statistical patterns in language at enormous scale, which turns out to produce surprisingly capable reasoning, writing, and problem-solving abilities. LLM is the technical term you will hear most often in AI discussions.
What Is Llama?
Llama is Meta’s family of open-weight language models. Meta releases the model weights publicly, which means anyone can download, run, and modify them. This makes Llama important for researchers, startups, and organizations that want to run AI models on their own infrastructure rather than depending on an API provider. Llama models are not as capable as the top closed models, but they are “good enough” for many applications and their openness has made them enormously influential.
What Is Machine Learning?
Machine learning is the broader field of computer science in which systems learn patterns from data rather than following explicitly programmed rules. All modern AI, including large language models and image generators, is built on machine learning techniques. The term has been around since the 1950s. Think of it as the parent category: machine learning is the discipline, deep learning is a technique within it, and LLMs are a specific application of deep learning.
What Is MCP (Model Context Protocol)?
MCP is an open protocol developed by Anthropic that lets AI models connect to external tools and data sources in a standardized way. Before MCP, every integration between an AI and an outside service (like your calendar, your email, or a database) required custom code. MCP provides a common language so that a single integration can work across different AI models and applications. Think of it like USB for AI tools: one standard connector instead of a different cable for every device.
What Is Midjourney?
Midjourney is an AI image generation service known for producing highly aesthetic, stylized images from text prompts. It originally operated entirely through Discord (a chat platform) and has since launched a web interface. Midjourney, along with DALL-E (OpenAI) and Stable Diffusion (Stability AI), represents the leading wave of text-to-image AI tools. If someone shows you an eerily beautiful AI-generated image, there is a good chance Midjourney made it.
What Is Multimodal AI?
Multimodal AI refers to models that can process and generate more than one type of data. A text-only model is “unimodal.” A model that can read text, look at images, listen to audio, and generate any of those outputs is multimodal. Most frontier models in 2026 are multimodal. Claude can read images and documents; GPT can process images and generate audio. The trend is toward models that handle all media types natively, just as humans do.
What Is a Neural Network?
A neural network is a computing system loosely inspired by the structure of biological brains. It consists of layers of interconnected nodes (“neurons”) that process information. Data enters through an input layer, passes through hidden layers that transform it, and exits through an output layer. Neural networks are the foundational architecture behind virtually all modern AI. Deep learning just means using neural networks with many layers (hence “deep”). You do not need to understand the internals, just that this is the engine under the hood.
What Does Open Source vs. Closed Source Mean in AI?
Open source in AI means the model’s code and weights are publicly available for anyone to use, modify, and distribute. Closed source (or proprietary) means the company keeps the model private and you can only access it through their products or API. Llama (Meta) and Mistral are leading open models. Claude (Anthropic), GPT (OpenAI), and Gemini (Google) are closed. There is also a middle ground, “open weight,” where the model weights are released but the training data and full methodology are not. The open vs. closed debate is one of the most consequential policy questions in AI right now.
What Is OpenAI?
OpenAI is the AI company that built ChatGPT and the GPT family of models. Founded in 2015 as a nonprofit, it restructured into a “capped profit” company and has since moved toward a more traditional corporate structure. Sam Altman is the CEO. OpenAI’s release of ChatGPT in November 2022 is widely credited with starting the current AI boom. The company is one of the three leading frontier AI labs, alongside Anthropic and Google DeepMind.
What Are Parameters?
Parameters are the internal values that a neural network learns during training. They are the “knowledge” of the model, encoded as numbers. When people say a model has “70 billion parameters” or “1 trillion parameters,” they are describing its size and, roughly, its capacity to learn complex patterns. More parameters generally means more capable, but also more expensive to train and run. Parameter count is a useful rough proxy for model size, though architecture and training data matter as much or more.
What Is Pre-Training?
Pre-training is the initial, large-scale training phase where a model learns from an enormous dataset of text, code, images, or other content. This is the most expensive and time-consuming part of building an AI model, often costing tens or hundreds of millions of dollars and taking weeks or months on massive clusters of specialized hardware. During pre-training, the model learns general patterns in language, facts about the world, reasoning abilities, and coding skills. After pre-training, the model typically goes through additional stages (like RLHF) to make it more helpful, safe, and conversational. When someone says a model was “pre-trained on internet text,” they are describing this foundational learning phase.
What Is a Prompt?
A prompt is the input you give to an AI model. In a chatbot, your message is the prompt. In an API call, the prompt is the text you send to the model. The quality of your output depends heavily on the quality of your prompt. A vague prompt gets a vague answer. A specific, well-structured prompt with context and examples gets dramatically better results. The term is simple but the concept is central to getting value out of any AI tool.
What Is Prompt Engineering?
Prompt engineering is the practice of crafting effective prompts to get better outputs from AI models. This ranges from simple techniques (being specific, providing examples) to sophisticated approaches (chain-of-thought reasoning, multi-step workflows, system prompts). Prompt engineering emerged as a skill set because small changes in how you ask can produce large changes in what you get back. Some people do this as a full-time job. For most users, learning a few core techniques makes an enormous difference.
What Is RAG (Retrieval-Augmented Generation)?
RAG is a technique that combines an AI model with a search step. Instead of relying only on what the model learned during training, RAG first retrieves relevant documents from a database or knowledge base, then passes those documents to the model along with the user’s question. This lets the model give answers grounded in specific, current, verified information rather than relying on memory alone. RAG is how most enterprise AI applications work, because businesses need the model to answer questions about their specific data, not just general knowledge.
What Is a Reasoning Model?
A reasoning model is an AI model specifically designed to “think through” complex problems step by step before giving an answer. OpenAI’s o1, o3, and o4-mini are the most prominent examples. These models trade speed for accuracy on hard problems: they take longer to respond but perform better on math, science, coding, and logic tasks. The model essentially generates an internal chain of reasoning before producing its final answer, somewhat like how you might work through a problem on scratch paper before writing your solution.
What Is RLHF?
RLHF stands for Reinforcement Learning from Human Feedback. It is a training technique where human evaluators rate the model’s outputs, and the model learns to produce responses that humans prefer. RLHF is a key reason modern AI assistants feel helpful and coherent rather than raw and unpredictable. Without RLHF (and related techniques like RLAIF, which uses AI feedback), a language model would generate text that is statistically plausible but not necessarily useful, safe, or aligned with what you actually want.
What Are Scaling Laws?
Scaling laws are the empirical finding that AI models get predictably better as you increase three things: the amount of training data, the number of parameters, and the amount of compute used for training. This discovery, formalized in research from OpenAI and others around 2020, is the intellectual foundation of the current AI boom. It is why companies are spending billions on larger models and data centers. The bet is that scaling will continue to produce more capable systems. Whether this trend will hit a wall is one of the central debates in AI.
What Is “Slop”?
Slop is the informal term for low-quality, obviously AI-generated content. It is the AI equivalent of spam. You know it when you see it: generic blog posts stuffed with phrases like “in the ever-evolving landscape,” images with too many fingers, LinkedIn posts that feel like they were written by a committee of robots. The term emerged from widespread frustration with the flood of mediocre AI content across the internet. When people use “slop” critically, they are usually pointing at content where someone used AI without any taste, editing, or quality control.
What Is a System Prompt?
A system prompt is a set of instructions given to an AI model that shapes its behavior for an entire conversation. Unlike a user’s message, the system prompt is typically hidden from the end user and set by the developer or application. It might tell the model to respond in a certain tone, stay on certain topics, or follow specific formatting rules. System prompts are how businesses customize AI behavior for their specific use case without fine-tuning the model itself.
What Is Temperature?
Temperature is a setting that controls how random or creative an AI model’s outputs are. A low temperature (like 0.0) makes the model more deterministic and predictable, choosing the most likely next word every time. A high temperature (like 1.0) makes it more creative and varied, but also more prone to strange or incoherent outputs. You can think of it as a “creativity dial.” Most AI interfaces set this automatically, but API users can adjust it for their specific needs.
What Is a Token?
A token is the basic unit of text that AI models process. Tokens are not exactly words. They are chunks of text, typically three to four characters. The word “hamburger” might be split into “ham,” “bur,” and “ger,” which counts as three tokens. Tokens matter because AI pricing, context windows, and rate limits are all measured in tokens. As a very rough rule of thumb, one token is about three-quarters of a word, or 750 words is roughly 1,000 tokens.
What Is Training Data?
Training data is the massive collection of text, images, code, and other content used to train an AI model. For large language models, this typically includes large portions of the public internet, books, academic papers, and code repositories. The quality and composition of training data heavily influences what the model knows, what it is good at, and what biases it carries. Debates about training data, including whether copyrighted material should be used, are among the most active legal and ethical questions in AI.
What Is a Transformer?
The Transformer is the neural network architecture behind virtually all modern large language models. Introduced in a 2017 paper titled “Attention Is All You Need,” the Transformer’s key innovation is a mechanism called “attention,” which lets the model weigh the relevance of every word against every other word in its input. This is what allows LLMs to understand context, follow long passages, and generate coherent text. The Transformer is to the current AI era what the internal combustion engine was to the automotive age: the foundational technology that made everything else possible.
What Is Vibe Coding?
Vibe coding is the practice of building software by describing what you want in natural language and letting an AI write the actual code. Instead of writing Python or JavaScript yourself, you tell Claude or another AI assistant what the program should do, review the output, and iterate through conversation. The term was coined by Andrej Karpathy (former OpenAI, Tesla AI lead) in early 2025 and quickly entered common usage. Vibe coding has made basic software development accessible to people with no traditional programming background, though complex projects still benefit from understanding what the code is actually doing.
What Are Weights?
Weights are the numerical values inside a neural network that determine how it processes information. When a model is “trained,” the training process adjusts millions or billions of weights until the model produces good outputs. The weights are, in a real sense, the model. When Meta releases “open weights” for Llama, they are releasing the actual trained model itself, which anyone can then run on their own hardware. “Weights” and “parameters” are closely related terms and often used interchangeably in casual conversation.
Frequently Asked Questions
What is the difference between AI, machine learning, and deep learning? AI is the broadest term, covering any system that performs tasks typically requiring human intelligence. Machine learning is a subset of AI where systems learn from data. Deep learning is a subset of machine learning that uses neural networks with many layers. Most current AI products use deep learning.
What is the difference between open source and closed source AI models? Open source models release their code and weights publicly. Closed source models keep them private. Open models (like Llama) offer flexibility and independence. Closed models (like Claude and GPT) tend to be more capable and come with managed infrastructure. The choice depends on your needs around control, capability, and cost.
Which AI model is the best? There is no single best model. Claude, GPT, and Gemini each have strengths. For most business users, the best model is the one that performs well on your specific tasks at a price you can afford. Try a few and compare results on your actual use cases rather than relying on benchmark scores alone.
Do I need to learn to code to use AI? No. Consumer products like Claude, ChatGPT, and Gemini require no coding at all. For more advanced use cases (automation, custom integrations, fine-tuning), some technical knowledge helps, but vibe coding and no-code tools are lowering that bar rapidly.
What is the difference between a model and a product? A model (like Claude Sonnet 4.5) is the underlying AI system. A product (like claude.ai or ChatGPT) is the interface and set of features built around the model. The same model can power many different products through an API.
This glossary was last updated in March 2026. AI vocabulary changes fast, so we will update this page as new terms enter common usage. Have a term you think we should add? Get in touch.
Arrow & Bell helps businesses navigate AI adoption with clear language, not jargon. If you are trying to figure out what AI means for your operations, start a conversation.