AI Glossary 2026: 80+ Terms Every Professional Must Know

Last Updated: August 11, 2026

This AI glossary defines over 80 essential artificial intelligence terms in plain English for business professionals - from foundational concepts like large language models and neural networks to 2026-specific terms like agentic AI, vibe coding, GEO, and reasoning models - with every definition written for executives, managers, and professionals who need to use these terms confidently, not engineers who already know them.

AI vocabulary has never moved faster. Terms that did not exist two years ago - agentic AI, vibe coding, GEO, reasoning models, physical AI - are now appearing in board presentations, vendor contracts, and job descriptions. Meanwhile, foundational terms like tokens, embeddings, and context windows are still misunderstood by most of the professionals being asked to make decisions about AI tools and budgets.

Part 1 covers foundational AI concepts, how AI models work, context and retrieval, and prompting. Part 2 covers agentic AI, performance metrics, AI platforms, enterprise governance, and 2026-specific terms.

Use Ctrl+F or Cmd+F to search for any specific term.

The terms in this section are the foundation of every AI conversation. If someone on your team or a vendor uses a term you do not recognize, it is almost certainly defined here.

Artificial Intelligence (AI)

The broad field of computer science focused on building systems that can perform tasks that previously required human intelligence - understanding language, recognizing images, making decisions, and solving problems. In 2026, AI typically refers to machine learning-based systems rather than the rule-based systems that dominated earlier decades. When a business leader says "we are deploying AI," they almost always mean machine learning, specifically generative AI or predictive AI.

Related: What Is Artificial Intelligence?

Machine Learning (ML)

A subset of AI in which systems improve their performance on tasks through exposure to data rather than through explicit programming. Instead of writing rules like "if the email contains this phrase, mark it as spam," machine learning systems learn the pattern of spam from thousands of examples. Most modern AI applications - from recommendation engines to fraud detection to language models - are built on machine learning. When executives talk about AI implementation, they are almost always referring to machine learning applications.

Related: What Is Machine Learning?

Deep Learning

A type of machine learning that uses neural networks with many layers to learn patterns from large amounts of data. Deep learning is the technology behind image recognition, speech recognition, and large language models. The "deep" refers to the many layers in the neural network, not to the sophistication of its understanding. Deep learning requires significantly more data and computing power than earlier machine learning approaches but produces dramatically better results on complex tasks like natural language and image understanding.

Related: What Is Deep Learning?

Neural Network

A computational architecture loosely inspired by the human brain - not in its biological mechanics but in its organizational principle of many interconnected nodes passing signals to each other. Neural networks learn by adjusting the strength of connections between nodes based on examples until the network produces correct outputs. Every modern AI application from ChatGPT to facial recognition to self-driving cars is built on neural networks. Understanding them requires no mathematics - they learn patterns from examples the same way a person learns to recognize a face from seeing many faces.

Related: What Is a Neural Network?

Generative AI

AI systems that create new content - text, images, video, audio, or code - rather than simply classifying or analyzing existing content. ChatGPT generating a business proposal, Midjourney creating a marketing image, and Veo 3.1 producing a video are all generative AI. The distinction matters for enterprise deployments: generative AI creates outputs that require human review, while predictive AI (recommending products, detecting fraud) typically produces decisions that get acted on automatically. Generative AI has the highest potential impact and the highest hallucination risk of any AI category.

Related: What Is Generative AI?

Large Language Model (LLM)

A type of AI model trained on massive amounts of text data to understand and generate human language. ChatGPT, Claude, Gemini, and Grok are all large language models. The "large" refers to the number of parameters - mathematical values learned during training - which for frontier models now range from tens of billions to over a trillion. LLMs do not understand language the way humans do. They predict the most likely next token given everything that came before. That prediction process, repeated billions of times across massive datasets, produces outputs that appear to demonstrate understanding, reasoning, and creativity.

Related: What Is an LLM?

Foundation Model

A large AI model trained on broad, general data that can be adapted to a wide range of specific tasks through fine-tuning or prompting. GPT-5.6, Claude Opus 4.8, and Gemini 3.1 Pro are all foundation models. The term "foundation" reflects that these models are the base layer on which many applications are built - a company building an AI customer service tool typically builds on top of a foundation model rather than training a model from scratch. Foundation models represent the most significant shift in enterprise AI economics because they make capability accessible without requiring billions of dollars in training infrastructure.

Related: What Is a Foundation Model?

Natural Language Processing (NLP)

The AI discipline focused on enabling computers to understand, interpret, and generate human language. NLP powers everything from voice assistants and chatbots to sentiment analysis and machine translation. Modern large language models are the most advanced NLP systems ever built. When your CRM flags a support ticket as high-priority based on the customer's language, or your email client summarizes a long thread, that is NLP. The distinction between NLP and LLMs: NLP is the broader field; LLMs are the current state-of-the-art technology within it.

Related: Natural Language Processing Explained

Computer Vision

The AI discipline enabling machines to interpret and understand visual information from images and video. Computer vision powers facial recognition, medical imaging analysis, quality control in manufacturing, autonomous vehicle navigation, and product search by image. Modern computer vision systems using deep learning achieve superhuman accuracy on specific tasks like identifying cancerous cells in medical scans. As cameras become ubiquitous in physical environments, computer vision is becoming one of the most economically significant AI applications.

Related: What Is Computer Vision?

Multimodal AI

An AI system that can process and generate multiple types of data - text, images, audio, and video - in a single unified model. GPT-5.6 Sol, Claude Opus 4.8, and Gemini 3.1 Pro are all multimodal. You can show them an image and ask questions about it, describe a chart in words and ask for analysis, or provide audio and request a transcript with sentiment analysis. In 2026, multimodal capability has become standard in frontier models. The business value: AI that works with information the way humans do, across formats, rather than requiring separate tools for each data type.

Related: Multimodal AI Explained

Artificial General Intelligence (AGI)

A hypothetical AI system with human-level intelligence across all cognitive domains - able to learn any task a human can learn, reason across domains, and adapt to genuinely novel situations without retraining. No AGI exists in 2026. Current AI systems are narrow specialists, extraordinarily capable in their trained domains but brittle outside them. AGI is the stated long-term goal of OpenAI, Anthropic, and Google DeepMind, though definitions of what constitutes AGI vary significantly across organizations. Anthropic CEO Dario Amodei has suggested AI broadly smarter than a Nobel Prize winner could arrive as early as 2026-2027.

Related: What Is AGI?

Understanding these terms helps you evaluate vendor claims, make sense of benchmark data, and ask the right questions when purchasing AI tools.

Training

The process of exposing an AI model to large amounts of data so it can learn patterns and improve its performance on tasks. Training a frontier language model like GPT-5.6 or Claude requires billions of examples, months of computation on thousands of specialized chips, and costs tens to hundreds of millions of dollars. Training is typically done once by the AI company. When you use ChatGPT or Claude, you are not training the model - you are using an already-trained model in inference mode. Understanding the distinction matters for enterprise AI: organizations that build on existing foundation models pay inference costs, not training costs.

Inference

Running a trained AI model to generate outputs - the process that happens every time you send a message to ChatGPT, Claude, or any AI tool. Inference is computationally cheaper than training but still requires significant infrastructure at the scale of millions of daily users. Inference costs have fallen dramatically - Google reported a 33x improvement in inference efficiency per query in 12 months per our AI energy statistics guide. Understanding inference versus training matters for enterprise AI: organizations that build on existing foundation models pay inference costs, not training costs.

Related: AI Inference Explained

Fine-tuning

Adapting a pre-trained foundation model to a specific task or domain by continuing to train it on a smaller, specialized dataset. A law firm might fine-tune Claude on its case files to produce a model that understands legal terminology and formats. A healthcare provider might fine-tune a model on clinical notes. Fine-tuning produces a model that performs better on specific tasks than the general foundation model while preserving the general capabilities acquired during pre-training. In 2026, fine-tuning has become accessible through cloud APIs, no longer requiring specialized ML teams.

Related: AI Fine-Tuning Explained

Parameters

The numerical values a neural network learns during training - the adjustable knobs that determine how the model processes and responds to inputs. When you hear that GPT-5.6 has hundreds of billions of parameters or that Llama 4 Maverick has 400 billion parameters, those numbers represent the scale of learned knowledge encoded in the model. More parameters generally means more capacity to learn complex patterns, though efficient architectures like Mixture-of-Experts can achieve strong performance with fewer active parameters than their total count suggests. For business leaders: parameter count is a rough proxy for capability, not a guarantee of it.

RLHF (Reinforcement Learning from Human Feedback)

The training technique that made modern AI assistants helpful, harmless, and honest by incorporating human preferences into the learning process. After initial pre-training, models are refined using feedback from human raters who evaluate responses for quality, safety, and helpfulness. The model learns to generate responses that humans rate favorably. RLHF is why ChatGPT and Claude feel helpful and conversational rather than outputting raw text prediction artifacts. It is also why models sometimes tell you what they think you want to hear rather than what is true - the alignment challenge that makes AI safety research important.

Constitutional AI

Anthropic's approach to AI alignment that trains models using a set of written principles - a "constitution" - rather than purely human preference ratings. The model evaluates its own outputs against the principles and revises them, reducing dependence on human feedback at scale. Constitutional AI is Anthropic's proprietary safety architecture and a key differentiator in enterprise sales conversations where regulated industries need documented safety frameworks. For compliance and procurement teams: Constitutional AI provides the kind of explainable governance documentation that traditional RLHF approaches do not.

Transformer

The neural network architecture that underlies virtually all modern large language models - introduced by Google researchers in 2017 in the paper "Attention Is All You Need." The Transformer's key innovation is the attention mechanism, which allows the model to consider all parts of the input simultaneously rather than processing it sequentially. This enables language models to maintain coherent context across long passages and understand complex relationships between words and concepts. The T in GPT stands for Transformer.

Related: Transformer Model Explained

Attention Mechanism

The mathematical technique inside Transformer models that determines which parts of the input are most relevant to each output. When generating the word "it" in a sentence, the attention mechanism calculates which earlier word "it" most likely refers to and weights that context heavily. In practical terms: attention is why modern language models can answer questions about specific sections of a 200-page document you uploaded - the model can attend to the relevant passages rather than processing everything with equal weight.

Temperature

A parameter controlling the randomness of AI model outputs - how creative or conservative the model is when choosing its next word. Temperature 0 makes the model deterministic, always choosing the most probable next token. Temperature 1 introduces significant randomness, producing more creative but less reliable outputs. For business applications requiring accuracy and consistency (contract review, data extraction, factual Q&A), lower temperature settings are preferable. For creative tasks (marketing copy, brainstorming, storytelling), higher temperature produces more varied and interesting outputs.

Mixture-of-Experts (MoE)

An AI architecture in which a large model is divided into specialized sub-networks, with a routing mechanism that activates only the most relevant experts for each input rather than using the entire model. Llama 4 Maverick uses MoE. Mistral's models use MoE. The advantage: MoE models can have a very large total parameter count for knowledge capacity while requiring far less compute per inference by only activating a fraction of parameters per query. For enterprise buyers: MoE architecture is why some large models are cheaper to run than their parameter count would suggest.

Inference Cost

The computational expense of running an AI model to generate a response, typically measured in dollars per million tokens of input and output. In 2026, inference costs have fallen dramatically - Grok 4.1 Fast processes a million input tokens for $0.20. OpenAI's GPT-5.6 Sol costs $2 per million input tokens. Understanding inference costs is essential for any business building AI-powered products because the per-query cost multiplies by query volume. A feature used by 100,000 users daily at 5,000 tokens per session produces very different economics than an internal tool used by 50 employees.

Related: AI Pricing Guide 2026

These terms explain how AI models handle information - what they can see, what they remember, and how they access knowledge beyond their training data.

Context Window

The maximum amount of text - including your input, any documents you provide, conversation history, and the model's responses - that an AI model can process at once in a single session. Think of it as the model's working memory. Anything outside the context window is invisible to the model. In August 2026, context windows range from 128K tokens on ChatGPT Free to 2 million tokens on Gemini Advanced and Grok 4.1 Fast. One token equals approximately three-quarters of a word in English. A 200K context window holds approximately 150,000 words - roughly 500 pages of standard text.

Related: What Is a Context Window?

Token

The unit AI models use to process text - roughly three-quarters of a word in English, though tokenization varies by language and content type. The word "unbelievable" might be two tokens. A space before a word is often a separate token. Code tokenizes differently from prose. Tokens matter because AI pricing is calculated per token (input and output separately), context windows are measured in tokens, and rate limits are set in tokens per minute. When a model's pricing shows $2 per million input tokens, every word you send costs approximately 1.33 tokens.

Embeddings

Mathematical representations of text, images, or other data as lists of numbers that capture their meaning and relationships. Words or concepts with similar meanings have embeddings that are mathematically close to each other. Embeddings are the technology that enables semantic search - searching by meaning rather than exact keywords - and power features like "find emails similar to this one" or "recommend products related to what this customer bought." In enterprise AI, embedding models are often the invisible layer underneath the visible AI features.

Related: What Are Embeddings?

Vector Database

A specialized database designed to store and search embeddings by semantic similarity rather than exact keyword matching. When an AI application needs to find the most relevant documents from a large knowledge base to include in a prompt, a vector database retrieves them based on meaning proximity. Pinecone, Weaviate, and Qdrant are leading vector database providers. Vector databases are the infrastructure layer underneath RAG systems and most enterprise AI knowledge management applications.

RAG (Retrieval-Augmented Generation)

A technique that connects an AI model to an external knowledge base, allowing it to retrieve relevant information before generating a response. Rather than relying solely on knowledge baked into the model during training, RAG pulls current and organization-specific information from databases, documents, or real-time sources. A RAG-enabled customer service bot retrieves the actual product documentation, current pricing, and real-time inventory data before answering a customer question - dramatically reducing hallucination and keeping information current without retraining the model. RAG is the most widely deployed enterprise AI architecture in 2026.

Related: What Is RAG?

Knowledge Cutoff

The date after which an AI model's training data does not include information - creating a blind spot for any events, products, or developments that occurred after that date. Claude's knowledge cutoff is early 2025. ChatGPT's standard models have an April 2024 cutoff. This matters practically: asking a model with a 2024 cutoff about a product launched in 2025 produces either a hallucinated answer or an "I don't have information about this" response. The fix: use models with real-time search capability (Perplexity, Grok DeepSearch, ChatGPT Deep Research) for any query where current information matters.

Grounding

Connecting an AI model's outputs to verifiable, real sources rather than relying on training data that may be outdated or incorrect. A grounded AI system cites specific sources for specific claims and retrieves information from authoritative databases rather than generating from memory. Perplexity's architecture is grounded - every claim links to a verifiable web source. NotebookLM is grounded in the documents you upload. Grounding is the technical implementation of the verification principle: the model cannot confidently state things it cannot point to a source for.

These terms cover how to get the best outputs from AI models - the growing professional skill of communicating with AI effectively.

Prompt

Any input you give to an AI model to generate a response. A prompt can be a question, an instruction, a template, or a combination of instruction and data. The quality of AI outputs is directly proportional to the quality of the prompt. Vague prompts produce vague outputs. Specific, structured prompts with context, constraints, and desired format produce significantly better results. The single most impactful prompt improvement: tell the AI what role to play, what task to perform, what constraints to follow, and what format to produce.

Related: AI Prompt Templates

Prompt Engineering

The discipline of designing, testing, and refining prompts to consistently produce high-quality AI outputs. Prompt engineering ranges from simple techniques (being specific, providing examples, specifying format) to advanced methods (chain-of-thought, system prompts, few-shot examples). In 2026, prompt engineering is a professional skill valued across marketing, operations, legal, finance, and engineering functions. The most important prompt engineering principle: tell the AI what role to play, what task to perform, what constraints to follow, and what format to produce.

Related: What Is Prompt Engineering?

System Prompt

Instructions given to an AI model before any user conversation begins - typically configured by the product or organization deploying the AI rather than the end user. A customer service chatbot's system prompt might say "You are a helpful assistant for our company. Only discuss our products. Always be professional. Escalate billing questions." System prompts set the AI's persona, constraints, knowledge scope, and behavioral rules. For enterprise AI deployments, well-crafted system prompts are as important as model selection for determining output quality and safety.

Zero-Shot Prompting

Asking an AI model to perform a task without providing any examples, relying entirely on the model's training to understand what is needed. "Classify this email as positive, negative, or neutral" with no examples is a zero-shot prompt. Modern frontier models handle many tasks well zero-shot. When zero-shot results are inconsistent or low quality, few-shot prompting typically improves them.

Few-Shot Prompting

Including a small number of examples in your prompt to show the AI model the format and quality of output you expect. Rather than just asking for email classification, you might show three examples of emails labeled positive, negative, and neutral before asking the model to classify a new one. Few-shot prompting dramatically improves consistency on specialized tasks, particularly formatting, style matching, and domain-specific classification. Most professional prompt engineering workflows use at least two to three examples for any task requiring specific output format.

Chain-of-Thought Prompting

Instructing an AI model to show its reasoning step by step before reaching a conclusion, rather than jumping directly to an answer. Adding "Think through this step by step before answering" to a complex analytical prompt consistently improves accuracy on reasoning-heavy tasks. Chain-of-thought prompting is particularly effective for math, logic, multi-step planning, and any task where intermediate reasoning steps affect the final answer. It is the conversational equivalent of asking a consultant to show their work rather than just deliver a recommendation.

Hallucination

When an AI model generates plausible-sounding but factually incorrect information - presenting fabricated statistics, non-existent sources, incorrect dates, or entirely invented facts with the same confident tone it uses for accurate information. Hallucination is the most significant quality risk in enterprise AI deployments. The technical cause: language models predict probable text rather than verifying facts. Mitigation: use grounded tools (Perplexity, NotebookLM) for factual research, verify specific claims against primary sources, and never use AI-generated citations without checking whether the cited source actually exists.

Related: AI Hallucinations: Causes and Solutions

Continue to Part 2 for agentic AI terms, AI platform definitions, enterprise governance vocabulary, and the 2026-specific terms now appearing in vendor conversations and job descriptions.