Glossary of AI Terms

Foundational AI Concepts Every Reader Should Know

The first cluster of any glossary of AI terms covers the four words that everything else builds on. Artificial intelligence describes systems that perform tasks normally requiring human cognition, including perception, reasoning, learning, and language use. A model is the trained mathematical function that turns input into output, while an algorithm is the recipe the model follows during training and inference. Data is the raw material the model learns from, and a parameter is one of the billions of internal numbers the model adjusts during training to fit that data. Together these four words frame every other concept in this guide.

Two further foundations matter for anyone trying to read 2026 product launches without losing time on the press release. Inference is the act of running a trained model on new input to produce an output, and it now drives most cloud AI cost. Training is the prior step where the model adjusts parameters by repeatedly minimizing error on labeled or unlabeled examples. A token is the smallest text unit a model processes, usually a fragment of a word, and pricing for commercial models is often quoted per million tokens. Latency is the time it takes for the model to return a response, and throughput is how many requests the system handles per second under load.

The third foundational cluster names the resources teams need to plan around. Compute means the processing power required for training and inference, usually measured in GPU hours or accelerator chips. A dataset is a structured collection of examples used for training, validation, or evaluation. A benchmark is a standardized test that compares models on the same task, with MMLU, GPQA, and SWE-Bench dominating headlines in 2026. Open source describes models with weights anyone can download, fine-tune, and deploy, while proprietary models stay behind APIs. Readers who want a deeper start can revisit our explainer on neural networks fundamentals for the math that underpins these foundations.

Machine Learning Vocabulary in Plain English

Building on those foundations, the machine learning cluster names the four main learning paradigms. Supervised learning trains a model on input examples that come paired with the correct output, so the model learns the mapping that turns one into the other. Unsupervised learning uses unlabeled data, leaving the model to find structure such as clusters or principal directions. Semi-supervised learning combines a small labeled set with a large unlabeled one. Reinforcement learning teaches a policy through trial, reward, and penalty, which is how robotics control and many agent loops get their behavior. Our paired explainers on supervised learning approaches walk through each paradigm with code-level examples.

The other essential ML words describe the parts of any pipeline. A feature is one column of input the model uses to predict, and feature engineering is the human work of designing those columns. Overfitting happens when a model memorizes training data and fails on new examples, while underfitting is the opposite weakness of failing to capture real signal. Regularization, dropout, and early stopping are the standard fixes most teams reach for when training overfits. Cross-validation splits the data multiple ways and averages the results to estimate generalization more honestly than a single hold-out set. Hyperparameters are the dials engineers set before training begins, such as learning rate or batch size. Transfer learning reuses weights from a model trained on a large general task to bootstrap a smaller specialized one. Our deeper write-up on transfer learning workflows traces this pattern across image, speech, and language work.

Deep Learning and Neural Network Terminology

Shifting focus to deep learning, the core building block is the neuron, a simple unit that multiplies inputs by weights, adds a bias, and applies a non-linear activation. A neural network is many of those neurons stacked into layers, with the depth of the stack giving deep learning its name and most of its modeling power. Forward propagation sends input through the layers to produce a prediction. Backpropagation then walks the gradient of the loss backward through the network to tell each weight how to change. Gradient descent applies those changes in small steps, and stochastic gradient descent does so on mini-batches rather than the full dataset.

The activation function decides whether a neuron fires and how strongly. Sigmoid and tanh dominated early networks, while ReLU and its variants run nearly every modern model because they avoid the vanishing gradient problem. The softmax activation in neural networks turns a vector of scores into a probability distribution, which is how classifiers and language models produce a final answer. Batch normalization stabilizes training by rescaling activations layer by layer. Dropout randomly zeros some activations during training so the network does not rely on any single neuron. Loss functions such as cross-entropy and mean squared error give the gradient something to follow.

Several architecture words inside the deep learning cluster define how layers connect inside a working model. A convolutional neural network shares weights across small filters that slide over images, which made image recognition tractable. A recurrent neural network feeds output from one step back as input to the next, fitting sequences. Long short-term memory adds gates that let the recurrence remember longer context. The transformer replaced recurrence in 2017 by letting every token attend to every other token through a mechanism called self-attention. Attention is now the dominant building block for language, vision, and audio. Multi-head attention runs several attention computations in parallel so the model can learn different relationships at once.

Two final deep learning ideas matter for 2026 procurement conversations. A foundation model is a large network pretrained on broad data, then adapted with fine-tuning or prompting to specific tasks. A mixture of experts splits a model into many specialist subnetworks and routes each token to the most relevant ones. That trick lets headline models like GPT-4o, Claude, and Gemini deliver large effective capacity at lower inference cost. Quantization compresses weights into low-precision integers so models fit on cheaper hardware. Distillation trains a smaller student model to mimic a larger teacher. These compression techniques explain why 2026 saw capable open weights run on a single laptop GPU.

Natural Language Processing and Speech AI Terms

Beyond the architecture vocabulary, NLP terms describe how text becomes something a model can compute on. Tokenization is the first step, splitting raw text into the integer ids the model actually processes, with sub-word algorithms like byte-pair encoding dominating modern systems. Our deep dive on tokenization in NLP walks through that process. Embeddings turn each token into a vector that captures meaning, so words with similar use land near each other in vector space. The companion explainer on word embeddings explained covers the math behind that vector space. A vocabulary is the fixed set of tokens the model knows, and out-of-vocabulary text gets handled through sub-word splits or special placeholder tokens.

Downstream NLP tasks each carry their own vocabulary that 2026 product teams now hear in vendor briefings every week. Named entity recognition tags people, places, and organizations inside any text span the model receives. Part-of-speech tagging labels words as nouns, verbs, or modifiers and underlies many simple grammar checkers. Sentiment analysis classifies text as positive, negative, or neutral, often with finer-grained categories. Machine translation maps text from one language into another using either neural sequence models or hybrid retrieval systems. Summarization compresses a long document into a short one, with abstractive variants generating new sentences and extractive variants pulling existing ones. Question answering produces a direct answer from a document or knowledge base. Each task now runs as a prompt to a general-purpose model rather than a custom pipeline, which is why 2026 NLP teams focus on evaluation rather than features.

Speech vocabulary follows the same pattern as text and image vocabulary when applied to audio workloads. Automatic speech recognition transcribes audio into text, with word error rate as the standard metric. Text to speech synthesizes audio from text and uses mean opinion score for quality. A spectrogram is the time-frequency representation models actually consume, since raw waveforms are too dense. Speaker diarization labels which voice spoke during each segment of a multi-party recording. Voice cloning trains a model to imitate a specific speaker, which is the technical capability behind both accessibility tools and deepfake risk. Wake words activate a device, and barge-in lets a user interrupt mid-response, two terms that show up in any voice product specification.

Computer Vision and Multimodal AI Vocabulary

Turning to vision, the core terms describe how pixels become labels. Image classification assigns a single label to a whole image, while object detection draws boxes around every object and labels each one. Semantic segmentation paints every pixel with a class label across the entire image surface. Instance segmentation goes further by separating individual objects of the same class. Pose estimation locates joints on people or hands so downstream systems can read movement. Optical character recognition turns text in pixels back into text. A bounding box is the rectangle a detector outputs, and intersection over union is the metric that scores how well predicted boxes match the truth. Our introduction to computer vision covers each of these tasks with implementation examples.

Multimodal AI unifies vision, text, and audio in a single model. A vision-language model uses one backbone to align images and text, which lets users ask questions about a picture and get a sentence back. A multimodal embedding lands images, captions, and audio clips in the same vector space, so search and retrieval can cross media. Diffusion models generate images by reversing a noise process, while autoregressive image models predict patches in sequence. Image-to-video, video-to-video, and text-to-3D extend that pattern to richer outputs. CLIP, SAM, and similar open models gave teams the building blocks to ship multimodal features without training from scratch. That access is why most 2026 product roadmaps include at least one vision-language workflow.

Generative AI Terms Reshaping Content and Code

Building on the multimodal vocabulary, generative AI names the class of models that produce new content rather than just classifying existing input. A large language model, or LLM, is a transformer trained on broad text data to generate language, write code, and reason step by step. A diffusion model generates images, video, or audio by learning to reverse noise. A generative adversarial network, or GAN, pits a generator against a discriminator and produces images by chasing the discriminator’s threshold. Our introduction to generative adversarial networks covers the architecture in detail for builders. A variational autoencoder learns a compressed latent that supports controlled generation. Each family has its own strengths, and 2026 systems often chain them together inside one product.

Several prompt and control terms shape the output that generative models actually produce in deployed applications. A prompt is the input text given to a generative model, and prompt engineering is the craft of writing prompts that yield reliable outputs. A system prompt sets the persistent role, instructions, and guardrails. Few-shot prompting includes example pairs in the prompt itself, while zero-shot relies on the instruction alone. Chain of thought asks the model to reason step by step in its visible output. Temperature controls randomness: low values produce focused output, high values produce variety. Top-p and top-k sampling cap the candidate token set, and a stop sequence ends generation at a specific marker.

Fine-tuning and adaptation vocabulary closes the generative cluster with the steps teams use to specialize a model. Pretraining is the heavy step that learns general language patterns from web-scale text. Fine-tuning continues training on a smaller domain-specific set so the model picks up your style or knowledge. Parameter-efficient fine-tuning, including LoRA and QLoRA, updates only a small fraction of weights so the cost stays low. Our hands-on guide to fine-tuning LLMs at home walks through the setup. Reinforcement learning from human feedback aligns the model to preferences by training on ranked outputs. Direct preference optimization is a newer recipe that skips the reward model and shortens the pipeline.

Agentic AI Vocabulary You Need for 2026

Stepping past static prompts, agentic AI describes systems that plan, act, observe, and adjust without waiting for a single human turn. An AI agent uses a language model as its reasoning core, calls tools to take actions in the world, reads the results, and loops until a goal is met. A tool, often called a function, is any callable code surface the agent can invoke: a calculator, a search API, a database query, a code executor, or a browser. Tool use is the act of choosing a tool, building arguments, and parsing the response. The agent loop is the gather-context, act, verify pattern that repeats until a stop condition fires.

Several planning vocabulary items matter inside any working agent stack that ships in 2026 production environments. A plan is the agent’s decomposition of a goal into ordered steps. ReAct, short for reason and act, interleaves reasoning steps with tool calls in one stream. A reflection step asks the model to critique its own output and try again. Memory is any state the agent keeps between turns: short-term scratchpads, long-term vector stores, or shared notebooks across agents. A scratchpad is the working transcript the model writes during a single task. Stateful agents persist memory across sessions, while stateless agents start clean every run.

Multi-agent systems extend the single agent pattern with specialist roles, supervisors, and shared memory layers. An orchestrator routes work to specialist agents and tracks completion across the full task graph. A planner builds the top-level plan and hands subtasks to executors. Critics review intermediate outputs before they ship to a downstream user or to another agent. Agent-to-agent protocols, including Anthropic’s Model Context Protocol and OpenAI’s tool-use schemas, let agents discover and call each other across vendors. Handoff is the act of transferring control from one agent to another with the relevant context attached. Our deep dive on Evaluating Amazon Bedrock agents with Ragas covers the metrics teams now use to grade these multi-agent systems before they reach production.

Production agent vocabulary closes the cluster with the runtime controls every 2026 deployment needs in place. A guardrail is a runtime check that blocks unsafe actions before they fire, such as a refund cap or a domain allowlist. A cost ceiling stops a long-running agent before token spend gets out of hand. Tracing captures every step of an agent run so engineers can replay and debug. Eval suites are recorded scenarios used to grade agent behavior at every release. Human-in-the-loop describes any step that pauses for a human decision, from approval clicks in workflow tools to escalation queues in support. These five words define the difference between a demo agent and an agent that touches a real customer account.

Retrieval, Memory, and Context Window Terminology

Looking at the data layer around the model, retrieval vocabulary describes how the right information gets into the prompt at the right time. Retrieval-augmented generation, or RAG, fetches relevant chunks of company data and inserts them into the prompt so the model can ground its answer in private knowledge. A chunk is a small segment of a source document, sized to fit alongside the question. Chunking strategy decides where to cut: fixed window, sentence boundary, semantic split, or layout-aware. A retriever is the component that ranks chunks by relevance, with dense retrievers using embeddings and sparse retrievers using keyword scores like BM25. Hybrid retrievers combine dense and sparse signals to deliver better recall than either approach alone.

The storage and search layer around retrieval brings its own words to learn for any architecture review. A vector database stores embeddings and runs approximate nearest neighbor search to return the closest matches. Common vector stores include Pinecone, Weaviate, Qdrant, Chroma, and Postgres with pgvector. A metadata filter narrows the search by document attributes such as date, author, or department. A rerank step uses a slower, more accurate model to reorder the top retrieved chunks before they reach the LLM. Embedding models like text-embedding-3-large, voyage-3, and nomic-embed turn text into the vectors those databases search. Indexing is the upfront step of running every document through an embedding model and storing the results.

Context vocabulary finishes the retrieval cluster with the words that describe how the model reads its prompt. A context window is the number of tokens a model can read at once, with 2026 frontier models offering one million tokens or more. Context engineering is the discipline of choosing what goes into that window: instructions, examples, retrieved chunks, tool definitions, and prior conversation. A KV cache stores the model’s intermediate computations so repeated context does not need to be processed twice, which is what makes long contexts affordable. Long-context evaluations, such as needle-in-a-haystack and RULER, measure how well a model uses its full window. Agentic RAG combines retrieval with autonomous decision making, letting the agent decide what to search for and when to stop, and it became the dominant 2026 enterprise pattern.

AI Safety, Alignment, and Risk Vocabulary

Shifting to the safety cluster, hallucination names the most discussed failure mode. A hallucination is an output the model presents as fact that has no basis in training data or retrieved sources. Modern frontier models still hallucinate on hard questions even after major 2026 progress. Grounding ties model output to real sources, often through citations to retrieved chunks. Calibration is how well a model’s expressed confidence tracks its real accuracy. Prompt injection is an attack where untrusted text in a tool result or document steers the model away from its system instructions. Indirect prompt injection hides those instructions inside web pages, emails, or attached files. Jailbreaks are prompts that talk the model past its safety policy.

Alignment vocabulary covers the work of pointing the model the right way. Alignment is the broad goal of making model behavior match human values and instructions. Reinforcement learning from human feedback, RLHF, trains a reward model on ranked outputs and then optimizes the policy against it. Constitutional AI from Anthropic uses written principles to grade outputs and reduce reliance on human raters. Red teaming is the structured practice of trying to break a model before release, and it now produces test suites that ship with each major version. Sandbagging describes a model that hides capability during evaluation, which is a documented risk for frontier systems and a 2026 research focus.

Bias, Fairness, Ethics, and Responsible AI Terminology

Building on the safety cluster, the bias and fairness vocabulary covers harms that show up in real deployments. Bias in AI describes systematic errors that produce different outcomes for different groups, and it can enter through data, labels, model design, or deployment context. Disparate impact measures whether an automated decision affects protected groups at different rates. Disparate treatment is the legal term for treating people differently based on a protected attribute. Equal opportunity, demographic parity, and equalized odds are three competing fairness definitions used in audits, and choosing among them requires real product judgment. Our explainer on the dangers of AI bias and discrimination walks through high-profile failure cases.

Several mitigation terms matter for product teams that ship customer-facing models inside regulated industries. Debiasing covers any technique that reduces measured disparity, from data reweighting to post-processing predictions. A representative dataset includes the populations and edge cases the model will serve in production. Annotation guidelines tell labelers how to make consistent judgments, and disagreements between labelers often reveal real ambiguity rather than human error. Fairness audits assess models on slices of users, often disaggregated by demographic group. Human review queues catch high-stakes decisions before they reach the user. Bias bounty programs invite external researchers to find harms a product team missed.

Trust vocabulary closes the responsible AI cluster with the words leaders use in board and regulator meetings. Explainability is the discipline of making model decisions inspectable, often through feature attribution, counterfactuals, or saliency maps. Explainable AI methods explained covers the methods most often required in regulated industries. Interpretability is the deeper research goal of understanding what individual circuits in a model actually compute, with mechanistic interpretability now an active 2026 research field. Transparency is the broader practice of publishing model cards, system cards, evaluation results, and incident reports. Accountability assigns responsibility for outcomes, and it depends on contracts, governance, and clear audit trails as much as on any technical method.

AI Governance, Policy, and Compliance Vocabulary

Shifting from individual model harms to organizational controls, governance vocabulary describes the rules of the road in 2026. AI governance is the set of policies, controls, and accountability structures organizations use to develop, deploy, and oversee AI responsibly across the entire lifecycle. A model card documents a model’s training data, intended use, evaluation results, and known limitations, and the Hugging Face standard for model cards is now widely adopted. A system card extends the model card idea to whole products and to multi-model pipelines. An AI inventory lists every AI system in use, often required by internal audit and emerging regulations. Tiered risk classifications, used by the EU AI Act and the NIST AI Risk Management Framework, decide which controls apply at which stakes. Our responsible AI governance frameworks guide unpacks the implementation details.

Several regulation and standard names now dominate 2026 procurement contracts across regulated industries worldwide. The EU AI Act creates four risk tiers from unacceptable to minimal, with general-purpose AI obligations layered on top. The NIST AI RMF is a voluntary US framework structured around govern, map, measure, and manage functions. ISO/IEC 42001 is the international management system standard for AI. Sovereign AI describes the push to run models on national infrastructure with local data residency, and it is reshaping cloud procurement across the EU, Gulf states, and parts of Asia. Data residency, model provenance, and training data disclosure are the three contractual hooks that buyers ask about first. Knowing these terms now decides whether a procurement conversation closes in a week or stalls for a quarter.

Implementation: How These AI Terms Map to Real Deployments

Looking at how the vocabulary lands in real teams, implementation maps each cluster to a role and a system component. A typical 2026 AI deployment has a model layer, a retrieval layer, an agent layer, an evaluation layer, and a governance layer. Each layer pulls vocabulary from the matching glossary cluster every day. The model layer holds the LLM, embedding model, and any classifiers, and the team that owns it uses foundations, deep learning, and generative vocabulary every day. The retrieval layer holds the vector store, chunking pipeline, and rerank step. The agent layer holds tools, planners, and memory and pulls its vocabulary from the agentic cluster. The evaluation layer holds prompts, eval suites, and dashboards that grade behavior at every release candidate.

Procurement workflows also draw on the vocabulary cluster by cluster as buyers evaluate AI vendors in 2026. A request for proposal now asks for SOC 2 reports, model cards, EU AI Act risk classifications, and data residency commitments. Pricing conversations cover input and output tokens, cached tokens, and per-request fees. Latency requirements get written as time-to-first-token and tokens-per-second targets in vendor responses. Vendor evaluations test model output against eval suites that mirror real customer workflows. A reference architecture diagram now routinely shows the retrieval layer, the agent loop, and the guardrail stack alongside classic application boxes. Our coverage of reinforcement learning with human feedback explains the training-side vocabulary procurement teams now hear during alignment reviews.

Build versus buy decisions hinge on the same vocabulary cluster a procurement lead carries to vendor briefings. Teams building from foundation models budget for compute, inference latency, and fine-tuning data. Teams buying packaged products check for tool integration, agent transparency, and audit log access. Open-source paths require operational fluency with quantization, LoRA, and serving frameworks. Hosted paths require fluency with prompt caching, batch inference, and rate-limit handling. Pilot success is now usually measured against three metrics: task completion rate from eval suites, cost per successful task, and intervention rate where a human had to take over. Each of those metrics is a vocabulary chain that runs from product to engineering to finance.

People and process vocabulary closes the implementation picture with the roles every 2026 AI program needs. An AI product manager pairs with an applied scientist or ML engineer to ship features. A governance lead writes policy and runs the internal risk committee. A red team probes the system before each release and files findings to a tracked queue. A trust and safety team triages incidents reported by users. Cross-functional rituals such as model release reviews and post-incident retrospectives keep the work coordinated. Naive Bayes classifiers still show up in older pipelines, and our naive Bayes classifiers primer is a useful reminder that not every production model is a transformer. The vocabulary above is the shared language those people use to move work from idea to production.

Future Outlook for AI Terminology Beyond 2026

Looking ahead beyond 2026, several emerging terms are already shaping next year’s procurement conversations. World models describe systems that learn an internal simulation of the environment and use it for planning, with robotics and self-driving research teams leading adoption. Reasoning models are LLMs trained to spend additional inference compute on multi-step problem solving, with OpenAI o-series, Anthropic Claude, and DeepSeek models pushing benchmarks across math, coding, and scientific reasoning. Continual learning lets a deployed model absorb new examples without full retraining. Mechanistic interpretability tries to read individual circuits inside a model, and progress here is the strongest evidence that opaque systems can be understood at the gear level.

Several agentic vocabulary terms are about to enter common use. Computer use agents, also called CUAs, control desktop and browser interfaces the way a human would, screenshot by screenshot and click by click. Long-horizon agents are designed to run for hours or days against open-ended goals, with checkpointing and resumption baked in. Agent marketplaces let teams discover, evaluate, and rent specialist agents the way they buy SaaS today. Multi-agent simulation environments grade agents in sandboxed economies before any production exposure. Together these terms are the next layer of the glossary of AI terms enterprises will need by mid-2027.

Policy vocabulary keeps evolving alongside the technical stack as regulators publish new guidance every quarter. Sovereign AI clusters, AI bills of materials, and content provenance standards such as C2PA are entering procurement contracts. AI assurance, the third-party audit category that mirrors financial audit, is forming around frameworks like ISO/IEC 42001 and the NIST AI RMF. Synthetic media labeling rules are spreading across regulators, and watermarking research is racing to make those rules technically enforceable. The glossary of AI terms that defines 2027 will read like 2026’s plus a richer agent vocabulary, a deeper safety vocabulary, and a longer list of compliance acronyms. Staying ahead means treating the glossary as a living document and updating it every quarter.