AI glossary

Rules describing permitted and prohibited uses of an AI product or model.

A nonlinearity applied in neural networks so stacked layers can model complex functions.

Stored state that agents use across steps or sessions, from scratchpads to vector stores.

Artificial general intelligence: hypothetical AI with broad, flexible competence across most cognitive tasks.

A system that pursues goals by choosing actions over multiple steps, often using tools, memory, and feedback from an environment.

Using AI to execute workflow steps that previously required manual effort.

Policies, roles, and controls for managing AI risk and value across an organization.

A confident fabricated or unsupported statement produced by a generative model.

NIST’s Artificial Intelligence Risk Management Framework for governing AI risks.

Making AI systems behave according to intended goals, preferences, and constraints.

Computer systems designed to perform tasks that typically require human cognitive abilities, such as perception, language use, planning, or decision support.

Automatic speech recognition; same family as speech-to-text.

A neural network component that weights which parts of an input sequence matter most for the current computation.

Generating sequences token-by-token, each conditioned on previous tokens.

The algorithm for efficiently computing gradients in neural networks via the chain rule.

A structured comparison of models or vendors on the same evaluation tasks.

A decoding algorithm that keeps multiple candidate prefixes instead of only greedy choices.

Systematic error or unfair differential performance that can harm people or distort decisions.

When fine-tuning erases useful behaviors the base model previously had.

Prompting or training patterns that encourage step-by-step intermediate problem solving.

Splitting documents into smaller passages before embedding and retrieval in RAG systems.

Linking generated claims to source passages in grounded Q&A systems.

When a model invents plausible-looking references that do not exist.

A supervised task that assigns inputs to discrete categories.

Unsupervised grouping of items by similarity.

Convolutional neural network: an architecture using local filters for grid-like data such as images.

AI techniques that extract meaningful information from images and video, such as objects, text, depth, or motion.

An Anthropic-popularized approach using written principles and AI feedback to critique and revise outputs.

Packing excessive text into a prompt, often harming relevance and raising cost.

The maximum number of tokens a model can consider in one request, including instructions, history, and retrieved text.

Rotating held-out data folds to estimate generalization more robustly.

Improperly sharing information between training and evaluation that inflates metrics.

Machine learning with multi-layer neural networks that learn hierarchical representations from large datasets.

Synthetic media that convincingly depicts people saying or doing things they did not.

Decoding settings that make outputs repeatable for the same inputs, such as greedy decoding.

A generative model that learns to reverse a gradual noising process to synthesize data such as images.

Training a smaller model to mimic a larger model’s behaviors or probability outputs.

A numeric vector representation of an item—such as a token, sentence, image, or document—used to capture similarity and feed neural models.

Degradation of retrieval quality when embedding models or data distributions change.

A model specialized for producing vectors used in similarity search.

A list of numbers representing an item so similar items land nearby in space.

The searchable structure storing vectors for nearest neighbor retrieval.

European Union legislation establishing a risk-based regulatory framework for AI systems.

An automated suite that scores model or prompt versions on fixed tasks.

A historic symbolic AI application encoding specialist rules for decision support.

Techniques and product designs that help humans understand why an AI system produced an output.

Transforming raw data into informative inputs for machine learning models.

Providing a handful of input–output examples in the prompt to steer model behavior.

Curated examples used to adapt a pretrained model to a task or style.

Additional training that adapts a pretrained model to a narrower task, domain, or preferred behavior using specialized data.

A large pretrained model adaptable to many downstream tasks via prompting, tools, or fine-tuning.

Structured tool invocation from an LLM.

A model’s ability to perform well on new data from the same distribution as training.

AI systems that create new content—such as text, images, audio, or code—by modeling patterns from training data.

A common file format for quantized local LLM weights.

An optimization method that updates parameters by stepping opposite the loss gradient.

Accepted correct labels or answers used for evaluation.

Connecting model outputs to verifiable evidence such as retrieved documents or tool results.

Technical and policy controls that constrain AI inputs and outputs to reduce harmful or out-of-policy behavior.

When a generative model produces fluent content that is fabricated, unsupported, or inconsistent with provided sources.

Human-in-the-loop oversight patterns.

Keeping people involved to review, label, or approve AI system steps.

A configuration choice set before learning parameters, such as learning rate or number of layers.

Generative transformation of an input image guided by a prompt or strength parameter.

The process of running a trained model on new inputs to produce predictions or generations, usually without updating model weights.

Editing a region of an image by generating new pixels constrained by surrounding context and a prompt.

Fine-tuning language models on examples of instructions paired with desired responses.

An international standard for artificial intelligence management systems.

A prompt attack that attempts to bypass an AI system’s safety policies.

The point in time after which a model’s frozen training data no longer covers world events.

Cached key/value attention states that speed up autoregressive decoding.

A large neural network, typically a transformer, trained on massive text corpora to predict tokens and perform language tasks.

The maximum acceptable response time for an AI feature in a product.

An internal representation space where models encode compressed features of inputs.

Abbreviation for large language model.

Running models on devices or private infrastructure you control rather than a third-party API.

Raw unnormalized scores a model outputs before converting to probabilities.

Low-Rank Adaptation: a parameter-efficient fine-tuning method that trains small adapter matrices instead of all model weights.

A numeric measure of prediction error that training aims to minimize.

A family of methods that improve model performance by learning patterns from data instead of relying only on hand-coded rules.

An architecture that routes each input to a subset of specialized expert networks.

Documentation describing a model’s intended use, limitations, and evaluation highlights.

Running a trained model on new inputs to produce predictions or generations.

Directing each request to an appropriate model based on complexity, cost, or policy.

Systems where multiple AI agents with roles collaborate on a workflow.

AI systems that process or generate more than one data type—such as text, images, audio, or video—within a unified model or pipeline.

A model that accepts or generates more than one data type, such as text and images.

AI systems designed for specific tasks rather than general human-like intelligence.

The AI subfield focused on analyzing, understanding, and generating human language for tasks like translation, search, and assistants.

In image generation, text describing what to avoid in the output.

A model made of layered interconnected units that transform inputs through weighted connections and nonlinear activations.

Optical character recognition: extracting text from images of documents.

Model parameters released for download, often under licenses that still restrict some uses.

When a model fits training data quirks and fails to generalize to new data.

A learned numeric value in a model, such as a neural network weight.

Parameter-efficient fine-tuning methods such as LoRA.

Personally identifiable information that can identify an individual.

In RL, the component that selects actions; in RLHF, the language model being optimized.

A model trained to score which responses humans prefer, used in RLHF-like pipelines.

Large-scale training on broad data that builds general capabilities before task specialization.

The input text or structured instructions given to a generative model to steer its output toward a desired task or style.

An attack where untrusted text manipulates an LLM into ignoring developer instructions or misusing tools.

A reusable prompt structure with variables filled per request.

Metadata and signals describing how content was created or modified, including AI involvement.

Reducing numeric precision of model weights (and sometimes activations) to save memory and speed up inference.

Retrieval-augmented generation: retrieving evidence to ground model answers.

A pattern that retrieves relevant documents and conditions a generative model on that evidence to improve grounded answers.

Adversarial testing that probes AI systems for unsafe or unintended behaviors.

Supervised learning that predicts continuous numeric targets.

Learning to choose actions in an environment to maximize cumulative reward through interaction rather than fixed labels.

A model that reorders retrieved candidates for higher precision before generation.

Reinforcement Learning from Human Feedback: aligning models using preference data and a reward model.

A structured argument with evidence that a system is acceptably safe for a given context.

Automated checks that block or modify disallowed inputs or outputs.

Creating training signals from the data itself, such as predicting masked tokens.

Retrieving documents by meaning using embeddings rather than only exact keywords.

A comparatively compact LM designed for lower cost, edge deployment, or specialized tasks.

A function that converts logits into a probability distribution over classes or tokens.

Labeling who spoke when in a multi-speaker audio recording.

Converting spoken audio into text.

Sending model output tokens to clients incrementally as they are generated.

Constraining model responses to schemas such as JSON for reliable downstream parsing.

Learning from labeled input–output examples to predict labels for new inputs.

AI approaches that manipulate explicit symbols and rules, such as logic systems and knowledge graphs.

Artificially generated data used to train or evaluate models when real data are scarce or sensitive.

Developer- or product-level instructions that set role, style, and policy before user messages.

A decoding parameter that controls how random or deterministic next-token sampling is.

Producing natural language outputs with a generative model.

A chunk of text (often a word piece) used as a basic unit for language model training, context limits, and generation.

Billing model inference by the number of tokens processed in prompts and completions.

Software that splits text into tokens and maps them to vocabulary IDs for a model.

A throughput metric for how quickly a system generates tokens during inference.

Letting an LLM request structured actions through developer-defined tools or functions.

Nucleus sampling that limits token draws to a probability mass cutoff.

The examples used to fit a model’s parameters, including inputs and any labels or supervisory signals required by the learning objective.

A neural network architecture that uses attention mechanisms to model relationships among tokens in a sequence, foundational to modern LLMs.

A neural architecture built around attention mechanisms, foundational to modern LLMs.

Text-to-speech synthesis of audio from text.

Finding structure in unlabeled data, such as clusters or compressed representations.

A database optimized to store embeddings and perform approximate nearest neighbor similarity search.

A multimodal model combining vision and language.

Synthesizing speech that resembles a specific person’s voice from reference audio.

Signals embedded in generated media or text to help indicate synthetic origin.

Synonym for learned model parameters.

A common speech recognition metric based on insertions, deletions, and substitutions.

Performing a task from instructions alone without in-prompt demonstrations.