Glossary

A

AI Token

Also known as: LLM token, language model token, input token, output token, subword token

An AI token is the small unit of text, often a subword, that a language model reads, generates, counts against its context window, and bills for.

AI Inference

Also known as: inference, model inference, machine learning inference, ML inference, AI inferencing, inference serving

AI inference is the execution phase where a trained AI model applies what it learned to new, unseen data and produces an output such as a prediction.

AI Red Teaming

Also known as: AI red team, AI red-teaming, LLM red teaming, GenAI red teaming, adversarial testing, AI adversarial testing

AI red teaming is a structured testing effort that uses adversarial methods to find flaws, vulnerabilities, and misuse risks in a deployed AI system.

AI Agent

Also known as: AI agents, agentic AI, autonomous agent, LLM agent, intelligent agent, software agent

An AI agent is a software system that uses an AI model to plan, pick tools and take actions toward a goal on a user’s behalf, with limited human oversight.

AI Hallucination

Also known as: AI hallucinations, LLM hallucination, AI confabulation, model hallucination

An AI hallucination is output a generative model states with confidence but that is factually wrong, unsupported, or contradicts its own prompt.

AI Jailbreak

Also known as: AI jailbreaking, LLM jailbreak, LLM jailbreaking, jailbreaking AI, jailbreak prompt, model jailbreak, GenAI jailbreak

An AI jailbreak is a direct prompting attack intended to circumvent the restrictions placed on a model’s outputs, such as its trained refusal behavior.

Automated Decision-Making

Also known as: ADM, automated decision-making, automated individual decision-making, solely automated decision

A significant decision about a person taken with no meaningful human involvement, defined by UK GDPR Articles 22A to 22C and subject to statutory safeguards.

C

Context Window

Also known as: context windows, context length, context size, context window size, token limit

A context window is all the text an AI model can reference when generating a response, measured in tokens and shared with the model’s own output.

D

Data Poisoning

Also known as: training data poisoning, data poisoning attack, dataset poisoning, poisoned training data, AI data poisoning

Data poisoning is a poisoning attack in which an adversary controls part of the training data an AI model learns from, altering its behavior.

F

Frontier Model

Also known as: frontier AI, frontier AI model, frontier models, frontier AI system, frontier system, frontier LLM

A frontier model is a highly capable general-purpose AI model that matches or exceeds today’s most advanced systems, and triggers safety obligations.

G

General-Purpose AI

Also known as: GPAI, general-purpose AI model, general purpose AI, GPAI model, general-purpose AI models

General-purpose AI (GPAI) is an EU AI Act term for a model with significant generality that can competently perform a wide range of distinct tasks.

L

LLM Parameters

Also known as: model parameters, LLM parameter count, parameter count, model size, LLM model size, billion parameter model

LLM parameters are the weights and biases a model learns during training. The count sets a model’s size, but training data and sparsity matter just as much.

M

Model Weights

Also known as: weights, AI weights, open weights, open-weight model, model parameters file, checkpoint, trained model weights

Model weights are the learned parameters a trained AI model uses to turn an input into an output, and the artifact a lab chooses to publish or withhold.

Model Card

Also known as: model cards, ML model card, AI model card, model reporting card

A model card is a short document released alongside a trained machine learning model that reports its intended uses, evaluation results and limitations.

Mixture of Experts

Also known as: MoE, sparse mixture of experts, MoE model, sparsely-gated mixture of experts, sparse model, mixture-of-experts

A mixture of experts (MoE) model splits its feed-forward layers into expert sub-networks and a router sends each token to only a few of them.

Multimodal AI

Also known as: multimodal model, multimodal models, multimodal machine learning, multimodal LLM, multimodal large language model, MLLM, large multimodal model

Multimodal AI is a single model that takes in and relates more than one type of input, such as text, images, audio or video, rather than only text.

P

Prompt Injection

Also known as: prompt injection attack, LLM prompt injection, indirect prompt injection, LLM01

Prompt injection is an attack in which crafted input makes a large language model follow the attacker’s instructions instead of its operator’s.

R

Responsible Scaling Policy

Also known as: RSP, responsible scaling policies, responsible scaling commitment

A responsible scaling policy is a voluntary commitment an AI developer publishes in advance, naming the capabilities that trigger stronger safeguards.