Note: many more terms found in the literature can be found in the NIST Glossary.

Accountability

Definition: 1) relates to an allocated responsibility. The responsibility can be based on regulation or agreement or through assignment as part of delegation; 2) For systems, a property that ensures that actions of an entity can be traced uniquely to the entity; 3) In a governance context, the obligation of an individual or organization to account for its activities, for completion of a deliverable or task, accept the responsibility for those activities, deliverables or tasks, and to disclose the results in a transparent manner.

Source: ISO/IEC TS 5723:2022(en) Trustworthiness — Vocabulary

Agentic AI

Definition: AI systems that autonomously perform multiple sequential steps – sometimes including actions likebrowsing the internet, sending emails, or sending instructions to physical equipment – to try and complete a high-level task or goal. 

Source: UK Government Frontier AI Discussion Paper

Application Layer

Definition: The application layer refers to the software applications and interfaces built on top of foundation models and infrastructure. This includes tools for content generation, chatbots, code assistants, and other user-facing AI applications that utilise underlying models and systems.

Source: Reframe Venture (inference from research)

Artificial general intelligence, or AGI

Definition: Artificial general intelligence (AGI) is a hypothetical type of AI that mimics human-like intelligence. Unlike regular AI, which is designed for specific tasks (such as playing chess, grammar correction, or speech translation), AGI is characterized by its general cognitive abilities. Which means it can perform any intellectual task that a human can do, adapt to new situations, and improve its performance over time.

Source: Syracuse University

Automation v Augmentation

Definition: Automation refers to AI systems that replace human tasks entirely, while augmentation refers to AI systems that enhance and support human capabilities rather than replacing them.

Source: Stanford Digital Economy Lab

Bias

Definition: Output errors caused by skewed training data. Such bias can cause models to produce inaccurate, offensive, or misleading predictions. Biased AI models arise when algorithms prioritize irrelevant or misleading data traits over meaningful patterns

Source: MIT

Black-box v White-box

Definition: A black box AI is an AI system whose internal workings are a mystery to its users. Users can see the system’s inputs and outputs, but they can’t see what happens within the AI tool to produce those outputs. White box AI, also called explainable AI (XAI) or glass box AI, is the opposite of black box AI. It is an AI system with transparent inner workings. Users understand how the AI takes in data, processes it and arrives at a conclusion. 

Source: IBM

Context Window

Definition: The context window is the maximum number of tokens (words or parts of words) that an AI model can process and consider simultaneously when generating a response. It is essentially the “memory” capacity of the model during an interaction or task. Models with larger context windows can handle larger attachments/prompts/inputs and sustain “memory” of a conversation for longer (Fogarty, 2023).

Source: MIT

Data Center

Definition: An AI data center is a facility that houses the specific IT infrastructure needed to train, deploy and deliver AI applications and services. AI data centers require high-performance graphics processing units (GPUs) and their IT infrastructure considerations, such as advanced storage, networking, energy and cooling capabilities.

Source: IBM

Explainability

Definition: The ability to provide a human interpretable explanation for a machine learning prediction and produce insights about the causes of decisions, potentially to line up with human reasoning.

Source: A Taxonomy and Terminology of Adversarial Machine Learning

Failure Modes

Definition: Failure modes refer to the various ways AI systems can malfunction, produce incorrect outputs, or cause harm. These include hallucinations (generating false information), bias amplification, security vulnerabilities, and unexpected behaviors that emerge from the statistical patterns encoded in the models.

Source: Microsoft

Frontier Model

Definition: A frontier model is a highly advanced, large-scale AI model that pushes the boundaries of AI in areas like NLP, image generation, video and coding. Frontier models are typically trained on extensive datasets with billions or even trillions of parameters. They have multimodal capabilities, meaning they can process and generate text, images, audio and video. They can also display abilities that were not explicitly programmed, such as reasoning, code generation, or creative writing.

Source: Iguazio

GDPR (General Data Protection Regulation)

Definition: GDPR (General Data Protection Regulation) is European Union legislation that governs data privacy and protection, requiring organisations to handle personal data responsibly and transparently.

Source: GDPR.eu

Generative AI (GenAI)

Definition: Generative AI, sometimes called gen AI, is AI that can create original content such as text, images, video, audio or software code in response to a user’s prompt or request.

Source: IBM

GPU (Graphics Processing Unit)

Definition: A graphics processing unit (GPU) is a specialised electronic circuit designed for digital image processing and to accelerate computer graphics. GPUs are designed for processing multiple tasks at the same time, having a large number of processing cores that can work on different parts of a task simultaneously. GPUs can accelerate the training and inference processes, allowing AI models to be developed and deployed more quickly and efficiently than using CPUs.

Source: IBM

GPT (Generative Pre-trained Transformer)

Definition: A generative pre-trained transformer (GPT) is a type of large language model (LLM) that is widely used in generative AI chatbots, based on a deep learning architecture called the transformer and pre-trained on large datasets of unlabeled content. Generative pre-trained transformers (GPTs) are a family of advanced neural networks designed for natural language processing (NLP) tasks, subjected to unsupervised pre-training on massive unlabeled datasets.

Source: IBM

Hallucination

Definition: Refers to outputs that deviate significantly from real data. These can be nonsensical creations, factual errors, or content biased by the training data. Hallucinations arise from limitations in the model's understanding of the underlying data distribution.

Source: Syracuse University

Interpretability

Definition: The ability to understand the value and accuracy of system output. Interpretability refers to the extent to which a cause and effect can be observed within a system or to which what is going to happen given a change in input or algorithmic parameters can be predicted.

Source: National Security Commission on Artificial Intelligence: The Final Report

Large Language Model (LLM)

Definition: A complex AI system trained on massive amounts of text data. These models leverage deep learning architectures like transformers to analyse and process information, enabling them to perform diverse tasks in natural language processing.

Source: Syracuse University

Machine Learning

Definition: A subfield of AI where algorithms improve their performance on a specific task through experience. They learn from data, identifying patterns and relationships without explicit programming. This allows them to make predictions, classifications, or decisions on new data, constantly refining their abilities.

Source: Syracuse University

Model Cards

Definition: Short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups (e.g., race, geographic location, sex, Fitzpatrick skin type) and intersectional groups (e.g., age and race, or sex and Fitzpatrick skin type) that are relevant to the intended application domains. [They] also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information.

Source: Model Cards for Model Reporting

Multimodal AI

Definition: Refers to AI systems that process and learn from multiple data types, like text, images, audio, and sensor data. It employs data fusion techniques to combine information from these modalities, leading to a richer understanding of the data compared to single-modality approaches. This enables multimodal AI to perform tasks like image captioning, video question answering, and robot perception in the real world.

Source: Syracuse University

Neural network

Definition: Computational models inspired by the structure and function of the brain. They consist of interconnected nodes (artificial neurons) arranged in layers. These nodes process information and transmit signals to other nodes, mimicking how neurons fire in the brain. By adjusting the connections between nodes (learning), neural networks can perform complex tasks like image recognition, speech understanding, and natural language processing.

Source: Syracuse University

Open-source v Closed-source

Definition: Open-source AI models are publicly and freely accessible systems that developers can use for various applications and purposes. Closed-source models are proprietary systems that keep their code confidential. By restricting access to underlying code, closed-source AI model developers ensure more control over the system.

Source: Mulitmodal

Parameters

Definition: Adjustable elements within a model, like weights and biases in neural networks. These values determine how the model transforms input data into generated outputs. Tweaking these parameters allows fine-tuning the model's behavior, influencing the style, creativity, and accuracy of the generated content.

Source: Syracuse University

Predictive AI

Definition: Predictive AI blends statistical analysis with machine learning algorithms to find data patterns and forecast future outcomes, extracting insights from historical data to make accurate predictions about the most likely upcoming event, result or trend. Predictive AI makes predictions, recommendations and decisions using various AI and machine learning (ML) techniques.

Source: IBM

Red-team

Definition: A group of people authorized and organized to emulate a potential adversary’s attack or exploitation capabilities against an enterprise’s security posture. The Red Team’s objective is to improve enterprise cybersecurity by demonstrating the impacts of successful attacks and by demonstrating what works for the defenders (i.e., the Blue Team) in an operational environment. Also known as Cyber Red Team.

Source: Information Technology Laboratory Computer Security Resource Center Glossary

Reliability

Definition: Reliability refers to the closeness of the initial estimated value(s) to the subsequent estimated values.

Source: Glossary of Statistical Terms

Resilience

Definition: The ability to prepare for and adapt to changing conditions and withstand and recover rapidly from disruptions. Resilience includes the ability to withstand and recover from deliberate attacks, accidents, or naturally occurring threats or incidents. The ability of a system to adapt to and recover from adverse conditions.

Source: A Taxonomy and Terminology of Adversarial Machine Learning

Safety

Definition: Property of a system such that it does not, under defined conditions, lead to a state in which human life, health, property, or the environment is endangered; [safety involves reducing both the probability of expected harms and the possibility of unexpected harms].

Source: ISO/IEC TS 5723:2022(en) Trustworthiness — Vocabulary

Small language models (SLMs)

Definition: Small language models (SLMs) have millions to low billions of parameters, substantially less than large language models. Engineers use less text data for training and design SLMs with fewer transformer layers, narrower hidden dimensions, and fewer attention heads, resulting in simpler mathematical transformations and faster, more efficient inference suitable for focused tasks.

Source: IBM

Transparency

Definition: Open, comprehensive, accessible, clear and understandable presentation of information; property of a system or process to imply openness and accountability

Source: ISO/IEC TS 5723:2022(en) Trustworthiness — Vocabulary

Trust

Definition: The system status in the mind of human beings based on their perception of and experience with the system; concerns the attitude that a person or technology will help achieve specific goals in a situation characterized by uncertainty and vulnerability.

Source: Technology Investment Strategy 2015-2018