AI Terminology

A

Adversarial AI

Techniques and attacks used to manipulate AI systems, causing them to make incorrect or unintended predictions or decisions. These techniques exploit vulnerabilities in AI models, often by subtly altering input data, training data, or model interactions to manipulate the AI system.

Agentic AI

A category of AI systems capable of independently making decisions, interacting with their environment, and optimizing processes without direct human intervention.

Agentic workflow

A structured sequence of tasks assigned to one or more AI agents that can plan, execute, and iterate toward a goal with minimal human involvement. Agentic workflows often combine tool use, memory, and multi-step reasoning to complete complex objectives autonomously.

AI Agent

A system that autonomously perceives its environment, decides what to do, and takes actions to achieve its goals.

AI drift/decay

The tendency for an AI model's performance to degrade over time when deployed in a real-world setting with differing conditions from those present in training and testing.

AI model/system exploitation

Adversarial actions that exploit vulnerabilities an AI model or system to force misperformance against its intended objectives, disrupt access to its outputs or functionality, or enable unauthorized access to restricted or proprietary information.

AI governance

The set of organizational policies, rules, frameworks, roles, and oversight processes that direct how AI is adopted, developed, deployed, and monitored within the organization, with the objective of ensuring AI-related risks are identified, managed, and monitored across the AI lifecycle.

AI lifecycle

The set of phases an AI system goes through. These are plan and design, collect and process data, build and use model, verify and validate, deploy and use, and operate and monitor. These phases are often iterative, and not necessarily sequential.

AI model

A component of an information system that implements AI technology and uses computational, statistical, or machine-learning techniques to produce outputs from a given set of inputs.

AI risk assessment

A risk-management process for identifying, estimating, and prioritizing risks arising from the operation and use of an AI system, incorporating threat and vulnerability analyses and considering mitigations provided by controls planned or in place.

AI as a service (AIaaS)

Cloud-based systems providing on demand services to organizations and individuals to deploy, develop, train, and manage AI models.

AI system

The term 'artificial intelligence system'

(A) means any data system, software, application, tool, or utility that operates in whole or in part using dynamic or static machine learning algorithms or other forms of artificial intelligence, whether

(i) the data system, software, application, tool, or utility is established primarily for the purpose of researching, developing, or implementing artificial intelligence technology; or

(ii) artificial intelligence capability is integrated into another system or agency business process, operational activity, or technology system; and

(B) does not include any common commercial product within which artificial intelligence is embedded, such as a word processor or map navigation system.

AI use case

A specific scenario in which AI is designed, developed, procured, or used to achieve a particular objective, such as delivering a product or service, enhancing decision making, or providing a defined benefit.

AI use case inventory

A maintained repository or listing of an organization's AI use cases, intended to support governance, transparency, and risk management by documenting where and how AI is designed, developed, procured, or used, and the purpose and outputs associated with those uses.

Algorithm

A clearly specified mathematical process for computation; a set of rules that, if followed, will give a prescribed result.

Algorithmic trading system

A system that fundamentally depends upon computerized algorithms, and the data and technological infrastructure through which they operate, to address various decisions and tasks associated with trading financial instruments.

Anomaly detection system

A system for identifying the occurrence of a condition that deviates from expectations based on requirements specifications, design documents, user documents, or standards, or from someone's perceptions or experiences.

Artificial general intelligence (AGI)

The currently hypothetical level of AI capability that is able to understand or learn an intellectual task as human being can. It is an AI system that can perform across diverse cognitive domains with versatility and proficiency, rather than being limited to a narrow task or domain.

Artificial intelligence (AI)

The term 'artificial intelligence' means a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. Artificial intelligence systems use machine and human-based inputs to

(A) perceive real and virtual environments;

(B) abstract such perceptions into models through analysis in an automated manner; and

(C) use model inference to formulate options for information or action.

Autonomous agent

An AI system that operates independently, taking actions in a defined environment based on its objectives and observations without requiring step-by-step human direction. Autonomous agents may use planning, memory, and tool access to complete multi-step tasks on behalf of a user or system.

B

Benchmarking

An alternative prediction or approach used to compare a model's inputs and outputs to estimates from alternative internal or external data or models.

Bias

A systematic distortion, as opposed to random error, that reduces the representativeness or accuracy of an AI system's outputs or performance for its intended purposes and operating conditions. Bias may be introduced inadvertently or purposely, and may also emerge as the AI system is used in an application; this could arise when the data used to develop or operate the system are not representative of the intended population or operating conditions. Common sources/subcategories of bias include statistical/computational, systemic, and human bias.

Black box

The nature of some AI techniques whereby the inferential operations are complex, hidden, or otherwise opaque to their developers and end users in terms of providing an understanding of how classifications, recommendations, or actions are generated and what overall performance will be.

C

Capability evaluation

A comprehensive assessment of an AI model's or system's overall capabilities, including both planned capabilities and unplanned, emerging, or malicious capabilities. Unlike specific task-focused evaluations this evaluation seeks to understand the full range of an AI's capabilities. This includes evaluating how an AI might adapt or evolve beyond its initial training, identifying both beneficial emergent behaviors and potential risks that could arise from its autonomous operation or interaction with complex environments.

Chain-of-thought prompting

A prompting technique in which a user or system instructs an AI model to reason through a problem step by step before delivering a final answer. Chain-of-thought prompting tends to improve accuracy on complex or multi-step tasks by making the model's intermediate reasoning explicit and checkable.

Computer vision

The digital process of perceiving and learning visual tasks in order to interpret and understand the world through cameras and sensors.

Context poisoning

An attack in which malicious content is injected into the information an AI model receives during a session, causing it to behave in unintended or harmful ways based on the corrupted context. Context poisoning is closely related to prompt injection but specifically targets the broader context window rather than a single prompt.

D

Data lineage

The history of processing of a data element, which may include point-to-point data flows and the data actions performed upon the data element.

Data poisoning

An attack that corrupts and contaminates training data to compromise an AI system's performance.

Data quality/validity

The usefulness, accuracy, and correctness of data for its application.

Deep learning

A machine learning implementation technique that uses large quantities of data, or feedback from interactions with a simulation or an environment, as training sets for a network with multiple hidden layers, called a deep neural network, often employing an iterative optimization technique called gradient descent, to tune large numbers of parameters that describe weights given to connections among units.

Deepfake

AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places or other entities or events and would falsely appear to a person to be authentic or truthful.

Deterministic (algorithm / model)

An algorithm/model that, given the same inputs, always produces the same outputs.

Diffusion models

A type of generative AI model that produces output to match a prompt by iteratively refining noise. These types of models require substantial computational resources and processing time.

Documentation

The collection of records that describe an AI system's purpose and intended use, key design choices, training and operational data characteristics and provenance, testing and evaluation results, limitations, and version history, maintained to support transparency, oversight, and accountability across the AI lifecycle.

E

Emergent behavior

Capabilities or behaviors that arise in an AI model that were not explicitly trained or anticipated, typically appearing as models scale in size or complexity. Emergent behaviors can be beneficial, such as improved reasoning, or potentially harmful, such as the ability to execute tasks the model was not designed or intended to perform.

Explainability

Property of an AI system that enables a given human audience to comprehend the reasons for the system's behavior; the ability to understand an AI system's output and decision given certain inputs.

F

Federated learning

A method of training AI models across multiple devices or organizations without sharing underlying data. This machine learning architecture helps preserve privacy while enabling collaborative machine learning.

Fine-tuning

The process of taking a pre-trained machine learning model and further training it on a smaller, task-specific dataset to improve performance for a defined use case. Fine-tuning adapts general-purpose models to organizational data, workflows, or objectives.

Foundation models

Large machine learning models trained on vast amounts of raw and unlabeled data through unsupervised learning that can be adapted and applied to versatile downstream tasks. Large language models are common subsets of foundation models and underpin many generative AI applications in the financial sector.

G

General purpose AI

AI designed for use across a broad array of tasks across many different applications rather than for a specific domain.

Generative Adversarial Networks (GANs)

A machine learning framework in which two neural networks contest with each other in the form of a zero-sum game, where one agent's gain is another agent's loss. A GAN learns to generate new data with the same statistics as the training set.

Generative AI

The class of AI that emulate the structure and characteristics of input data in order to generate derived synthetic content. This can include images, videos, audio, text, and other digital content.

Guardrails

Layered safeguards to prevent access to bad information and behavior in an AI system. These may encompass policies, technical controls, and monitoring mechanisms, and may exist at the data, model, application, and infrastructure levels. These safeguards aim to ensure generative AI systems behave ethically, safely, and within organizational or regulatory boundaries by filtering training data, aligning model behavior, and enforcing post-deployment controls.

H

Hallucination

A phenomenon when AI produces output that is erroneous or flawed but is still in the form of a convincing narrative or presentation. Generative AI can still produce flawed information even if underlying data is free of defects.

Human-AI collaboration

A working model in which humans and AI systems jointly perform tasks, combining machine efficiency with human judgment, oversight, and decision-making responsibility.

Human biases

These biases reflect systematic errors in human thought based on a limited number of heuristic principles and predicting values to simpler judgmental operations. These biases are omnipresent in the institutional, group, and individual decision-making processes across the AI lifecycle, and in the use of AI applications once deployed.

Human-in-the-Loop (HITL)

A risk-control approach for AI where a human is integrated within the AI's decision-making process.

I

Inference guardrails

Controls applied at the point when an AI model generates a response, designed to filter, block, or modify outputs that violate safety, policy, or accuracy requirements. Inference guardrails operate after the model has processed a prompt and serve as a last line of defense before output reaches the end user.

Interpretability

Transparency into the inner workings of AI output in the context of their designed functional purposes, which helps users gain deeper insights into the functionality and trustworthiness of the system and its outputs.

K

Knowledge cutoff

The date beyond which a large language model has no training data and therefore no awareness of events, publications, or developments. A model's knowledge cutoff is a key limitation to understand when using AI for research, compliance, or time-sensitive decisions.

L

Large language model (LLM)

A subset of machine learning that uses algorithms trained on large amounts of data through self-supervised machine learning to recognize patterns and respond to user requests in natural language.

LLM inference

The phase in which a large language model (LLM) generates outputs such as responses, summaries, or predictions based on new input data. Inference represents the operational use of the model after training.

M

Machine learning (ML)

An AI learning method that enables computational systems to learn patterns, make predictions, and optimize decisions from large amounts of data without being explicitly programmed for each task. Machine learning encompasses supervised, unsupervised, and reinforcement learning paradigms, serving as the technical foundation for data-driven intelligence and automation.

Model alignment

The process of ensuring that an AI system's outputs and behavior are consistent with human values, organizational policies, and intended objectives. Alignment is achieved through training techniques, guardrails, and evaluation processes.

Model context window

The maximum amount of text or data, measured in tokens, that an AI model can process at one time when generating a response. The size of the context window determines how much information the model can consider in a single interaction.

Model integrity

The process of protecting a model against improper information modification or destruction and ensuring information non-repudiation and authenticity.

Model risk

The potential for adverse consequences from decisions based on incorrect or misused model outputs and reports. Model risk can be from individual models and be in the aggregate. Aggregate model risk is affected by interaction and dependencies among models; reliance on common assumptions, data, or methodologies; and any other factors that could adversely affect several models and their outputs.

Multi-modal model

A model that processes and relates information from multiple data modalities, such as text, images, audio, and sensor data, among others.

Multiagent orchestration

The coordination of multiple AI agents working in parallel or in sequence to complete tasks that exceed the capabilities or scope of a single agent. Orchestration systems manage task delegation, communication between agents, and aggregation of results into a coherent output.

N

Natural language processing (NLP)

The ability of a machine to process, analyze, and mimic human language, either spoken or written.

Neural network

A computational model loosely inspired by the structure of the human brain, consisting of interconnected nodes (neurons) organized in layers. Neural networks learn to perform tasks by adjusting the strength of connections between nodes based on exposure to training data. They serve as the foundational architecture for deep learning and most modern AI systems.

O

Output validation

Systematic process of verifying and confirming that AI system outputs meet specified requirements, accuracy standards, and quality criteria before being used for downstream processes.

Overfitting / Underfitting

Overfitting occurs when a model learns the training data too closely, including its noise and anomalies, causing it to perform poorly on new data it hasn't seen before. Underfitting is the opposite: the model is too simple to capture meaningful patterns in the training data and performs poorly on both training and new data. Both conditions indicate a model that is not properly generalized for real-world use.

Override

Output or input that is ignored, altered, rejected, or reversed.

P

Parameter

A numerical value within a machine learning model that is adjusted during training to minimize prediction error. Parameters define a model's learned knowledge; the count of parameters is commonly used to describe a model's size and capacity. Large language models may contain billions of parameters.

Performance monitoring

Ongoing activities that confirm an AI system is implemented appropriately, used as intended, and continues to perform as intended over time.

Performance threshold

A particular value or range of values of a performance measure or diagnostic that determines the acceptance or rejection of a model's performance.

Predictive analytics

A discipline within AI that leverages historical data, statistical algorithms, and machine learning techniques to identify patterns and forecast future outcomes, behaviors, or events. This discipline is distinguished by emphasis on forward-looking insights rather than descriptive analysis.

Prompt

Natural language text describing the task an AI should perform.

Prompt injection

An attack on an AI system that exploits how an application combines untrusted input with a prompt written by a higher-trust party, such as the application designer, so the system follows the untrusted instructions.

Prompt jailbreak

An attempt to bypass an AI system's safeguards by crafting inputs that cause the system to produce restricted, unsafe, or unintended outputs.

R

Reasoning model

A class of large language model specifically trained or prompted to work through problems in a structured, step-by-step manner before producing a final answer. Reasoning models are particularly useful for tasks involving logic, mathematics, multi-step analysis, and complex decision-making.

Red-teaming (AI)

A structured testing process in which security professionals or researchers deliberately attempt to find failure modes, vulnerabilities, or policy violations in an AI system before deployment. AI red-teaming typically involves adversarial prompting, model probing, and attempts to elicit harmful or unintended outputs.

Reinforcement learning

A type of machine learning in which a model learns to optimize its behavior according to a reward function by interacting with and receiving feedback from an environment.

Representation learning

Also known as feature learning, a set of techniques for automatically detecting feature patterns that replaces manual feature engineering.

Responsible AI

Conscientious design, deployment, and governance of AI systems aligned with ethical principles, societal values, and legal requirements.

Retrieval augmented generation (RAG)

A type of generative AI system in which a model is paired with a separate information retrieval system (or "knowledge base"). Based on a user query, the RAG system identifies relevant information within the knowledge base and provides it to the generative AI model in context for the model to use in formulating its response. RAG systems allow the internal knowledge of a generative AI model to be modified without the need for retraining.

S

Shadow AI

The use of AI tools or systems within an organization without formal approval, governance, or oversight. Shadow AI introduces risk related to data exposure, compliance gaps, and inconsistent security controls.

Service level agreement (SLA)

Contractually binding terms, often incorporated into a broader services contract, between a service provider and a customer that specify the services to be delivered and the measurable performance and service-quality commitments, such as availability and response times. SLAs also typically define each party's responsibilities and provisions for monitoring/reporting, issue resolution, and remedies if service levels are not met.

Supervised learning

A process for training algorithms by example. The training data consists of inputs paired with the correct outputs. During training, the algorithm will search for patterns in the data that correlate with the desired outputs and learn to predict the correct output for newly presented input data over iterative training and model updates.

Structured data

Data that is divided into standardized pieces that are identifiable and accessible by both humans and computers.

Swarm

Swarm shows up in a few distinct AI contexts:

(A) Swarm intelligence is the oldest use, as it's a subfield of AI inspired by collective behavior in nature (ant colonies, bird flocking, bee swarms). Algorithms like Particle Swarm Optimization (PSO) and Ant Colony Optimization fall here. It's been around since the 1990s.

(B) Multi-agent swarms is the more current, buzzy usage. With the rise of agentic AI, "swarm" now commonly refers to networks of AI agents working in parallel or in coordination to complete complex tasks. OpenAI released an experimental framework called Swarm in late 2024 specifically for orchestrating multi-agent workflows.

Synthetic data

Data that has been generated using a purpose built mathematical model or algorithm, that is statistically realistic but artificial, that can be used for activities like model development and training.

Synthetic identity

The use of a combination of real and fake personally identifiable information (PII) to fabricate a person or entity.

T

Text/word embedding

A numerical vector representation of text that machine learning and artificial intelligence systems use to work with meaning in text, such as comparing similarity between pieces of text.

Third-party AI risk

Risk that arises when an organization relies on another entity to develop, provide, host, operate, or support AI systems or key AI components such as models, data, and related infrastructure.

Tokens

Units of text processed by a language model, which may represent whole words, partial words, or characters. Token usage is commonly used to measure input size, output size, and cost in AI systems.

Traditional AI

Traditional AI, also referred to as symbolic or rule-based AI, is a subset of AI that focuses on performing discrete, preset tasks using predetermined algorithms and rules. These AI applications are designed to excel in a single activity or a restricted set of tasks, such as playing chess, diagnosing diseases, or translating languages.

Training data

A subset of input data samples used to train a machine learning model.

U

Unstructured data

Data that does not have a predefined data model or is not organized in a predefined way. This may also include data that is more free form, such as multimedia files, images, sound files, or unstructured text. Unstructured data does not necessarily follow any format or hierarchical sequence, nor does it follow any relational rules.

Unsupervised learning

A learning strategy that consists in observing and analyzing different entities and determining that some of their subsets can be grouped into certain classes, without any correctness test being performed on acquired knowledge through feedback from external knowledge sources.

V

Validation

Confirmation, through objective evidence, that an AI system or model meets requirements for a specific intended use or application and achieves its intended use in its intended operational environment.

Vector database

A database optimized for storing and querying vector representations of data, enabling fast similarity search. Vector databases are commonly used to support retrieval augmented generation and semantic search.

Version control

Systematic practice of tracking, managing, and documenting changes to AI assets through their development and deployment lifecycle.

Z

Zero-shot / Few-shot learning

Zero-shot learning refers to an AI model's ability to perform a task it was not explicitly trained on, relying solely on general knowledge acquired during pre-training. Few-shot learning extends this by providing the model with a small number of examples within the prompt to guide its response. Both approaches are commonly used when task-specific training data is unavailable or impractical to collect.