LangProtect Glossary

Explore clear, concise definitions of essential AI and machine learning terms. Use the search bar below to find any concept quickly.

Activation Function

A mathematical function used in neural networks that determines whether a neuron should be activated, enabling models to learn complex patterns and relationships.

Active Learning

A machine learning approach where the model selectively requests labels for the most informative data samples to improve learning with minimal labeled data.

Adversarial Machine Learning

A field focused on studying and defending against attacks that manipulate machine learning models through carefully crafted inputs or poisoned training data.

AGI (Artificial General Intelligence)

A theoretical form of AI capable of understanding, learning, and performing any intellectual task at a human level across diverse domains.

AI Acceptable Use Policy

A document that defines how employees or users can safely and responsibly use AI systems while complying with organizational security, legal, and ethical requirements.

AI Accountability

The practice of assigning responsibility for AI decisions, outcomes, and risks through governance, oversight, documentation, and compliance with organizational or regulatory requirements.

AI Agent / Agentic AI

An AI system that can autonomously plan, reason, use tools, and execute multi-step tasks with minimal human intervention to achieve specific goals.

AI Code Assistants

AI-powered development tools that help developers write, review, debug, and optimize code using natural language prompts and contextual understanding.

AI Fairness

The practice of designing AI systems that make unbiased, equitable, and non-discriminatory decisions across different individuals and demographic groups.

AI Observability

The ability to monitor, analyze, and understand AI system behavior using logs, metrics, traces, and performance insights throughout the model lifecycle.

AI Pipeline

The end-to-end workflow for developing, deploying, monitoring, and maintaining AI models, including data preparation, training, testing, and production operations.

AI Security

The discipline of protecting AI models, applications, and data from attacks, misuse, unauthorized access, manipulation, and other security threats throughout their lifecycle.

Annotation

The process of labeling or tagging data with meaningful information so it can be used to train, validate, and evaluate machine learning models.

Anomaly Detection

A machine learning technique used to identify unusual patterns, behaviors, or data points that differ significantly from expected or normal activity.

Artificial Neural Network (ANN)

A machine learning model inspired by the human brain that consists of interconnected neurons capable of learning patterns from data.

Audit Logging

The process of recording AI system activities, user interactions, and security events to support monitoring, investigations, compliance, and forensic analysis.

Augmentation (Data Augmentation)

A technique that expands training datasets by creating modified versions of existing data to improve model accuracy and generalization.

Autoencoder

A neural network that learns to compress data into a compact representation and reconstruct it, commonly used for anomaly detection and dimensionality reduction.

AutoML

A collection of tools and techniques that automate machine learning tasks such as feature engineering, model selection, hyperparameter tuning, and deployment.

Autoregressive Model

A model that generates outputs sequentially by predicting each new element based on previously generated or observed elements.

Average Precision (AP)

A performance metric that summarizes the precision-recall curve by measuring how well a model identifies relevant positive predictions.

Backpropagation

A training algorithm that updates neural network weights by propagating prediction errors backward through the network to minimize loss.

Bagging (Bootstrap Aggregating)

An ensemble learning technique that trains multiple models on randomly sampled datasets and combines their predictions to improve accuracy and reduce variance.

Baseline Models

Simple reference models used to establish a performance benchmark before evaluating more advanced machine learning models.

Batch Normalization

A neural network technique that normalizes intermediate layer outputs during training, improving stability, convergence speed, and overall model performance.

Bayesian Inference

A statistical method that updates the probability of an outcome as new evidence becomes available using Bayes' theorem.

BERT (Bidirectional Encoder Representations from Transformers)

A transformer-based language model developed by Google that understands text by analyzing context from both directions simultaneously.

Bias

Systematic errors in AI models that produce unfair, inaccurate, or discriminatory outcomes due to biased data, algorithms, or model design.

Bias-Variance Tradeoff

The balance between underfitting and overfitting, where reducing one source of prediction error often increases the other.

Biases in AI

Systematic unfairness in AI systems caused by biased data, algorithms, or design choices that result in discriminatory or inaccurate outcomes.

Bigotry

Content or behavior that promotes hatred, discrimination, or prejudice against individuals or groups based on protected characteristics such as race, religion, or gender.

Binary Classification

A machine learning task where a model predicts one of two possible classes, such as spam or not spam.

Binomial Distribution

A probability distribution that models the number of successful outcomes in a fixed number of independent trials with the same success probability.

Black Box Model

A machine learning model whose internal decision-making process is difficult or impossible for humans to interpret or explain.

BLEU (Bilingual Evaluation Understudy)

An evaluation metric that measures the quality of machine-translated text by comparing it with one or more human-generated reference translations.

Bounding Box

A rectangular outline used in computer vision to identify and locate objects within an image or video.

Bring-Your-Own Policy (BYOP)

An organizational policy allowing employees to use their preferred AI tools or models while enforcing security, compliance, and governance requirements.

Calibration Curve

A graph that compares predicted probabilities with actual outcomes to evaluate how well a machine learning model's confidence matches reality.

Canonical Schema

A standardized data structure that ensures information is represented consistently across systems, enabling reliable data exchange, integration, and interoperability between applications.

Catastrophic Forgetting

A problem where a machine learning model loses previously learned knowledge after being trained on new data, reducing performance on earlier tasks.

Categorical Variables

Variables that represent discrete categories or labels rather than numerical values, such as colors, countries, or product types.

Chatbots

AI-powered conversational systems that interact with users through natural language to answer questions, provide assistance, or automate customer interactions.

ChatGLM

An open-source large language model developed for multilingual conversational AI, designed to support efficient dialogue generation and natural language understanding.

Chief AI Officer

A senior executive responsible for overseeing an organization's AI strategy, governance, adoption, risk management, and responsible AI initiatives.

Child Safety

AI safety practices designed to prevent the generation, promotion, or distribution of content that exploits, harms, or endangers children.

CI/CD

Continuous Integration and Continuous Deployment (CI/CD) is a software development practice that automates code integration, testing, and deployment for faster, more reliable releases.

Class Imbalance

A situation where one class in a dataset contains significantly more samples than another, making model training and evaluation more challenging.

Classification Threshold

The probability cutoff used by a classification model to determine whether an input belongs to a particular class.

Classifier

A machine learning model that categorizes input data into predefined classes based on learned patterns from labeled training data.

Clustering Algorithms

Machine learning algorithms that group similar data points together without predefined labels based on shared patterns or characteristics.

CNN (Convolutional Neural Networks)

A deep learning architecture designed to process visual data by automatically learning spatial features from images and videos.

Collaborative Filtering

A recommendation technique that predicts user preferences by analyzing similarities between users, items, or historical interactions.

Common Crawl

A publicly available web dataset containing billions of webpages that is widely used to train and evaluate large language models.

Complex Event Processing (CEP)

A technology that analyzes multiple real-time events to identify meaningful patterns, detect anomalies, and trigger automated actions or decisions.

Computer Vision

A field of artificial intelligence that enables computers to analyze, interpret, and understand images, videos, and other visual information.

Content Moderation

The process of detecting, reviewing, and removing harmful, illegal, or policy-violating content from AI systems or online platforms.

Content Safety Classifier

An AI model that detects and categorizes unsafe, harmful, or policy-violating content before it reaches users or downstream applications.

Continuous Integration Model

A development approach where code changes are frequently integrated, automatically tested, and validated to detect issues early in the software lifecycle.

Continuous Validation

The ongoing process of evaluating AI models using new data to ensure they remain accurate, reliable, and aligned with expected performance.

Convex Optimization

A mathematical optimization technique used to find the best solution to problems where the objective function has a single global optimum.

CPU (Central Processing Unit)

The primary processor in a computer responsible for executing instructions, performing calculations, and managing overall system operations.

Cross-Entropy Loss

A loss function used in classification tasks that measures the difference between predicted probabilities and the actual class labels.

Cross-Validation Modeling

A model evaluation technique that repeatedly splits data into training and validation sets to assess performance and improve generalization.

Dangerous Substances

Content involving hazardous chemicals, drugs, explosives, or other harmful materials that AI systems should restrict or carefully moderate.

Data Cleaning

The process of identifying and correcting inaccurate, incomplete, duplicate, or inconsistent data to improve its quality for analysis or model training.

Data Decomposition

The process of breaking complex datasets into smaller, meaningful components to simplify analysis, storage, or machine learning workflows.

Data Drift

A change in the statistical properties or distribution of input data over time that can reduce the accuracy of machine learning models.

Data Mining

The process of discovering meaningful patterns, relationships, and insights from large datasets using statistical and machine learning techniques.

Data Science

An interdisciplinary field that combines statistics, machine learning, programming, and domain expertise to extract insights and support data-driven decision-making.

Data Versioning

The practice of tracking and managing changes to datasets over time, enabling reproducibility, collaboration, and recovery of previous data versions.

Data-Centric AI

An AI development approach that focuses on improving data quality, labeling, and management rather than primarily modifying model architectures.

Dataset

A structured collection of related data used for analysis, training, testing, or validating machine learning models and other data-driven applications.

Deceptive AI Behavior

AI behavior that intentionally or unintentionally misleads users through false claims, fabricated information, hidden intentions, or manipulative responses.

Decision Boundary

The boundary that separates different classes in a machine learning model, determining how new data points are classified.

Decision Tree

A supervised machine learning algorithm that makes predictions by splitting data into branches based on feature values until a final decision is reached.

Deep Belief Networks (DBNs)

A class of deep neural networks composed of multiple layers of hidden units that learn hierarchical representations of data through unsupervised pretraining.

Deep Blue

IBM's chess-playing supercomputer that became the first AI system to defeat a reigning world chess champion in a match under standard tournament conditions.

Deep Learning

A subset of machine learning that uses multi-layered neural networks to automatically learn complex patterns from large amounts of data.

Deep Reinforcement Learning (DRL)

A machine learning approach that combines deep learning with reinforcement learning to enable intelligent decision-making in complex environments.

Deepfake

AI-generated or manipulated media that realistically imitates a person's appearance, voice, or actions, often making fabricated content appear authentic.

Denial of Wallet

An attack that intentionally increases AI usage or inference costs by triggering excessive requests, expensive operations, or unnecessary model executions.

DenseNet

A convolutional neural network architecture where each layer connects directly to every subsequent layer, improving feature reuse and reducing training complexity.

Density-Based Clustering

A clustering technique that identifies groups of closely packed data points while separating them from sparse regions or noise.

Dimensionality Reduction

The process of reducing the number of input features while preserving important information, making models faster, simpler, and easier to visualize.

DLP for AI

Data Loss Prevention for AI uses policies and automated controls to prevent sensitive information from being exposed through AI prompts, responses, or connected applications.

Drift Monitoring

The continuous monitoring of data and model behavior to detect changes that could reduce prediction accuracy or model performance over time.

Early Stopping

A training technique that prevents overfitting by stopping model training when performance on validation data no longer improves.

Embedding

A dense numerical representation of data, such as words or images, that captures meaningful relationships and semantic similarities in a lower-dimensional space.

Ensemble Learning

A machine learning approach that combines predictions from multiple models to improve overall accuracy, robustness, and generalization.

Epoch

One complete pass through the entire training dataset during the training process of a machine learning model.

Ethical AI

The development and use of AI systems that prioritize fairness, transparency, accountability, privacy, and respect for human rights.

EU AI Act

The European Union's comprehensive AI regulation that classifies AI systems by risk and establishes legal requirements for trustworthy AI development and deployment.

Evolutionary Algorithms

Optimization algorithms inspired by natural evolution that use selection, mutation, and reproduction to find high-quality solutions to complex problems.

Excessive Agency

The condition where an AI system is granted unnecessary autonomy, permissions, or capabilities that increase operational and security risks.

Excessive Model Curiosity

A risky AI behavior where a model unnecessarily requests, infers, or collects more sensitive information than required to complete a task.

Explainable AI (XAI)

An approach to AI that makes model decisions understandable and transparent, helping users trust, validate, and audit AI-generated outcomes.

F-score (F1 Score)

A performance metric that combines precision and recall into a single score, providing a balanced measure of classification accuracy.

Facial Recognition

A computer vision technology that identifies or verifies individuals by analyzing unique facial features from images or video.

Factual Inconsistency

A situation where an AI model generates information that contradicts verified facts, trusted sources, or previously established context.

False Negative

A prediction error where a model incorrectly classifies a positive instance as negative, failing to detect the expected outcome.

False Positive Rate

The proportion of negative instances incorrectly classified as positive, commonly used to evaluate the performance of classification models.

FDA

The U.S. Food and Drug Administration, responsible for regulating AI-enabled medical devices and ensuring their safety, effectiveness, and regulatory compliance.

Feature Selection

The process of identifying and selecting the most relevant input features to improve model accuracy, efficiency, and interpretability.

Feature Vector

A numerical representation of an object or data instance where each value corresponds to a measurable characteristic or feature.

Federated Learning

A machine learning approach where models are trained across multiple devices or organizations without sharing the underlying data, improving privacy and security.

Feed-forward Neural Networks (FNNs)

A type of neural network where information flows only from the input layer to the output layer without loops or feedback connections.

Feedback Loop Optimization

The continuous improvement of AI models by using user feedback, performance data, or system outcomes to refine future predictions and responses.

Few-Shot Learning

A machine learning technique where a model learns new tasks using only a small number of labeled examples.

Fine-Tuning

The process of adapting a pre-trained AI model using additional domain-specific data to improve performance for a particular task or application.

Firewall for AI

A security layer that inspects AI prompts and responses to detect, block, and prevent threats such as prompt injection, data leakage, and unsafe outputs.

Foundational Model

A large AI model trained on diverse datasets that serves as a general-purpose base for multiple downstream applications and specialized tasks.

Fuzzer for AI

A testing tool that automatically generates unexpected or malicious inputs to identify vulnerabilities, safety failures, or security weaknesses in AI systems.

GAN (Generative Adversarial Network)

A deep learning architecture consisting of a generator and discriminator that compete to create realistic synthetic data.

Garbage In, Garbage Out (GIGO)

A principle stating that poor-quality or inaccurate input data leads to unreliable or inaccurate outputs from an AI or computer system.

GenAI (Generative AI)

Artificial intelligence that creates new content such as text, images, code, audio, or video by learning patterns from existing data.

GenAI Deployment

The process of integrating, configuring, securing, and operating generative AI models within production applications or enterprise environments.

GenAI Inventory

A centralized record of generative AI models, applications, tools, and services used across an organization for governance, visibility, and compliance.

GenAI Red Teaming

A structured security assessment that tests generative AI systems against attacks, misuse scenarios, and safety failures before deployment.

GenAI Security

The practice of protecting generative AI models, applications, and data from cyber threats, misuse, prompt attacks, and unauthorized access.

GPU (Graphics Processing Unit)

A specialized processor designed to perform parallel computations, making it ideal for training and running machine learning models.

Gradient Descent

An optimization algorithm that iteratively updates model parameters to minimize prediction error and improve model performance.

Graphic Violence

Content depicting extreme physical harm, injury, or death that AI systems typically detect, restrict, or moderate for user safety.

Ground Truth

The verified, accurate data used as the reference standard for training, validating, and evaluating machine learning models.

GRU (Gated Recurrent Unit)

A type of recurrent neural network that efficiently captures sequential patterns while reducing computational complexity compared to LSTMs.

Guardrails

Policies, controls, and safety mechanisms that restrict AI behavior, helping prevent harmful outputs, policy violations, and security risks.

Hallucination

An AI-generated response that appears believable but contains fabricated, inaccurate, or unsupported information not grounded in reliable sources.

Hate Speech

Language or content that attacks, degrades, or promotes hatred against individuals or groups based on protected characteristics.

Hidden Layer

An intermediate layer in a neural network where data is transformed and meaningful patterns are learned before reaching the output layer.

HIPAA

The Health Insurance Portability and Accountability Act, a U.S. law that protects sensitive patient health information and governs healthcare data privacy.

Human Exploitation

Content or AI behavior that promotes, facilitates, or normalizes the abuse, trafficking, coercion, or exploitation of individuals.

Human-in-the-Loop (HITL)

An AI workflow where human oversight is included to review, validate, or approve important decisions before automated actions are completed.

Hyperparameter

A configurable value set before training that controls how a machine learning model learns, such as learning rate or batch size.

IID (Independent and Identically Distributed)

A statistical assumption that each data sample is independent of others and follows the same probability distribution.

Illegal Activity

Content that encourages, facilitates, or provides instructions for unlawful acts, requiring detection and restriction by AI safety systems.

Image Collection

A structured set of images used to train, validate, or evaluate computer vision and machine learning models.

Image Pre-processing

The process of preparing images for machine learning by resizing, normalizing, enhancing, or transforming them to improve model performance.

Image Recognition

A computer vision task where AI models identify and classify objects, people, scenes, or other visual elements within an image.

Image Segmentation

A computer vision technique that divides an image into meaningful regions by assigning labels to individual pixels or groups of pixels.

ImageNet

A large-scale labeled image dataset widely used for training, benchmarking, and advancing computer vision and deep learning models.

Impersonation Attacks

Attempts to deceive users or AI systems by pretending to be another person, organization, or trusted entity for malicious purposes.

Indirect Prompt Injection

An attack where hidden instructions embedded in external content manipulate an AI model without direct interaction from the user.

Inference

The process of using a trained machine learning model to generate predictions or decisions from new, unseen input data.

Information Retrieval

The process of searching, finding, and ranking relevant information from large collections of documents, databases, or other data sources.

Input Obfuscation

Techniques that hide malicious prompts or sensitive instructions using encoding, formatting, or indirect language to bypass AI security controls.

Instance Segmentation

A computer vision task that identifies individual objects within an image and precisely outlines each object's boundaries.

Intelligent Document Processing (IDP)

An AI-powered technology that automatically extracts, classifies, and processes information from structured and unstructured documents.

IP & Copyright Infringement

The unauthorized use, reproduction, or distribution of copyrighted or proprietary content through AI-generated outputs or training data.

ISO/IEC 42001:2023

An international standard that defines requirements for establishing, implementing, and improving an Artificial Intelligence Management System (AIMS).

Jailbreaking

A prompt-based attack that attempts to bypass an AI model's safety controls and force it to produce restricted or harmful outputs.

K-means

An unsupervised machine learning algorithm that groups similar data points into a predefined number of clusters based on feature similarity.

K-nearest Neighbour (KNN)

A supervised machine learning algorithm that classifies or predicts values based on the closest labeled data points in the feature space.

Keras

An open-source deep learning framework that provides a simple, high-level interface for building and training neural networks.

Labeled Data

Data that has been tagged with the correct output or category, enabling supervised machine learning models to learn accurate predictions.

LangChain

An open-source framework that helps developers build AI applications by connecting large language models with tools, APIs, databases, and multi-step workflows.

Learning Rate

A hyperparameter that determines how much a machine learning model updates its parameters during each training step to minimize prediction errors.

LightGBM

An open-source gradient boosting framework that builds fast, efficient tree-based models for classification, regression, and ranking tasks.

Linear Regression

A supervised learning algorithm that predicts continuous values by modeling the linear relationship between input variables and a target variable.

LLaMA

A family of open-source large language models developed by Meta for research and commercial applications in natural language processing.

LLM (Large Language Model)

A deep learning model trained on massive text datasets to understand, generate, summarize, and reason using natural language.

LLM Agents

AI systems powered by large language models that can reason, plan, use external tools, and autonomously complete multi-step tasks.

LLM Debugger

A tool that helps developers inspect, analyze, and troubleshoot large language model behavior, prompts, outputs, and performance issues.

LLM Evaluation

The process of measuring the quality, accuracy, safety, reliability, and effectiveness of large language models using benchmarks, tests, and human feedback.

LLM Parameters

The learned numerical values within a large language model that determine how it processes information and generates predictions or responses.

LLMOps

A set of practices for developing, deploying, monitoring, and managing large language models throughout their operational lifecycle.

Local Interpretable Model-Agnostic Explanations (LIME)

A model interpretation technique that explains individual predictions by approximating the behavior of complex machine learning models with simpler, interpretable models.

Log Loss

A performance metric that measures how well predicted probabilities match actual outcomes, with lower values indicating better classification performance.

Logistic Regression

A supervised learning algorithm used for binary classification that predicts the probability of an input belonging to a particular class.

Low-Rank Adaptation (LoRA)

A parameter-efficient fine-tuning technique that adapts large language models by training a small set of additional parameters instead of updating the entire model.

LSTM (Long Short-Term Memory)

A type of recurrent neural network designed to capture long-term dependencies in sequential data such as text, speech, and time-series data.

Machine Learning Workflow

The structured process of collecting data, training models, evaluating performance, deploying solutions, and continuously monitoring machine learning systems.

Machine Translation

An AI technique that automatically translates text or speech between languages while preserving the original meaning and context.

MCP Gateway

A gateway that manages, secures, and controls communication between AI applications and Model Context Protocol servers or connected tools.

Mean Absolute Error (MAE)

A regression metric that measures the average absolute difference between predicted values and actual values, regardless of error direction.

Mean Square Error (MSE)

A regression metric that calculates the average squared difference between predicted values and actual values, giving greater weight to larger errors.

Memory Injections

An attack that manipulates an AI system's stored memory or persistent context to influence future responses or behavior.

Meta-Learning

A machine learning approach where models learn how to learn, enabling them to adapt quickly to new tasks using limited training data.

ML (Machine Learning)

A branch of artificial intelligence where systems learn patterns from data to make predictions, classifications, or decisions without explicit programming.

ML Diagnostics

The process of identifying, analyzing, and resolving issues affecting the accuracy, reliability, or performance of machine learning models.

ML Interpretability

The ability to understand and explain how a machine learning model reaches its predictions, improving transparency, trust, and regulatory compliance.

ML Model Deployment

The process of integrating a trained machine learning model into a production environment where it can generate predictions for real-world applications.

ML Model Management

The practice of organizing, versioning, monitoring, updating, and maintaining machine learning models throughout their entire operational lifecycle.

ML Model Validation

The process of evaluating a machine learning model to ensure it performs accurately, reliably, and generalizes well to unseen data.

ML Orchestration

The coordination and automation of machine learning workflows, including data preparation, model training, deployment, monitoring, and maintenance.

ML Scalability

The ability of a machine learning system to maintain performance and efficiency as data volumes, workloads, or user demand increase.

MLOps

A set of practices that automate and manage the development, deployment, monitoring, and maintenance of machine learning models throughout their lifecycle.

Model Calibration

The process of adjusting a model so its predicted probabilities accurately reflect the true likelihood of real-world outcomes.

Model Context Protocol (MCP)

An open protocol that enables AI models to securely connect with external tools, data sources, and applications using a standardized interface.

Model Context Protocol Server

A server that exposes tools, resources, or data through the Model Context Protocol, allowing AI models to securely access external capabilities.

Model Degradation

The gradual decline in a model's performance over time due to changing data patterns, concept drift, or evolving real-world conditions.

Model Fairness

The principle of ensuring machine learning models make equitable decisions without introducing unfair bias against individuals or demographic groups.

Model Monitoring

The continuous tracking of a deployed model's accuracy, performance, data quality, and operational health to detect issues early.

Model Registry

A centralized repository for storing, versioning, tracking, and managing machine learning models and their associated metadata.

Model Selection

The process of evaluating multiple machine learning models and choosing the one that delivers the best performance for a specific task.

Model Training

The process of teaching a machine learning model by exposing it to training data and adjusting its parameters to minimize prediction errors.

Model Weight Exposure

The unauthorized disclosure or extraction of a model's learned parameters, increasing the risk of theft, replication, or adversarial attacks.

Model-Centric AI

An AI development approach that focuses on improving model architectures and algorithms rather than primarily enhancing the quality of training data.

Moderation Layer

A security layer that evaluates AI inputs and outputs against safety policies before content reaches users or downstream systems.

Multi-Class Classification

A classification task where a machine learning model predicts one label from three or more possible output classes.

Multi-Task Learning

A machine learning approach where a single model learns multiple related tasks simultaneously, improving efficiency and knowledge sharing.

Multi-Turn Simulation

A testing technique that evaluates AI behavior across multiple conversational exchanges to identify risks that emerge over extended interactions.

Multimodal AI

AI systems that can understand, process, and generate information across multiple data types, including text, images, audio, and video.

Naive Bayes Models

A family of probabilistic machine learning algorithms based on Bayes' theorem that assume features are conditionally independent.

Natural Language Understanding (NLU)

A branch of natural language processing that enables AI systems to interpret the meaning, context, and intent behind human language.

NCII (Non-Consensual Intimate Imagery)

Intimate images or videos shared without a person's consent, including AI-generated content, requiring strict detection and removal.

Negative Class

The category representing the absence of a target condition or event in a binary classification problem.

Neural Network Tuning

The process of optimizing a neural network by adjusting hyperparameters, architecture, or training settings to improve performance.

Neuron

The fundamental computational unit of a neural network that receives inputs, applies a mathematical function, and passes the output to other neurons.

NIST AI RMF

The National Institute of Standards and Technology AI Risk Management Framework, providing guidance for building trustworthy and responsible AI systems.

NLP (Natural Language Processing)

A branch of AI focused on enabling computers to understand, interpret, generate, and interact with human language.

Noise

Irrelevant, incorrect, or random data that obscures meaningful patterns and can reduce the accuracy and performance of machine learning models.

Normalization

A data preprocessing technique that scales numerical values to a common range, improving model training speed, stability, and performance.

NSFW Content

Content containing explicit sexual material, nudity, or other adult themes that typically requires filtering or access restrictions.

Null Accuracy

A baseline metric that measures the accuracy achieved by always predicting the most common class in a classification dataset.

Object Detection

A computer vision task that identifies, classifies, and locates one or more objects within an image or video.

Object Tracking

A computer vision technique that continuously follows the position and movement of objects across consecutive video frames.

Observability

The ability to monitor, analyze, and understand AI system behavior through logs, metrics, traces, and performance insights.

Observation

A single data record or instance within a dataset, consisting of one or more feature values used for analysis or model training.

Off-Policy Behavior

AI behavior that violates predefined policies, safety rules, or operational instructions despite being explicitly configured to follow them.

Off-Topic Output

An AI response that is unrelated or only loosely connected to the user's request, reducing accuracy and usefulness.

One-Hot Encoding

A technique that converts categorical values into binary vectors, allowing machine learning models to process non-numerical data effectively.

Optimizers

Algorithms that adjust a model's parameters during training to minimize the loss function and improve prediction accuracy.

Outlier

A data point that differs significantly from the rest of the dataset and may represent anomalies, errors, or rare events.

Output Filtering

The process of inspecting and modifying AI-generated responses to block unsafe, sensitive, or policy-violating content before delivery.

Output Layer

The final layer of a neural network that produces the model's prediction, classification, or generated output.

Output Obfuscation

Techniques that disguise harmful or restricted AI-generated content using encoding, formatting, or indirect wording to evade detection.

Overfitting

A modeling problem where a machine learning model learns training data too closely, resulting in poor performance on unseen data.

Pandas and NumPy

Popular open-source Python libraries used for data manipulation, numerical computing, and preparing datasets for machine learning applications.

Panoptic Segmentation

A computer vision technique that combines semantic and instance segmentation to identify every object and background region in an image.

Parameter

A value learned during model training that determines how a machine learning model processes inputs and generates predictions.

Parameter-Efficient Fine-Tuning (PEFT)

A fine-tuning approach that adapts large AI models by updating only a small subset of parameters, reducing computational and memory requirements.

Part-of-Speech Tagging (POS Tagging)

A natural language processing technique that assigns grammatical categories, such as nouns or verbs, to each word in a sentence.

Pascal (PASCAL VOC)

A widely used computer vision benchmark dataset for evaluating object detection, image classification, and image segmentation models.

Pattern Matching

The process of identifying predefined patterns or sequences within data, text, or images for analysis, search, or automation.

Pattern Recognition

A branch of machine learning focused on identifying recurring patterns and relationships in data for classification or prediction tasks.

PII Leakage

The accidental or unauthorized exposure of personally identifiable information through AI inputs, outputs, logs, or connected systems.

Policy-Adaptive Controls

Dynamic AI security controls that automatically adjust enforcement based on organizational policies, user roles, regulatory requirements, or risk levels.

Polygon

A series of connected points used to outline the precise shape and boundaries of objects in computer vision annotation.

Pooling Layers in CNN

Neural network layers that reduce the spatial dimensions of feature maps, improving computational efficiency and reducing overfitting.

Positive Class

The target category representing the presence of a condition, event, or characteristic in a binary classification problem.

Pre-trained Transformer

A transformer-based neural network that has already been trained on large datasets and can be adapted for specific downstream tasks.

Precision

A classification metric that measures the proportion of predicted positive results that are actually correct.

Predictive Model Validation

The process of testing a predictive model on independent data to verify its accuracy, reliability, and ability to generalize.

Probabilistic Classification

A classification approach where models predict the probability of each possible class rather than only the most likely outcome.

Profanity

Offensive or inappropriate language that AI systems may detect, filter, or moderate depending on platform policies and user settings.

Prompt Engineering

The practice of designing and refining prompts to improve the accuracy, reliability, and usefulness of AI model responses.

Prompt Filtering

The process of inspecting user prompts to detect and block malicious, unsafe, or policy-violating instructions before model execution.

Prompt Leak

The unintended or malicious disclosure of an AI model's hidden system prompt, instructions, or confidential prompt engineering logic.

Proxy Labels

Labels generated from indirect sources or heuristics instead of manual annotation, often used when obtaining true labels is expensive or impractical.

PyTorch

An open-source machine learning framework widely used for building, training, and deploying deep learning models with flexible and dynamic computation.

RAG (Retrieval-Augmented Generation)

An AI technique that combines language models with external knowledge retrieval to generate more accurate and context-aware responses.

Random Forest

A supervised machine learning algorithm that combines multiple decision trees to improve prediction accuracy and reduce overfitting.

Random Initialization

The process of assigning random starting values to a model's parameters before training begins, helping optimize the learning process.

Recall

A classification metric that measures the proportion of actual positive instances correctly identified by a machine learning model.

Rectified Linear Unit (ReLU)

A widely used neural network activation function that outputs zero for negative values and the input value for positive values.

Recurrent Neural Network (RNN)

A type of neural network designed to process sequential data by retaining information from previous inputs through internal memory.

Regression Model

A machine learning model used to predict continuous numerical values based on relationships between input features and target variables.

Regression Testing

A software testing practice that verifies existing functionality continues to work correctly after code changes, updates, or bug fixes.

Regular Expression (Regex)

A sequence of characters that defines a search pattern for matching, extracting, validating, or replacing text within strings.

Regularization

A technique that reduces model complexity by adding constraints during training, helping prevent overfitting and improve generalization.

Regularization Rate

A hyperparameter that controls the strength of regularization applied during model training, balancing model complexity and performance.

Reinforcement Learning

A machine learning approach where an agent learns optimal actions by interacting with an environment and receiving rewards or penalties.

Reinforcement Learning from AI Feedback (RLAIF)

A training approach where AI-generated evaluations replace or supplement human feedback to improve model performance and alignment.

Reinforcement Learning from Human Feedback (RLHF)

A model training technique that uses human preferences and evaluations to improve AI behavior, helpfulness, and alignment.

Reproducible AI

The practice of ensuring AI models, experiments, and results can be consistently recreated using the same data, code, and configurations.

ResNet

A deep convolutional neural network architecture that uses residual connections to enable the training of very deep neural networks.

Responsible AI (RAI)

The development and deployment of AI systems that prioritize fairness, transparency, accountability, privacy, safety, and ethical principles.

Retrieval Augmentation Abuse

An attack that manipulates retrieved external information to influence AI responses with malicious, misleading, or unauthorized content.

Ridge Regression

A regression algorithm that applies L2 regularization to reduce overfitting while maintaining all input features in the model.

Robotic Process Automation (RPA)

A technology that automates repetitive, rule-based business processes by using software bots to perform routine digital tasks.

ROC (Receiver Operating Characteristic) Curve

A graphical evaluation tool that measures a classification model's performance by comparing true positive and false positive rates across thresholds.

Root Mean Square Error (RMSE)

A regression metric that measures the average magnitude of prediction errors, giving greater weight to larger errors than MAE.

Rotating Proxies

A network of proxy servers that automatically changes IP addresses between requests to improve anonymity and distribute network traffic.

ROUGE

A set of evaluation metrics used to measure the quality of automatically generated text by comparing it with reference text.

Scikit-Learn

An open-source Python library that provides machine learning algorithms and tools for data preprocessing, model training, evaluation, and analysis.

Segmentation

The process of dividing data or images into meaningful groups or regions to improve analysis, classification, or computer vision tasks.

Selective Sampling

An active learning technique that chooses the most informative data samples for labeling to improve model performance while reducing annotation effort.

Semantic Segmentation

A computer vision technique that assigns a class label to every pixel in an image, identifying different object categories and regions.

Semi-Supervised Learning

A machine learning approach that trains models using a small amount of labeled data combined with a larger amount of unlabeled data.

Sensitive Information Exfiltration

The unauthorized extraction or disclosure of confidential business, customer, or personal data through AI systems.

Sensitivity

A classification metric that measures a model's ability to correctly identify actual positive instances, also known as recall or true positive rate.

Sentiment Analysis

A natural language processing technique that identifies and classifies the emotional tone or opinion expressed in text.

Sextortion

A form of online exploitation where victims are threatened with the release of intimate content unless they comply with demands.

Shadow AI

The unauthorized use of AI tools or models by employees without organizational approval, increasing security, compliance, and data privacy risks.

Softmax Function

An activation function that converts a set of numerical values into probabilities, commonly used for multi-class classification tasks.

Sparse Representation

A data representation technique where most values are zero, improving storage efficiency and computational performance for machine learning models.

Specificity

A classification metric that measures a model's ability to correctly identify actual negative instances, also known as the true negative rate.

Static Inference

The process of generating predictions from a trained machine learning model without updating its parameters or learning from new data.

Stochastic Gradient Descent (SGD)

An optimization algorithm that updates model parameters using one or a few training samples at a time, improving training efficiency.

Suicide & Self-Harm

Content involving self-harm or suicide that requires specialized detection, safety interventions, and responsible AI handling.

Superintelligence (ASI)

A hypothetical form of AI that surpasses human intelligence across virtually every cognitive task, including reasoning, creativity, and problem-solving.

Supervised Learning

A machine learning approach where models learn from labeled data to predict outputs or classify new, unseen inputs.

Support Vector Machine (SVM)

A supervised machine learning algorithm that separates data into classes by finding the optimal decision boundary with the largest margin.

Surrogate Model

A simplified model used to approximate the behavior of a more complex model for optimization, interpretation, or computational efficiency.

Symbolic AI

An AI approach that relies on explicit rules, logic, and knowledge representations rather than learning patterns from large datasets.

Synthetic Data

Artificially generated data that replicates real-world patterns while reducing privacy risks and supporting AI training or testing.

Synthetic Data Generation

The process of creating artificial datasets that replicate the statistical properties of real data for training, testing, or privacy preservation.

Synthetic Feature

An artificially created feature derived from existing data to improve the predictive performance of a machine learning model.

System Prompt Override

An attack that attempts to ignore, replace, or manipulate a model's hidden system instructions to alter its intended behavior.

Tabular Data

Structured data organized into rows and columns, commonly stored in spreadsheets or databases for analysis and machine learning.

Take It Down Act

A U.S. law addressing the removal of non-consensual intimate imagery, including AI-generated deepfakes, and establishing related platform obligations.

TensorFlow

An open-source machine learning framework developed by Google for building, training, and deploying deep learning and AI models.

Test Set

A separate portion of a dataset used to evaluate the performance and generalization ability of a trained machine learning model.

Text Generation Inference

The process of using a trained language model to generate text responses or predictions from input prompts in real-time applications.

Threat Intelligence

The collection, analysis, and sharing of information about emerging cyber threats to improve AI and enterprise security defenses.

Token Smuggling

An attack technique that hides malicious instructions within encoded or fragmented tokens to bypass AI security filters.

Tokenization

The process of breaking text into smaller units called tokens that AI models use for training, processing, and generating language.

Toxicity Detection

The process of identifying abusive, offensive, hateful, or harmful language in AI inputs or outputs to enforce content safety policies.

Training Dataset

The portion of a dataset used to teach a machine learning model by allowing it to learn patterns and relationships from labeled or unlabeled data.

Training Loss

A metric that measures how well a machine learning model fits the training data by quantifying prediction errors during training.

Transfer Learning

A machine learning technique where knowledge gained from one task is reused to improve performance on a related task with less training data.

Transformer Models

A family of deep learning models based on the transformer architecture, widely used for natural language processing, vision, and multimodal AI tasks.

Transformer Neural Network

A neural network architecture that uses self-attention mechanisms to efficiently process sequential data such as text, images, and audio.

Triplet Loss

A loss function that trains models to distinguish similar and dissimilar data by minimizing the distance between related samples while maximizing separation from unrelated ones.

Trust & Safety

The policies, technologies, and operational practices used to protect users from harmful, abusive, or unsafe AI interactions.

Underfitting

A modeling problem where a machine learning model is too simple to capture underlying patterns, resulting in poor performance on both training and test data.

Unimodal AI

AI systems designed to process and understand a single type of data, such as text, images, audio, or video.

Unlabeled Data

Data that does not include predefined output labels or categories, commonly used in unsupervised and semi-supervised learning.

Unsupervised Learning

A machine learning approach where models identify patterns, structures, or relationships in unlabeled data without predefined target outputs.

US Executive Order on AI

A U.S. government directive establishing priorities and requirements for the safe, secure, trustworthy, and responsible development and use of artificial intelligence.

User Access Abuse

The misuse of legitimate user accounts or permissions to perform unauthorized actions against AI systems or connected resources.

Validation Loss

A metric that measures a model's prediction error on the validation dataset, helping assess generalization performance during training.

Validation Set

A portion of data used during model development to evaluate performance, tune hyperparameters, and prevent overfitting.

Variance

A statistical measure of how much data points or model predictions vary from their average, influencing model stability and generalization.

Variational Autoencoder (VAE)

A generative deep learning model that learns compact data representations and generates realistic synthetic data by sampling from a learned latent space.

VGG

A deep convolutional neural network architecture known for using small convolutional filters and achieving strong image classification performance.

Violent Extremism

Content promoting or supporting extremist ideologies, terrorist organizations, or acts of violence that AI systems should detect and restrict.

Vision-Based Injection

An attack that embeds malicious instructions within images or visual content to manipulate multimodal AI systems.

Weight

A learned parameter in a machine learning model that determines the importance of an input when generating predictions.

XGBoost

An optimized gradient boosting framework that builds highly accurate decision tree models for classification, regression, and ranking tasks.

YOLO (You Only Look Once)

A real-time object detection algorithm that identifies and locates multiple objects in a single pass through an image.

Z-Mean

The standardized mean of a dataset after Z-score normalization, typically equal to zero, indicating the data has been centered around its average.

Z-Score Normalization

A data normalization technique that transforms values by subtracting the mean and dividing by the standard deviation, producing standardized features.

Zero-Shot Learning

A machine learning technique where a model performs new tasks without task-specific training by leveraging previously learned knowledge.

Sanitized Prompt

A prompt that has been cleaned to remove sensitive information, malicious instructions, or unsafe content before it is sent to an AI model.

Grounding

The process of providing an AI model with relevant and trusted information so it can generate accurate, context-aware, and fact-based responses.

Indexing

The process of organizing data so it can be quickly searched and retrieved by AI applications when responding to user queries.

Inferencing

The process where a trained AI model analyzes new input and generates a prediction, decision, or response without further training.

Prompt

The input, instruction, or question given to an AI model that tells it what task to perform or what response to generate.

Response

The output generated by an AI model after processing a prompt, such as text, code, images, or other types of content.

Prompt Injection

An attack that uses malicious instructions to manipulate an AI model into ignoring its intended rules or performing unintended actions.

Sensitive Data

Confidential or private information that requires protection from unauthorized access, exposure, modification, or misuse.

API Key

A unique credential used to authenticate and authorize access to an API, application, or connected service.

Toxic Content

Content containing harmful, abusive, hateful, threatening, or offensive language that may violate safety or moderation policies.

Malicious URL

A web link designed to direct users or systems to harmful destinations associated with phishing, malware, scams, or other security threats.

Policy Violation

An action, input, or output that breaks predefined organizational, security, compliance, or AI usage policies.