Feb 9, 2026
Lizzy Herzer
Executives face the challenge of evaluating AI initiatives and training their teams without having the necessary technical vocabulary themselves. Complex AI terms are often used inconsistently or misunderstood, leading to project misunderstandings and ineffective decisions.
Our practice-oriented AI glossary explains the most important artificial intelligence terms in an understandable and management-relevant way, so you can make informed decisions and successfully lead your team through the AI transformation.
Download our free AI glossary as a PDF and share it with your team for a common knowledge base.
A/B Testing - Method for comparing different model versions in real applications through systematic testing with different user groups.
Accuracy - Accuracy metric that measures the proportion of correctly classified predictions to the total number of predictions.
Adversarial Learning - A training approach where models are made robust against specially developed "attacks" or disruptive inputs. The goal is to improve the security and reliability of AI systems.
Agentive (Agents) - Systems or models that exhibit their own agency to autonomously pursue goals. An agentive model can act without constant supervision, such as a highly autonomous vehicle.
AGI (Artificial General Intelligence) - A hypothetical form of AI that reaches or exceeds human cognitive abilities in all areas and can develop independently. AGI does not currently exist.
AI (Artificial Intelligence) - A subfield of computer science that deals with developing systems that can perform tasks that normally require human intelligence, such as learning, problem-solving, and pattern recognition.
AI Ethics - Principles and guidelines aimed at preventing harm to humans through AI. Includes topics such as fairness, transparency, accountability, and responsible handling of data and algorithms.
AI Governance - A structured framework of policies, processes, and controls for the responsible development, implementation, and monitoring of AI systems in organizations.
AI Management System - Comprehensive platform for central management and orchestration of AI projects throughout their entire lifecycle. An AI Management System coordinates data flows, model development, deployment and monitoring, while simultaneously enforcing governance policies and ensuring compliance. E.g. basebox
AI Model - A mathematical model that has been trained through machine learning to perform specific tasks.
AI Safety - An interdisciplinary research field that deals with the risks and long-term impacts of AI. Examines both short-term problems like bias and long-term scenarios like the possible development of superintelligence.
AI Sovereignty - The ability of an organization or state to maintain autonomous control over their AI systems, data, and digital infrastructures without being dependent on external providers.
AI Stack - Technological architecture of interconnected software layers that together form a complete AI development environment. The AI Stack typically includes the data layer (preprocessing, feature engineering), framework layer (TensorFlow, PyTorch), model management layer, API layer for deployment and the application layer - an integrated environment that seamlessly orchestrates these components and provides a unified platform for the entire ML workflow. E.g. basebox
AI System - A complete system that includes AI models, data processing, user interfaces, and other components.
Air Gapped - A security concept where AI systems are physically separated from external networks to prevent unauthorized access or data leakage.
Algorithm - A series of mathematical instructions and rules that enable a computer to solve specific problems or perform tasks. In AI, algorithms form the basis for learning from data.
Algorithmic Accountability - Responsibility for the decisions and impacts of algorithms, including transparency about their functioning and accountability for their results.
Alignment - Adjusting an AI so that its goals and behavior align with human values and intentions. This includes both technical aspects like content moderation and philosophical questions about desired outcomes.
Anthropomorphism - The human tendency to attribute human characteristics, emotions, or consciousness to non-human objects. In AI, this often leads to people perceiving chatbots as sentient.
API (Application Programming Interface) - Interfaces through which various software applications can communicate with AI models.
Attention Mechanism - A technique in neural networks that allows models to focus on relevant parts of input data rather than weighting all information equally.
Autoencoder - Neural networks that compress and reconstruct data, often used for dimensionality reduction or anomaly detection.
Automation/RPA (Robotic Process Automation) - Technology for automating repetitive, rule-based business processes through software "robots" that mimic human actions in digital systems. RPA automates structured tasks like data entry, report generation, or email processing without changing existing IT systems. Combined with AI (Intelligent Process Automation), unstructured data can also be processed and more complex decisions made. RPA offers rapid implementation, high accuracy, and 24/7 availability with low initial investments.
Autonomous Agents - AI systems that can act independently in their environment to achieve specific goals. They have sensors for perception, decision algorithms, and actuators for executing actions.
Backpropagation - The algorithm by which neural networks learn by propagating errors backward through the network.
Base Models - See Foundation Models.
Benchmark - Standardized tests for evaluating and comparing AI models based on uniform criteria.
BERT - Bidirectional Encoder Representations from Transformers; influential language model from Google that enables bidirectional context processing.
Bias - Systematic distortions in AI models that can lead to unfair or discriminatory results. Usually arise from unbalanced or prejudiced training data.
Bias Mitigation - Strategies and techniques for identifying, measuring, and reducing prejudices in AI systems to achieve fairer results.
Big Data - Extremely large and complex datasets that require special technologies and methods for processing, storage, and analysis. Often characterized by the "5 V's": Volume, Velocity, Variety, Veracity, and Value.
Certification - Formal confirmation that AI systems meet certain standards or requirements.
Chain of Thought - Technique where AI models explicitly present their thinking steps to develop comprehensible problem solutions.
Chatbot - A computer program that can conduct natural language conversations with humans, typically through text or voice interfaces.
ChatGPT - A conversational AI assistant developed by OpenAI that is based on the GPT architecture (Generative Pre-trained Transformer) and optimized for dialogue-oriented applications.
Cloud - A model for providing IT resources over the internet that enables flexible and scalable use of AI services and infrastructures.
Clustering - An unsupervised learning method that automatically organizes similar data points into groups (clusters) without prior knowledge of which groups should exist. Clustering algorithms like K-Means or hierarchical clustering find hidden patterns and structures in data. Applications include customer segmentation, market research, anomaly detection, and recommendation systems. The algorithm independently determines which data points are similar and groups them accordingly. This enables insights into natural data structures without human specifications.
Cognitive Computing - An approach to AI that aims to simulate human thought processes, including learning, reasoning, and self-correction.
Computer Vision (CV) - The field of AI that deals with the interpretation and analysis of visual data.
Constitutional AI - Approach to developing AI systems with built-in ethical principles and values.
Continual Learning - Ability of models to continuously learn new tasks without forgetting old knowledge.
Conversational AI - AI systems specialized in conducting natural conversations with humans, often through chatbots or voice assistants.
Convolutional Neural Networks (CNNs) - A specialized architecture of neural networks that is particularly effective for processing image data. CNNs use convolution operations to recognize local patterns and features in images.
Cross-Validation - A method for evaluating model performance through systematic division of data into training and test sets.
Data - The foundation for AI systems; structured or unstructured information used for training and operating algorithms.
Data Augmentation - Techniques for artificially expanding training datasets through transformations of existing data (e.g., rotation, scaling of images) to improve model performance.
Data Governance - Policies and processes for data management in AI projects.
Data Labeling - Process of annotating data with correct answers or categories for Supervised Learning.
Data Mining - The process of discovering patterns and insights in large datasets using statistical and machine learning methods.
Data Pipeline - Automated sequence of data processing steps from source to model.
Data Protection and Data Security (Data Privacy) - Protection of personal data in AI systems.
Decision Trees - An interpretable machine learning model that represents decision rules in a tree-like structure. Each node represents a decision based on a feature, leading to comprehensible decision paths.
Deep Learning - A subfield of machine learning that uses deep neural networks with many layers to recognize and learn complex patterns in large amounts of data.
Diffusion Models - A class of generative models that learn to generate data by reversing a process that gradually adds noise to data. Particularly successful in image generation.
Dimensionality Reduction - Techniques for reducing the number of features in datasets to decrease complexity and enable visualization.
Dropout - Regularization technique in neural networks to prevent overfitting by randomly deactivating neurons during training.
Edge AI - AI processing directly on end devices instead of in the cloud.
Embedding - Numerical representations of words, images, or other data in a multi-dimensional space that capture semantic relationships.
Emergent Behavior - Unexpected abilities or behaviors that occur in AI models without being explicitly programmed or trained. Arise through complex interactions in large models.
End-to-End Encryption - A security procedure where data remains encrypted throughout transmission and can only be decrypted by intended recipients.
End-to-End Learning - An approach in deep learning where a model learns a complete task from raw input to final output in a single, integrated system.
Ensemble Methods - Combination of multiple models for better predictions.
Epoch - A complete pass through all training data during model training.
Ethical Considerations - The systematic consideration of moral and societal impacts of AI systems, including privacy, fairness, and potential misuse.
EU AI Act - The EU regulation for regulating AI systems that defines risk categories and establishes corresponding requirements.
Explainable AI (XAI) - AI systems and methods that can make their decision processes transparent and understandable to humans.
F1-Score - Harmonic mean of Precision and Recall for balanced evaluation of classification models.
Fairness - Ensuring that AI systems treat all groups equally.
Feature Engineering - Process of selecting and transforming input features for better model performance.
Federated Learning - Distributed learning where models are trained on decentralized data without needing to centrally collect raw data.
Few-Shot Learning - The ability of a model to learn new tasks with only a few examples.
Fine-Tuning - The process of adapting a pre-trained model to specific tasks.
Foundation Models - Large, versatile models that can serve as a basis for various applications.
GDPR - The EU General Data Protection Regulation that regulates the protection of personal data and has important implications for AI systems.
Generative Adversarial Networks (GANs) - An architecture of two competing neural networks: a generator that creates new data and a discriminator that distinguishes real from generated data.
Generative AI - AI systems that can create new content like text, images, audio, or code based on patterns they learned from training data.
Google Gemini - Google's multimodal AI model that can process text, images, and other data types.
GPT - Generative Pre-trained Transformer; a family of language models from OpenAI based on the Transformer architecture.
GPU - Graphics Processing Unit; specialized processors that are particularly efficient for training and running AI models.
Gradient Boosting - Ensemble method that sequentially combines weak learners to create strong predictive models.
Gradient Descent - An optimization algorithm used to improve model parameters.
Grounding/Ground Truth - Verified, correct reference data that serves as a standard for evaluating AI models.
Guardrails - Safety mechanisms and guidelines implemented in AI systems to prevent harmful, unethical, or unwanted outputs.
Hallucination - The phenomenon where AI models (especially language models) generate information that is factually incorrect but appears convincing and plausible.
Hosting - The provision of infrastructure and services for operating AI applications.
Human-in-the-Loop - Approach where humans remain involved in the AI decision process to ensure quality and control.
Hybrid AI - Systems that combine different AI approaches or connect AI with traditional programming methods.
Hyperparameter - Configuration settings that are set before training and control learning behavior.
In-Context Learning - Learning through examples in the prompt without retraining the model.
Inference - The process where a trained AI model generates predictions or outputs for new, unseen input data.
Integration - The process of incorporating AI systems into existing IT infrastructures and business processes.
Interpretability - The ability to understand and explain AI decisions.
IoT - Internet of Things - network of physical objects with embedded sensors and connectivity, often in connection with AI for intelligent automation.
Language Models - AI models specifically developed for processing and generating natural language.
Llama - A family of open-source language models from Meta (formerly Facebook).
LLM (Large Language Model) - Large neural networks trained on extensive text data to understand and generate human language. Examples are GPT, BERT, and LaMDA.
Lock-In - The dependence on a specific vendor or technology that makes switching to alternatives difficult.
Loss Function - A mathematical function that measures how far a model's predictions deviate from actual results.
LSTM (Long Short-Term Memory) - Special RNN architecture for processing long sequences that solves the vanishing gradient problem.
Machine Learning (ML) - A subfield of AI that uses algorithms and statistical models to enable computers to learn from data without being explicitly programmed for each task.
MCP - Model Context Protocol - a standard for communication between AI models and external data sources.
Microsoft Bing - Microsoft's search engine that has integrated AI technologies (especially from OpenAI) to provide enhanced search functions and conversational responses.
Mistral - A French AI company that develops open-source language models.
Mixture of Experts (MoE) - Architecture that combines different specialized models, where a gating network decides which experts are activated for a specific input.
MLOps - Practices for deploying and maintaining machine learning systems in production.
Model Deployment - The process of implementing trained models in productive environments.
Model Drift - Deterioration of model performance over time due to changing data or environmental conditions.
Multiagent Systems - Systems with multiple interacting AI agents that collaborate or compete.
Multimodal AI - AI systems that can simultaneously process and understand different types of input data (text, images, audio, video).
MVP (Minimum Viable Product) - The simplest version of an AI application that contains just enough features to provide real value to users and collect feedback. An MVP in AI projects focuses on core functionality and is developed quickly to obtain early user feedback. It contains only essential features and is iteratively improved based on user experiences. This enables validation of market needs and minimization of development risks before extensive resources are invested.
Neural Architecture Search (NAS) - Automated search for optimal neural network architectures.
Neural Networks (NN) - A computational model inspired by the structure of the human brain consisting of interconnected nodes (neurons) that can recognize and learn patterns in data.
NLP (Natural Language Processing) - A subfield of AI that deals with the interaction between computers and human language, including understanding, interpreting, and generating natural language.
Normalization - Scaling data to uniform ranges for better training and comparability.
Object Detection - Computer vision task for recognizing and localizing objects in images.
On-Prem (On Premises) - IT infrastructure operated locally in a company's own facilities, as opposed to cloud-based solutions.
Open Source - Software whose source code is publicly accessible and can be freely modified and distributed.
Optimization - Process of improving model parameters to minimize errors or maximize performance.
Overfitting - A problem in machine learning where a model is too specifically adapted to training data and therefore generalizes poorly to new, unseen data.
Parameters - Numerical values in a model that are adjusted during training and determine the model's behavior and predictions.
Perplexity - Both a metric for evaluating language models (measures uncertainty in predictions) and an AI-powered search engine that cites sources for its answers.
PoC (Proof of Concept) - A small, experimental project to demonstrate the feasibility of an AI solution before making larger investments. A PoC tests the basic functionality of an AI application with limited resources and data. The goal is to identify technical risks, determine initial performance metrics, and convince stakeholders of viability. PoCs typically last 4-12 weeks and use a small dataset. They help decide whether a project moves to the pilot phase or full development.
Precision and Recall - Metrics for evaluating the quality of classification models. Precision measures the accuracy of positive predictions, Recall measures the completeness of recognized positive cases.
Predictive Analytics - Use of data and algorithms to predict future events.
Preprocessing - Preparation and cleaning of raw data before training or analysis.
Privacy - The protection of personal information and control over how this data is collected and used.
Prompt - The input (text, question, or instruction) given to an AI model to obtain a desired response or output.
Prompt Chaining - A technique where multiple consecutive prompts are used, with the output of one prompt serving as input for the next.
Prompt Engineering - The art and science of optimally formulating inputs for AI models to achieve desired results.
Quantization - Technique for reducing model size by decreasing numerical precision to save memory and computational effort.
RAG (Retrieval-Augmented Generation) - An advanced AI technique that incorporates external knowledge sources into text generation. RAG combines the capabilities of language models with a search component that retrieves relevant information from a knowledge database. The process works in three steps: 1) Retrieval - searching for relevant documents or information based on the query, 2) Augmentation - enriching the original prompt with the found information, 3) Generation - creating a response based on the extended context. This enables AI systems to give more current and factually correct answers by accessing external, verified knowledge sources instead of relying only on their training data.
Random Forest - Ensemble method that combines many decision trees to make robust predictions.
Recommendation Systems - Algorithms that provide personalized recommendations based on user behavior.
Recurrent Neural Networks (RNNs) - A neural network architecture specifically developed for processing sequential data.
Regularization - Techniques to prevent overfitting by constraining model complexity.
Reinforcement Learning - A learning method where an agent learns through interaction with an environment, receiving rewards for desired actions.
RLHF (Reinforcement Learning from Human Feedback) - Training models based on human feedback to improve output quality.
Robotics - Application of AI in physical robot systems for autonomous movement and manipulation.
ROI (Return on Investment) - A metric for evaluating the profitability of AI investments that measures the ratio between achieved benefits and invested costs. ROI in AI projects is calculated as (Benefits - Costs) / Costs × 100%. Benefits can arise from cost savings, revenue increases, or efficiency gains. In AI projects, ROI calculation is often complex because advantages like improved decision-making or customer experience are difficult to quantify. Typical ROI drivers are automation of manual processes, better predictions, and personalized customer approaches.
Scaling - The ability of a system to handle growing requirements, both in terms of data volume and user numbers.
Security by Design - An approach where security aspects are integrated into AI system development from the beginning.
Semantic Segmentation - Computer vision task for pixel-precise classification of image content.
Sentiment Analysis - NLP task for determining emotional evaluations or opinions in texts.
Sovereignty (Digital/AI Sovereignty) - The ability of an organization or state to maintain autonomous control over their AI systems, data, and digital infrastructures without being dependent on external providers.
Speech Recognition - Technology for converting spoken language into text.
Speech-to-Text - See Speech Recognition.
Stochastic Parrot - A critical term that suggests language models may only reproduce statistical patterns without true understanding.
Style Transfer - A technique that transfers the artistic style of one image to another.
Supervised Learning - A learning method where models are trained on labeled data where both input and desired output are known.
Support Vector Machine (SVM) - Classic ML algorithm for classification and regression that finds optimal separating hyperplanes.
Swarm Intelligence - An approach where the collective behavior of simple agents enables complex problem solutions.
Synthetic Data - Artificially generated data to supplement or replace real training data.
Temperature - A parameter in language models that controls the randomness and creativity of output.
Text-to-Image Generation - The ability of AI systems to create new images based on text descriptions.
Time Series Analysis - Analysis of temporally ordered data for pattern recognition and prediction.
Token - The smallest processing units in language models. A token can be a word, word fragment, or character.
Training Data - The datasets used to train AI models.
Transfer Learning - Use of knowledge from one domain for tasks in another domain.
Transformer Model - A neural network architecture based on attention mechanisms that is particularly successful in language processing.
Transparency - Disclosure of how AI systems function and make decisions.
Turing Test - A test proposed by Alan Turing for evaluating machine intelligence.
Unsupervised Learning - A learning method where models must find patterns in data without given labels.
Use Cases - Specific application scenarios or business cases where AI technologies are deployed.
Validation Set - Dataset for evaluating model performance during development, separate from training and test data.
Vendor Risk - The risks associated with dependence on external providers for AI services.
Weak AI (Narrow AI) - AI systems developed for specific, limited tasks.
Weight - Parameters in neural networks that determine the strength of connections between neurons.
Zero-Shot Learning - The ability of a model to perform tasks it was not explicitly trained for, based on its general understanding and prior knowledge.
Are you missing an important term? We gladly accept your suggestions.
Stay Up to Date