We all know what AI stands for—but if you are looking for a deeper dive into the technical side of artificial intelligence, this glossary decodes and serves as your tech speak translator for what’s behind everyday tools.
Agent (Autonomous / Software Agent)
A program that can perceive its environment, make decisions, and act without constant human direction (e.g., an e‑mail subject‑line optimizer that runs unattended).
Artificial Intelligence (AI)
The umbrella discipline focused on building systems that perform tasks normally requiring human intelligence—perception, language, reasoning, prediction, or autonomous action.
A/B (Multivariate) Testing
A statistical method that compares two or more variants (subject lines, web layouts, etc.) on a live audience to identify the highest‑performing version.
Bias (Algorithmic)
Systematic, unfair skew in predictions or outputs caused by non‑representative data or model design. Can manifest as exclusion, stereotyping, or disparate impact.
Chain‑of‑Thought (CoT)
A prompt‑engineering technique for large language models (LLMs) that asks the model to “show its work” step‑by‑step, often improving reasoning and transparency.
Context Window
The maximum number of tokens (roughly words or word‑pieces) an LLM can ingest at once. Limits how much text you can feed or retrieve in a single prompt.
Computer Vision (CV)
Field of AI that enables machines to “see” and interpret images or video—logo detection, event badge scanning, or visual QA of print collateral.
Data Lake / Warehouse
Centralized storage for structured and unstructured data, used to train models or feed dashboards. A lake holds raw files; a warehouse enforces schema for fast queries.
Deep Learning
Sub‑field of machine learning that uses multi‑layer neural networks to automatically learn high‑level abstractions from data (images, text, audio).
Differential Privacy
A mathematical framework that injects statistical noise so insights about groups can be shared without revealing any individual’s data.
Embedding
Numeric vector representation of text, images, or users that capture semantic meaning. Powers semantic search, recommendations, and clustering.
Explainability / Interpretability
Techniques that make model decisions understandable to humans (e.g., SHAP values showing which features drove a churn prediction).
Few‑Shot / Zero‑Shot Learning
LLM ability to perform tasks with a handful (few) or no (zero) specific examples by leveraging prior pre‑training.
Fine‑Tuning
Additional targeted training of a pre‑trained model on domain‑specific data to improve performance (e.g., an LLM fine‑tuned on association bylaws).
Foundation Model
Large, pre‑trained model (text, image, code, multimodal) that can be adapted to many downstream tasks—LLMs, vision transformers, etc.
Generative AI
Models that create new content—text, images, audio, code—rather than simply classifying or predicting. Includes LLMs and diffusion models.
Guardrail
Policy or technical control that confines model behavior—rate limits, content filters, human‑approval steps.
Hallucination
Confident but factually incorrect output from a generative model (e.g., invented citation). Requires validation layers.
Human‑in‑the‑Loop (HITL)
Workflow where humans review, correct, or override AI actions to ensure quality and accountability.
Inference
The act of running a trained model to generate predictions or content in production (as opposed to training, which learns parameters).
Large Language Model (LLM)
Transformer‑based neural network with billions of parameters trained on massive text corpora to perform language tasks—chatting, summarizing, translating.
Machine Learning (ML)
Subset of AI where algorithms learn patterns from data instead of being hard‑coded with rules. Includes supervised, unsupervised, and reinforcement learning.
Model Drift
Degradation of model accuracy over time due to changes in data, business process, or environment. Requires monitoring and periodic retraining.
MLOps
Engineering practice that applies DevOps principles to ML: versioning data & models, continuous integration / deployment, automated testing, monitoring.
Multimodal
Models or applications that accept and/or generate multiple data types (text + image + audio), enabling richer interactions (e.g., a slide generator from text prompts and brand images).
Natural Language Processing (NLP)
Field that enables computers to understand and manipulate human language—sentiment analysis, entity extraction, summarization.
Natural Language Generation (NLG)
Sub‑field of NLP focused on producing new human‑like text (e.g., automated report writing). Gen‑AI advances have super‑charged this area.
Prompt Engineering
Crafting and refining instructions, examples, or constraints fed to an LLM to elicit the desired output.
Recommender System
Algorithm that suggests the next best product, article, or action based on user behaviour and item similarity (e.g., suggesting webinars to members).
Reinforcement Learning (RL)
Training paradigm where an agent learns by receiving rewards or penalties from its environment, suitable for pricing, scheduling, or ad bidding.
Reinforcement Learning from Human Feedback (RLHF)
Fine‑tuning method that aligns model outputs with human preferences by training on human‑rated examples.
Retrieval‑Augmented Generation (RAG)
Architecture that first retrieves relevant documents from a knowledge base, then feeds them into a generative model to ground answers in verified facts.
Robotic Process Automation (RPA)
Software “bots” that mimic keyboard & mouse actions to automate repetitive desktop tasks—data entry, invoice matching, CRM updates.
Sentiment Analysis
NLP technique that classifies text as positive, negative, or neutral, often with emotion nuance—useful for social listening and member feedback.
Supervised / Unsupervised Learning
- Supervised: model learns from labelled examples (input → target).
- Unsupervised: model finds patterns without labels (clustering, anomaly detection).
Synthetic Data
Artificially generated data that mimics real distributions, useful for training when actual data is scarce or sensitive.
Token (LLM Tokenization)
The smallest unit of text an LLM processes—roughly a word or word‑piece. Prompt length and pricing are often measured in tokens.
Transformer
Deep‑learning architecture based on self‑attention, underpinning LLMs and many modern vision models.
Vector Database
Specialized store optimized for similarity search on embeddings; common backbone for semantic search and RAG.
Workflow Orchestration
Coordinating tasks, data, and approvals across systems—used to stitch AI
