AI Lexicon

The language of AI, explained.

AI conversations move fast, and the vocabulary moves with them. This is a working glossary of the 54 terms we hear most often in client conversations, from foundational concepts to the language of large language models, explained without the jargon.

ai · lexicon

An abstract close-up render of a glowing neural network, standing in for the concepts explained in this lexicon

Jump to a letterA B C D E F G H I J K L M N O P Q R S T U V W X Z

Agent (AI Agent): A system that uses a model to plan and take actions toward a goal, calling tools, APIs, or other software with limited human oversight.
AGI (Artificial General Intelligence): A hypothetical AI that can understand, learn, and perform any intellectual task a human can, as opposed to today's systems, which are specialised for particular tasks.
Algorithm: A defined set of rules or steps a computer follows to solve a problem or complete a task, the underlying logic that any model or program runs on.
Alignment (AI Alignment): The practice of designing and training AI systems so their behaviour matches human intentions and values, rather than optimising for an unintended proxy goal.
Attention Mechanism: A technique that lets a model weigh which parts of its input matter most for a given output, the core idea behind the Transformer architecture.

Backpropagation: The algorithm used to train neural networks, calculating how much each internal parameter contributed to an error and adjusting it to improve the next prediction.
Bias (Algorithmic Bias): Systematic errors in a model's output that unfairly favour or disadvantage particular groups, usually inherited from patterns in its training data.

Chain-of-Thought: A prompting technique where a model is asked to reason through intermediate steps before giving a final answer, often improving accuracy on complex tasks.
Chatbot: A conversational application, increasingly built on a large language model, designed to answer questions or complete tasks through natural language dialogue.
Chunking: The practice of splitting a document into smaller passages before generating embeddings for it, so a retrieval system can return the specific section that answers a query instead of the whole document.
Computer Vision: The field of AI focused on interpreting images and video, covering tasks like object detection, classification, and image generation.
Context Window: The amount of text, measured in tokens, a language model can consider at once when generating a response, including both the prompt and its own output so far.

Deep Learning: A subset of machine learning that uses multi-layered neural networks to learn patterns directly from large volumes of data, powering most modern AI systems.
Diffusion Model: A generative model that creates images or other data by learning to reverse a process of gradually adding noise, the technique behind most modern image generators.
Drift (Model Drift / Data Drift): The gradual decline in a model's accuracy as real-world inputs or behaviour shift away from the data it was trained on, one of the main reasons production AI systems need ongoing monitoring.

Embedding: A numerical representation of text, images, or other data that captures meaning in a form a model can compare and search over.
Ethics (AI Ethics): The study and practice of ensuring AI systems are developed and used fairly, safely, and transparently, covering issues like bias, privacy, and accountability.
Explainable AI (XAI): Techniques and tools that make a model's decisions understandable to humans, important for trust, debugging, and regulatory compliance.

Few-shot Learning: Prompting or training a model with only a small number of examples to guide it toward the desired behaviour, rather than a full retraining pass.
Fine-tuning: Further training a pretrained model on a smaller, specific dataset so it performs better on a particular task or domain.
Foundation Model: A large model trained on broad data that can be adapted, via fine-tuning or prompting, to many different downstream tasks.

Generative AI: AI systems that create new content, such as text, images, audio, or code, rather than only classifying or predicting from existing data.
GPU (Graphics Processing Unit): Specialised hardware that performs many calculations in parallel, making it far faster than a standard CPU for training and running AI models.
Graph Neural Network (GNN): A model architecture built to learn directly from graph-structured data, such as networks of connected people, products, or transactions, rather than the grid- or sequence-shaped data most other models expect.
GraphRAG: A retrieval-augmented generation approach that organises source material into a knowledge graph first, so a model can reason over explicit relationships between entities rather than retrieving isolated chunks of similar text.
Guardrails: Rules, filters, or checks placed around an AI system to keep its outputs and actions within safe, intended boundaries.

Hallucination: When a model generates information that is plausible-sounding but factually incorrect or entirely fabricated.
Human-in-the-Loop: A system design where a person reviews, approves, or corrects an AI's output before it takes effect, balancing automation with oversight.

Inference: The process of running a trained model on new input to produce an output, as distinct from training, when the model's parameters are being learned.

Jailbreak (Prompt Jailbreak): A prompt crafted to bypass a model's safety guardrails and produce content it would normally refuse to generate.

Knowledge Graph: A structured network of entities and their relationships, used to give AI systems factual, queryable context beyond what's in their training data.

Large Language Model (LLM): A neural network trained on vast amounts of text to predict and generate language, the technology behind tools like ChatGPT and Claude.

Machine Learning: The broader field of building systems that improve at a task by learning from data, rather than following only explicitly programmed rules.
Multimodal AI: A model that can understand or generate more than one type of data, such as text, images, and audio, within a single system.

Natural Language Processing (NLP): The field of AI concerned with understanding, interpreting, and generating human language.
Neural Network: A model architecture loosely inspired by the brain, made of layered nodes that learn to transform input into output through training.

Orchestration (Agent Orchestration): The coordination of multiple AI models, agents, or tools so they work together on a larger task or workflow.
Overfitting: When a model learns its training data too closely, including its noise and quirks, and performs poorly on new, unseen data as a result.

Parameter: An internal value a model learns during training; modern large language models have anywhere from millions to trillions of them.
Pretraining: The initial, broad training phase where a model learns general patterns from a large dataset before any task-specific fine-tuning.
Prompt Engineering: The practice of crafting inputs to an AI model to reliably get the output you want, an increasingly important skill for working with LLMs.

Quantization: A technique that reduces the precision of a model's parameters to shrink its size and speed up inference, usually with a small trade-off in accuracy.

RAG (Retrieval-Augmented Generation): An approach where a model retrieves relevant documents or data before generating a response, grounding its answers in real, current information.
Reinforcement Learning: A training method where a model learns by trial and error, receiving rewards or penalties for its actions to improve future decisions.

Supervised Learning: Training a model on labelled examples, where each input has a known correct output, so it learns to map one to the other.
Synthetic Data: Artificially generated data used to train or test models, often to supplement real data or protect privacy.

Token: A chunk of text, roughly a word or part of one, that a language model processes as its basic unit of input and output.
Transformer: The neural network architecture, built around the attention mechanism, that underlies virtually all modern large language models.

Unsupervised Learning: Training a model on unlabelled data, letting it find patterns or groupings on its own without a known correct answer to learn from.

Vector Database: A database optimised for storing and searching embeddings, commonly used to power RAG systems and semantic search.
Vision-Language Model: A model trained to understand both images and text together, able to answer questions about an image or generate captions for it.

Weights (Model Weights): The learned numerical values inside a neural network that determine how it transforms input into output; training is the process of adjusting them.

XGBoost: A widely used gradient-boosting algorithm for structured, tabular data, valued in industry for its accuracy and speed on classic prediction problems.

Zero-shot Learning: Asking a model to perform a task it has never explicitly been trained or shown examples for, relying on what it learned during pretraining.

Have a term we should add?

This lexicon grows as the field does. Tell us what's missing, or talk to us about what any of this means for your business.

Start a conversation