AI Glossary

Software that learns patterns from large amounts of data and uses those patterns to make predictions, generate content, or take actions — without being explicitly programmed with rules for every situation. The key word is learned: AI adapts from examples, not from a rulebook.

Example: A spam filter that improves over time by seeing millions of emails is AI. A rule that blocks all emails containing "WINNER" is not.

A step-by-step set of instructions a computer follows to solve a problem or complete a task. In AI, algorithms are the mathematical processes used to learn from data and produce outputs. Not all algorithms are AI — a recipe is an algorithm; so is the logic that sorts your inbox by date.

Using technology to perform tasks with minimal or no human involvement. Automation and AI are often confused, but they are not the same thing. A calendar reminder that fires at 9am every morning is automation. An AI would learn when you tend to need reminders and suggest them proactively.

Rule of thumb: Automation follows fixed rules you set. AI learns patterns and adapts.

When an AI system produces results that systematically favour or disadvantage certain groups of people. AI bias is almost always caused by patterns in training data that reflect historical inequalities — the AI learns to repeat them, often without anyone intending it to discriminate.

Example: A credit-scoring AI trained on historical South African lending data may learn that certain postal codes predict loan default — because apartheid-era exclusion from formal credit markets produced those patterns in the data.

A software program designed to simulate conversation with humans. Older chatbots followed rigid decision trees ("Press 1 for billing"). Modern AI chatbots — like the ones built on large language models — generate responses dynamically and can handle a much wider range of inputs. The two types are often confused because they look similar.

A branch of AI that enables computers to interpret and understand images and video. Computer vision systems are trained on millions of labelled images until they can identify objects, faces, text, or anomalies with high accuracy — in some cases, higher than humans.

Examples: Face ID, security camera monitoring, medical imaging analysis (detecting tumours on X-rays), document scanning and extraction.

Known limitation: Performance often drops significantly for faces and scenarios underrepresented in training data — many systems have lower accuracy on darker skin tones because they were trained primarily on lighter-skinned faces.

In AI contexts, data refers to the raw material an AI system learns from — text, images, numbers, audio, or records that contain patterns the model is trained to recognise. The quality, quantity, and representativeness of data directly determines what the AI learns and how it performs.

Key question to ask about any AI system: Where did the training data come from, and who is or isn't represented in it?

A type of machine learning that uses neural networks with many layers ("deep" layers) to identify complex patterns in large datasets. Most modern AI — including image recognition, speech recognition, and large language models — is built on deep learning. The "depth" refers to the number of processing layers, not complexity of understanding.

AI that creates new content — text, images, audio, code, or video — by predicting what should come next based on patterns learned from training data. It does not retrieve from a database of facts; it generates outputs that are statistically consistent with what it has seen. This is why it can be fluent and wrong at the same time.

Examples: ChatGPT, Claude, Gemini, Microsoft Copilot, Midjourney (images), ElevenLabs (audio).

When a generative AI produces information that sounds plausible and confident but is factually incorrect or entirely fabricated. Hallucination is a structural property of how LLMs work — they generate statistically likely sequences of text, not verified facts. The AI cannot tell when it is wrong, and its tone gives no indication of accuracy.

Common examples: Inventing citations and journal articles that don't exist, fabricating statistics, stating incorrect historical dates with complete confidence.

A type of AI trained on vast quantities of text to understand and generate human language. "Large" refers to the scale of both training data and the model's internal parameters — modern LLMs are trained on a significant portion of publicly available text on the internet. LLMs are the technology behind ChatGPT, Claude, Gemini, and similar tools.

What they can do well: Drafting, summarising, translating, brainstorming, explaining concepts.
What they cannot do: Retrieve verified facts, know what has happened after their training cutoff, or reliably self-correct.

A branch of AI where systems improve their performance by learning from data, rather than being given explicit rules. Instead of a programmer writing "if postal code = X, then flag as high risk," a machine learning system is shown thousands of past decisions and figures out the patterns itself. Deep learning is a subset of machine learning.

The mathematical structure produced when an AI system is trained on data. Think of it as the "trained brain" of the AI — all of the learned patterns, weights, and relationships compressed into a deployable file. When a company says they are "running a model," they mean they are using this trained structure to make predictions or generate outputs on new inputs.

A computing architecture loosely inspired by the structure of the human brain, made up of layers of interconnected nodes ("neurons") that process and pass information. Neural networks are the foundation of most modern AI. The analogy to the brain is loose — neurons in AI are mathematical functions, not biological cells, and the network has no understanding of what it is processing.

A numerical value inside an AI model that gets adjusted during training to make the model's outputs more accurate. Parameters are how a model "stores" what it has learned. The number of parameters in a model is often used as a shorthand for its scale and capability — GPT-4 reportedly has over a trillion parameters.

AI that analyses an input and outputs a score, classification, or decision. Rather than generating new content, predictive AI assesses whether something falls into a category or predicts a likely outcome. It is particularly common in high-stakes decisions where organisations need to process large volumes of cases consistently.

Examples: Credit scoring, fraud detection, CV shortlisting systems, spam filtering, disease risk screening, insurance pricing models.

Key risk: Predictive AI inherits bias from its training data. If past decisions were discriminatory, the model learns to replicate them at scale.

The text input you give to a generative AI system. The quality, specificity, and framing of your prompt significantly affects the quality and relevance of the AI's response. A vague prompt produces a generic response; a well-structured prompt with context, constraints, and a clear goal produces a much more useful one.

Weak prompt: "Write an email about leave."
Strong prompt: "Write a brief, professional email to my team informing them that the office will be closed on 16 June for Youth Day. Tone: friendly but clear. Max 3 sentences."

The practice of crafting prompts deliberately to get more accurate, relevant, or useful outputs from AI systems. Prompt engineering is not about knowing magic words — it is about giving the AI the right context, constraints, and role to work within. It is an emerging professional skill across most sectors.

A technique that connects a generative AI model to a specific, controlled document set or database. When a query comes in, the system first retrieves relevant source documents, then gives those to the LLM to generate a response grounded in that content. This reduces hallucination compared to a standard LLM responding from memory alone.

Example: A company's internal HR chatbot that answers questions by first retrieving sections from the official policy documents, rather than generating answers from general training data.

The practice of designing, deploying, and using AI in ways that are ethical, fair, transparent, and accountable. Responsible AI asks not just "does this work?" but "who does this harm?", "can we explain how it decides?", and "who is accountable when it goes wrong?" It is both a professional practice and, increasingly, a regulatory requirement.

The basic unit of text that a language model processes — roughly equivalent to a word or part of a word. "Johannesburg" is one token; "un-be-liev-able" might be four. AI models have a "context window" measured in tokens, which limits how much text they can process in one go. Context windows have grown significantly — modern models can process hundreds of thousands of tokens.

The dataset used to teach an AI model. The patterns, gaps, biases, and representations present in training data directly shape what the model learns and how it performs on new inputs. Training data is the single biggest factor in AI quality — and the single biggest source of AI failure. Understanding where training data comes from is fundamental to understanding what any AI system can and cannot do.

Critical question: Is the group of people you are making decisions about fairly represented in the training data this AI was built on?