AI Glossary

The capacity of an AI system to take actions independently in pursuit of a goal — not just respond to a single prompt, but plan, decide, and act across a sequence of steps without human approval at each one. An AI system with high agency can browse the web, send emails, run code, and chain multiple actions together on your behalf. As AI systems gain more agency, questions of oversight and control become increasingly urgent.

Example: An AI agent given the task "book me a flight to Cape Town next Thursday under R5,000" that independently searches flight sites, compares options, and completes the booking — without prompting you at each step.

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.

The study and practice of ensuring AI systems are developed and used in ways that are fair, transparent, accountable, and aligned with human values. AI ethics is not just academic philosophy — it is increasingly a professional and regulatory discipline. It asks: who benefits from this AI, who might be harmed by it, who is responsible when it goes wrong, and whether the people it affects can understand or contest its decisions.

Key questions AI ethics asks: Is the system's decision-making explainable? Does it produce fair outcomes across different groups? Who bears accountability when it causes harm? Can affected people challenge the outcome?

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.

Technology — including AI — embedded invisibly into the physical environment, rather than accessed through a screen or an explicit interface. Ambient AI uses sensors, microphones, cameras, and connectivity built into rooms, buildings, vehicles, or wearables to monitor and respond to what is happening around you, often without you needing to initiate an interaction.

Examples: Smart home systems that adjust lighting and temperature based on who enters a room. Retail stores that track shopper movement without a traditional checkout. Hospital rooms that detect patient falls via ceiling cameras.

Privacy note: Because ambient technology is passive and often invisible, consent and transparency are significant ethical concerns — people may not be aware they are being monitored or that AI is acting on what it detects.

An American AI safety company founded in 2021 that created Claude, one of the leading large language models. Anthropic was founded by Dario Amodei, Daniela Amodei, and other former members of OpenAI, with a stated focus on AI safety research and interpretability — understanding how and why AI systems produce the outputs they do. Its name comes from "anthropic," meaning relating to human existence.

Relevance: Claude is one of the AI tools referenced across the AfriversalAI curriculum. Anthropic publishes research on AI safety and responsible deployment practices.

The use of AI to enhance human capabilities rather than replace them. Unlike full automation, augmentation keeps a human in the decision loop — the AI provides suggestions, analysis, or information, while the human applies judgment and takes action. Augmentation is often contrasted with automation to describe a model where AI makes workers more capable, not redundant.

Example: A doctor using AI to analyse X-rays receives highlighted areas of concern — but still makes the diagnosis. A lawyer using AI to surface relevant case law still drafts the argument. The AI augments; the human decides.

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.

A decision made solely or significantly by an algorithm, without meaningful human review of the individual case before the outcome is finalised. Under POPIA Section 71, data subjects in South Africa have two rights when an automated decision affects them: (1) the right to request that a human reviews the decision before it is implemented; and (2) the right to a reasoned explanation of the specific factors used to reach the outcome.

Example: An applicant tracking system that screens and rejects a job application without any human reading the CV constitutes an automated decision. The applicant can invoke POPIA s71 to request both human review and an explanation of which factors led to the rejection.

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.

Extremely large and complex datasets that traditional computing tools cannot process efficiently. Big data is often described by the "three Vs": Volume (enormous amounts), Velocity (generated at high speed), and Variety (structured numbers alongside unstructured text, images, and audio). Most modern AI depends on big data for training — the scale of available training data is one of the primary reasons current AI capabilities were not possible a decade ago.

Examples: Every transaction across a major bank's network in one day, all GPS pings from ride-hailing apps across a city, or all social media posts globally in a single hour.

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.

AfriversalAI's five-question judgment framework for evaluating any AI situation in a professional context. Introduced in Module 1 and applied throughout the course, the Funda Five™ gives every learner a consistent, repeatable decision process that requires no technical background. The name draws on the isiZulu and Sesotho word funda, meaning "to learn."

The five questions are: 1. Task — Is this the right kind of job for AI? 2. Data — Is any of this sensitive or POPIA-protected? 3. Tool — Is this AI type right for the task and approved by my organisation? 4. Trust — How would I detect an error in this output? 5. Human — Who checks this, and who is accountable if it goes wrong?

See also: The Funda Five™ page for the full interactive framework. Related terms: Responsible AI, AI Ethics.

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).

A type of computer chip originally designed for rendering graphics, now fundamental to AI training and inference. GPUs can perform thousands of mathematical operations simultaneously in parallel, making them far faster than standard CPUs for the matrix calculations at the core of neural network training. Access to sufficient GPU power is a major bottleneck in AI development — which is why cloud AI costs are high and why GPU supply chains have become geopolitically significant.

Scale reference: Training a large language model can require thousands of GPUs running continuously for weeks. Consumer GPUs can run smaller AI models locally — which is part of why the gap between what large companies and individuals can build is so wide.

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.

The South African legal term for a rule or practice that appears neutral but produces systematically worse outcomes for people with a protected characteristic. Prohibited under PEPUDA (Act 4 of 2000) and the Employment Equity Act Section 6 in employment contexts. When an AI system uses a proxy variable — like postal code — that correlates with race, disability, or gender due to historical inequality, the outcome may constitute indirect discrimination even when the system was not designed to discriminate.

Note: US law uses the term “disparate impact” for the same concept. SA courts and statutes use “indirect discrimination.” Learners will encounter both terms in global AI ethics literature — they refer to the same structural harm.

See also: Bias (AI Bias), Proxy Discrimination, PEPUDA.

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 property of compression where some information is permanently discarded to reduce file size or complexity, with the result that the decompressed version is an approximation of the original — not an exact copy. In AI, the concept applies to how models compress knowledge from training data. An LLM doesn't memorise every source it was trained on; it compresses patterns across billions of documents. What specific facts get "lost" in that compression is unpredictable, which is one structural reason models hallucinate.

Everyday examples: A JPEG photo saved at low quality loses fine detail that cannot be recovered when you zoom in. An MP3 discards audio frequencies the human ear is least sensitive to. Both are lossy — the file is smaller, but some of the original is gone permanently.

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.

An American AI research company founded in 2015 that created ChatGPT, GPT-4, and the image model DALL-E. OpenAI began as a nonprofit with the stated goal of safely developing artificial general intelligence (AGI), then adopted a "capped-profit" structure in 2019 to attract investment. It is the company most responsible for bringing generative AI into mainstream workplaces globally, primarily through the November 2022 launch of ChatGPT — which reached 100 million users in two months, faster than any consumer product in history at the time.

Relevance: ChatGPT is the most widely recognised generative AI tool in South African workplaces. Many third-party AI applications across sectors are built on OpenAI's GPT models via its API. See also: Anthropic.

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.

A type of machine learning where an AI system learns by trial and error — taking actions, receiving feedback (rewards or penalties), and gradually adjusting its behaviour to maximise positive outcomes. Rather than learning from a labelled dataset, the system learns from the consequences of its own decisions. A variant called RLHF (Reinforcement Learning from Human Feedback) was key to making tools like ChatGPT feel conversational rather than mechanical — human reviewers rated outputs, and the model learned from those ratings.

Examples: DeepMind's AlphaGo, which learned to play the board game Go by playing millions of games against itself. Robotics systems trained to walk or grasp objects through simulated trial and error. Recommendation systems that optimise for which content keeps you watching.

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 ability of an AI system to understand, interpret, and reason about physical space — including the position, shape, depth, and movement of objects in a three-dimensional environment. Spatial intelligence enables robots to navigate rooms, self-driving cars to understand roads, augmented reality tools to anchor digital content accurately onto physical surfaces, and manufacturing AI to inspect components in three dimensions.

Examples: The room-mapping in AR/VR headsets. LiDAR processing in self-driving cars. Robotic arms in warehouses that assess object size and orientation before picking. Medical imaging tools that reconstruct 3D models from CT scan slices.

Key distinction: Spatial intelligence is a separate capability from language understanding or pattern recognition. Most large language models have no spatial awareness — they process text, not three-dimensional space.

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?