Appendix · 13

Glossary

86 terms in plain language. Star the ones you keep having to explain, they are usually the ones that decide the room.

Artificial intelligence (AI): Computer systems that do tasks we usually link to human thinking, such as predicting, sorting, writing, or planning.
Machine learning: A way of building systems that learn patterns from examples, instead of being given every rule.
Model: A system that has learned from data and turns an input into an output.
Large language model (LLM): A model trained on huge amounts of text to predict and write language.
Generative AI: AI that creates new text, images, audio, code, or other content.
Agentic system / AI agent: A model connected to tools, memory, and instructions, so it can carry out a task in several steps.
Automation: Letting a system do a task with little direct help from a person.
Augmentation: Using AI to help people work, instead of replacing them.
Human-in-the-loop: A person checks, approves, or steps in during a process that uses AI.
Human-on-the-loop: A person watches over a system but does not check every action.
Human-out-of-the-loop: A system acts without a person checking each decision or action.
Training data: The examples a model learns from.
Dataset: An organised collection of data used to train, test, or study something.
Data provenance: Where data came from, and how it was collected or changed.
Data lineage: The path data takes through different systems, and what happens to it along the way.
Structured data: Data in fixed fields, like a spreadsheet or database.
Unstructured data: Text, images, audio, or video that does not fit into a table.
Synthetic data: Made-up data created to look like real data.
Ground truth: The answer taken as correct when testing a system.
Knowledge base: A chosen collection of information that a system can search.
Parametric knowledge: What a model has stored in its learned numbers, as opposed to what it looks up.
Factual knowledge: Facts and relationships a model may have picked up from its training data.
Retrieval-augmented generation (RAG): The system first looks up documents or records, then writes its answer from them.
Local knowledge: Knowledge that comes from the experience of a particular place or community.
Indigenous knowledge: Knowledge built up and kept by Indigenous peoples and communities.
Data sovereignty: The idea that data should be governed by the laws, rights, and authority of the people or place it is about.
Data minimisation: Collecting and keeping only the data you need for a stated purpose.
Sensitive data: Data that could cause harm if it is shared or misused.
Input / prompt: The instruction or material you give a model.
Output: What the system produces: text, a prediction, a category, an image, or an action.
Inference: Running a trained model to get an answer.
Context window: How much text or other material a model can take in at once.
Token: A small piece of text, often part of a word, that a model reads and writes.
Embedding: A list of numbers that represents a piece of text or an image, so a computer can compare meanings.
Classification: Putting an item into one or more categories.
Prediction: A best guess at an outcome, based on the data available.
Forecast: A prediction about the future.
Recommendation: An output that suggests what to do.
Optimisation: Searching for the best result, given a goal and some limits.
Objective function: The exact number a system tries to make as high or as low as possible.
Proxy: Something easy to measure used in place of a goal that is hard to measure.
Benchmark: A standard test used to compare models.
Evaluation: Measuring how well a model or system does against set criteria.
Uncertainty: How likely it is that an output is wrong or incomplete.
Confidence score: A number showing how sure a system is about a prediction.
Calibration: Whether a system's confidence matches how often it is actually right.
Explainability: Being able to give a clear account of how a system reached an output.
Interpretability: How well people can understand what goes on inside a system.
Black box: A system whose inner workings are hard to see or check.
Hallucination: A fluent answer that is false, made up, or not supported by any source.
Bias: Errors that consistently favour or disadvantage some groups, coming from the data, the design, or how the system is used.
Fairness: Ways to measure and fix unequal treatment or unequal results.
Robustness: Working reliably even when conditions change or inputs are messy.
Adversarial attack: An input designed on purpose to make a system fail or misbehave.
Neural network: A computer system made of connected layers that turn inputs into outputs.
Perceptron: An early maths model of a single brain cell. Later neural networks grew out of it.
Transformer: The design behind most modern language models. It works out which words in a text matter to each other.
Weights: The numbers a model learned during training. In practice, the weights are the model.
Open weights: Weights published so anyone can run, study, or host the model themselves.
Closed weights: Weights kept on the company's servers. You can only use the model through their service.
Pretraining: The first, largest stage of training, usually on predicting the next piece of text.
Fine-tuning: Extra training on a smaller dataset or a specific task.
Post-training: Training after pretraining that makes a model more useful and better behaved.
Reinforcement learning (RL): Training that rewards some actions and penalises others.
Reinforcement learning from human feedback (RLHF): Reinforcement learning where the rewards come from people's ratings.
Reasoning model: A model that spends extra effort working through multi-step problems before it answers.
Chain of thought: The step-by-step working a model writes out while solving a problem.
Scaling laws: The finding that models get better in a predictable way as they get bigger, see more data, and use more computing power.
Mixture of experts (MoE): A model made of many specialist parts, where only a few are used for each input. This makes it cheaper to run.
Alignment: Making a system behave according to chosen goals or values.
AI safety: Research and practice that aims to stop AI systems causing harm or getting out of control.
AI governance: The rules, institutions, and processes that shape how AI is built and used.
Responsible AI: Building and using AI with care for its social, ethical, legal, and environmental effects.
Value-sensitive design: Designing technology with human values and the interests of affected people in mind.
Sovereignty-aligned AI: AI built and governed so that a particular community or place keeps control over it.
Fine-grained permissions: Limits on exactly which tools, files, data, or actions a system can use.
Sandbox: A closed-off space where a system can run with limited access.
Human approval gate: A step where a person must approve before an action goes ahead.
Audit log: A record of what went into a system, what came out, and what it did.
Model card: A document describing what a model is for, its limits, and how it was tested.
System card: A document describing a whole deployed system, including its risks and safeguards.
Model drift: When a model's performance changes over time because the data or conditions change.
Versioning: Keeping track of changes to models, data, prompts, and software.
Rollback: Going back to an earlier version of a system.
Self-hosting: Running a model on hardware you or your institution control.
Zero data retention: A written promise from a provider that it stores nothing you send.