AI basics: the essential terminology, in plain language

Walk into almost any meeting about artificial intelligence, and the jargon arrives within a minute. Models, tokens, agents, hallucinations, fine-tuning. Some of it is precise and useful. Some of it is marketing. Many people nod along without being sure what a word means, and then make decisions on that uncertainty.

This article is a glossary written for people who are about to plan AI in their organisation. It is grouped by theme, with short, plain definitions, and ends with the pairs of terms that are most often confused. Read it once, then return to it whenever a new word appears.

The big picture

• Artificial intelligence (AI). The broad field of making computers do tasks that normally need human intelligence, such as understanding language, recognising images or making predictions.
• Machine learning (ML). A way of building AI in which a computer learns patterns from examples, instead of following rules that a person wrote by hand.
• Deep learning. A kind of machine learning that uses large neural networks with many layers. It powers modern image, speech and language systems.
• Neural network. A system of connected units, loosely inspired by the brain, that learns by adjusting the strength of its connections.
• Generative AI. AI that creates new content, such as text, images, audio, code or 3D models, rather than only classifying or predicting.
• Algorithm. A set of steps for solving a problem. In machine learning, the algorithm describes how the system learns.

How a model is built

• Data. The examples a system learns from or works on: records, images, text, sensor readings.
• Dataset. A collection of data prepared for a purpose, such as training or testing.
• Features. The pieces of information fed to a model, such as the size of a field or the month of the year.
• Label. The correct answer attached to an example, such as “flooded” or “not flooded”. Creating labels is often the slowest part of a project.
• Supervised learning. Learning from examples that have labels.
• Unsupervised learning. Finding groups or patterns in data with no labels.
• Model. The result of training: the thing that takes an input and gives an output.
• Training. The process of fitting a model to data, by adjusting its settings to reduce errors.
• Parameters (or weights). The adjustable numbers inside a model that training sets. Large language models have billions of them.
• Validation and test sets. Examples held back from training, used to choose settings and to measure real performance honestly.
• Inference. Using a trained model to produce an answer for new input. Training is learning. Inference is working.
• Pre-trained model. A model already trained on a large amount of data by someone else, ready to use or adapt.
• Fine-tuning. Continuing the training of a pre-trained model on your own examples, to specialise it for your task.

Language AI: the words in every chatbot conversation

• Large language model (LLM). A very large model trained on huge amounts of text, which can read, write, summarise, translate and answer questions in natural language.
• Prompt. The instruction or question you give a language model.
• Prompt engineering. Writing prompts carefully, with context, examples and constraints, to get better results.
• System prompt. Standing instructions that set the model’s role and rules, usually hidden from the user.
• Token. A small chunk of text, roughly a short word or part of a word, that models read and write. Costs and limits are usually counted in tokens.
• Context window. How much text a model can consider at one time, including your question, the documents you supply and its answer.
• Temperature. A setting that controls randomness. Low values give steadier, more predictable answers. High values give more varied ones.
• Hallucination. When a model states something confidently that is false or unsupported. It is a normal failure mode of language models, not a rare glitch.
• Grounding. Tying an answer to real sources, so that it can be checked.
• Multimodal. Able to work with more than one kind of input, such as text, images and audio.

Giving AI your knowledge and letting it act

• Embedding. A list of numbers that captures the meaning of a piece of text or an image, so that similar things sit close together. It powers meaning-based search.
• Vector database. A store built to find the closest embeddings quickly.
• Retrieval-augmented generation (RAG). Finding relevant passages in your documents and giving them to a language model, so that its answer is based on them.
• Chunking. Splitting documents into pieces for indexing and retrieval.
• Knowledge base. The collection of documents and data an assistant can draw on.
• Tool use (or function calling). Letting a model call outside functions, such as search, a calculator or a database lookup.
• AI agent. A system that pursues a goal over several steps, deciding what to do next, using tools and checking results.
• Workflow automation. Fixed, rule-based sequences of steps. Predictable and cheap, and often a better fit than an agent when the steps never change.
• Orchestration. Coordinating models, tools and data so that the parts work together.

Measuring whether it works

• Accuracy. The share of answers that are right. It can mislead when one outcome is rare.
• Precision. Of the things the system flagged, how many really were what it said.
• Recall. Of the things that were really there, how many the system found.
• Baseline. How the task performs today, without the new system. It is what you compare against.
• Benchmark. A standard test used to compare models. Useful, but not a substitute for testing on your own tasks.
• Evaluation (or eval). A structured test of a system on a set of realistic examples.
• Overfitting. When a model memorises its training data and fails on anything new.
• Drift. When data or the world changes over time, so that a model that was accurate becomes less so.
• Confidence. How sure a model says it is. It is not always well matched to how often it is right.

Responsible and safe use

• Bias. Systematic unfairness that can come from unrepresentative data or design, and cause worse results for some people or places.
• Explainability. How well a person can understand why a system gave a result.
• Human in the loop. A design in which a person reviews or approves the system’s output before it has an effect.
• Guardrails. Rules and limits that keep a system within safe behaviour, such as restricted topics and approval steps.
• Personal data (PII). Information that can identify a person. It needs special care and often legal protection.
• Data governance. The rules for who owns data, who may use it, how it is kept and for how long.
• AI governance. The policies, roles and review processes for how AI is chosen, used and monitored.
• Prompt injection. An attack in which text that a system reads contains instructions meant to override its own.
• Data residency. The requirement that data stays in a particular country or region.
• Audit trail. A record of what a system did and why, for review.

Project and organisation words

• Use case. A specific task or decision where AI could help, such as answering routine policy questions.
• Proof of concept (PoC). A small test to show that an idea can work.
• Pilot. A limited trial of a working solution with real users and a clear success measure.
• MVP. The smallest version of a product that delivers real value and teaches you what to build next.
• Deployment. Putting a system into real use.
• Integration. Connecting a new system to the ones you already use.
• API. A way for software systems to talk to one another, for example to send a question to a model and get an answer.
• Cloud, on-premise and edge. Where a system runs: on a provider’s servers, on your own, or on a local device.
• Vendor lock-in. Becoming so dependent on one supplier’s tools or formats that changing becomes costly.
• Open source. Software whose code is shared under licences that allow use and modification, with conditions.
• Return on investment (ROI). The value gained compared with the cost, and best judged against a baseline.
• Readiness. How prepared an organisation is, in data, people, process and risk, to use AI well.

Pairs of terms that are often confused

• AI versus automation. Automation follows fixed rules. AI learns or reasons from data. Many useful systems combine the two.
• Chatbot versus agent. A chatbot replies. An agent plans and acts, using tools.
• Training versus prompting. Training changes the model. Prompting only changes what you ask it.
• Fine-tuning versus RAG. Fine-tuning teaches style and skills. RAG gives the model current facts to read. Facts that change belong in RAG.
• Dashboard versus model. A dashboard shows data. A model produces predictions or classifications from it.
• Pilot versus production. A pilot is a controlled test. Production is the supported, monitored version that people depend on.
• Accuracy versus usefulness. A system can be accurate on average and still unhelpful where it matters.

Questions that cut through buzzwords

When someone uses an unfamiliar term, you do not need to bluff. A few questions work almost everywhere.

1. 01Can you explain that in one sentence, without the jargon?
2. 02What would it look like in a real example from our work?
3. 03How would we measure whether it works for us?
4. 04What could go wrong, and who would notice?
5. 05What do we need to provide, in data and in people’s time?

How to use this glossary

• Share it with the whole planning group before your first workshop, so that everyone starts with the same words.
• Add your own terms: the names of your systems, datasets and teams.
• Revisit it after your first pilot, and see which terms became real for you.
• Use it when reading vendor material, and look for terms that are used loosely.

Frequently asked questions

Is ChatGPT an LLM or an agent?

A chat product is built on a large language model. Depending on its features, it may also use tools and act in several steps, which makes it behave like an agent. The names describe different layers of the same product.

Do we need to understand all this to plan AI?

You do not need the technical depth. You do need the shared vocabulary, so that decisions about data, risk and cost are made with the same meaning in everyone’s head.

What is the difference between AI and machine learning?

AI is the goal and the broad field. Machine learning is one method of reaching it, by learning from data. Most modern AI relies on machine learning.

Why does everyone talk about tokens?

Because language models read and write in tokens, and providers measure cost and limits in them. It is a practical unit for estimating price and for fitting text into a model’s context window.

Vocabulary is not the goal. It is the foundation for better conversations. With these terms in hand, you can read a proposal critically, brief a team clearly and plan your own organisation’s use of AI with more confidence and fewer surprises.