The AI & Machine Learning Glossary — Interactive Reference

☰ Filters, A–Z & search

Primer · Orientation

How AI learns — in plain language

Understanding AI's strengths and limits means understanding how the technology works. Before the definitions, three ideas that most of the vocabulary hangs from: what "AI" spans, how a model learns, and how it drives its errors down.

The stages of AI

"Artificial intelligence," coined in the 1950s, covers what exists today and what is still aspirational — sorted here by the machine's level of intelligence and what it can do.

Current AI · here today

Narrow (weak) AI

Reactive systems respond identically to the same input every time — IBM's chess engine Deep Blue. Limited-memory systems keep some memory of past performance and improve over time — ChatGPT, Gemini. Narrow in application, powerful within their lane.

Future AI · not yet built

Artificial general intelligence (AGI)

Intelligence typical of an average adult human — not tied to a single task, and potentially able to acquire genuinely novel abilities. A further step, a theory of mind, would model human mental states and emotions.

Aspirational AI · theoretical

Artificial superintelligence

General intelligence exceeding human capability, and — further still — self-aware AI that is conscious, or behaves indistinguishably from a human. Speculative, and central to a great deal of the risk debate.

Four ways a machine learns

Machine learning gives a computer a task, a measure of performance, and data — then lets it find the patterns itself. The initial learning phase usually takes one of four shapes.

Supervised

Humans label the data in advance — "these images are birds." The model learns to map labelled inputs to known outputs.

Unsupervised

The data is unlabelled. The model is free to find whatever structure it can — clusters, associations, outliers — with no instruction on what matters.

Semi-supervised

A blend: a little labelled data guides learning over a much larger pool of unlabelled data.

Reinforcement

The model takes a sequence of steps in an environment to maximise a reward — e.g. optimising fuel use across a set of bus routes.

Learning by minimising error

No model is right every time; the goal is to make the errors small. A loss function measures how far predictions sit from known answers. Adjusting the model's parameters step by step to push that loss to its lowest point — its minimum — is gradient descent, and it is, in essence, what "learning" means here.

Gradient descent. Each step adjusts the parameters to reduce error, walking down the curve toward the lowest loss. Too large a learning rate overshoots the minimum; too small and progress crawls. Related failure modes — overfitting and underfitting — are defined below.

Known challenges

Underfitting (too little signal to generalise) and overfitting (memorising the training data) both cripple prediction on new cases. Add errors in the data (especially systematic ones, which bake in invisible bias), the black-box problem (outputs with no explainable path), hallucination in large language models, and unresolved ethical and regulatory questions — and you have the practitioner's real agenda. Each term is defined in the glossary below.

Reference · A to Z

The glossary

81 of 81 terms shown

Praxis Lab ◆ Companion references

From vocabulary to obligation.

Knowing the terms is step one. Knowing which rules bind you — and when — is step two. The Praxis Lab trackers map the live regulatory picture on both sides of the Atlantic.

The EU AI Act, mapped US AI Regulation Tracker