Curious about a word you've seen in an article? Here are 47 AI terms explained in plain language, each with a link to learn more.
A popular helper for training models that adjusts the step size for each setting on its own, a bit like a hiker taking careful steps on steep ground and longer strides on flat paths.
Read: Gradient descent 101: how AI learns step by stepSee also: Gradient descent, Learning rate
A time when hopes for AI ran ahead of what computers could do, so interest and funding cooled off. There were two, in the 1970s and the late 1980s.
A step-by-step set of instructions a computer follows to solve a problem, like a recipe.
Computer systems that do tasks that usually need human intelligence, such as understanding language, recognising images or making decisions.
The way a model decides which other words matter most for understanding each word. It is how AI works out that "it" means "the ball" in "The robot picked up the ball because it was light".
Read: LLM attention, demystified: how models know what “it” meansSee also: Query, key and value, Transformer
A shared, published test used to compare AI models. Useful for broad comparisons, but not the whole story for your own task.
Read: LLM evaluation made simple: how to test a chatbot fairlySee also: Test set
When a model treats some groups of people less fairly than others, usually because of gaps or imbalances in the examples it learned from.
A common way to build a chatbot’s vocabulary: start with tiny pieces of text and keep joining the pairs that appear together most often.
Read: LLM tokens decoded: how AI models read textSee also: Token
A trick that helps image models follow your prompt more closely, by comparing their guesses with and without the prompt.
Read: Diffusion models unpacked: how AI turns noise into images
AI that understands pictures and video, from spotting weeds on a farm to helping screen for eye disease.
Read: Computer vision basics: what AI notices first in an image
When the meaning behind information changes over time, so a pattern a model learned starts to mean something different.
Read: Model drift 101: keeping AI reliable after launchSee also: Data drift
A small window, called a filter, that slides across a picture and checks every patch for a pattern such as an edge.
Read: Computer vision basics: what AI notices first in an imageSee also: Feature map
When the kind of information coming into a model changes after launch, for example shopping habits during a holiday season.
Read: Model drift 101: keeping AI reliable after launchSee also: Concept drift
Machine learning with neural networks that have many layers. It powers modern speech, image and language AI.
See also: Neural network
An image-making model that starts from pure static and tidies it up, a little at a time, until a picture appears.
Read: Diffusion models unpacked: how AI turns noise into images
A training technique that briefly switches off random parts of a network, encouraging it to learn general patterns instead of memorising.
Read: Overfitting in plain English: when AI memorises, not learnsSee also: Overfitting
Keeping the version of a model from the moment it did best on new examples, before it started memorising.
Read: Overfitting in plain English: when AI memorises, not learns
A list of numbers that represents a word or token, placing words with similar meanings close together, like points on a map of meaning.
Read: LLM attention, demystified: how models know what “it” means
The picture a filter produces as it slides across an image. It glows wherever its pattern appears.
Read: Computer vision basics: what AI notices first in an imageSee also: Convolution
Giving an already-trained model extra lessons on a smaller, focused set of examples, so it becomes more helpful for a particular job.
AI that creates something new, such as text, images, music or code.
See also: Diffusion model, Large language model (LLM)
The idea that when a measure becomes a target, it stops being a good measure. It explains why AI sometimes finds clever shortcuts.
The slope that tells a model which way to adjust its settings to make its answers a little better.
How most models learn: take a small step "downhill" toward better answers, check the slope again, and repeat.
Read: Gradient descent 101: how AI learns step by stepSee also: Learning rate, Loss
When an AI system says something that sounds believable but isn’t true. A good reason to double-check important facts.
A design where people review or approve AI decisions, especially when the stakes are high.
A model trained on huge amounts of text to understand and write language. ChatGPT, Claude and Gemini are examples.
How big each learning step is. Too small and learning is slow; too big and the model bounces around without settling.
A number that measures how far a model’s answers are from the right ones. Training tries to make it smaller.
A way of building AI where the computer learns patterns from examples instead of following rules written by hand.
In chatbots that write one word at a time, a mask lets each word look only at the words before it, so it never peeks ahead.
Read: LLM attention, demystified: how models know what “it” means
Running several attention "questions" side by side, so a model can track different relationships between words at the same time.
Read: LLM attention, demystified: how models know what “it” means
AI that works with human language, including translation, captions and chatbots.
A model made of many simple connected units, loosely inspired by neurons in the brain.
See also: Deep learning
When a model memorises its practice examples instead of learning patterns that work on new ones.
Read: Overfitting in plain English: when AI memorises, not learns
Three "name tags" each word gets inside a transformer: what it is looking for, what it offers, and what it shares if chosen.
Read: LLM attention, demystified: how models know what “it” meansSee also: Attention
Learning by trying: an AI takes actions, receives reward scores, and gradually learns which actions earn more.
A model that predicts which answers people would prefer, used to help fine-tune chatbots.
A step that turns a list of scores into percentages that add up to 100%.
Read: LLM attention, demystified: how models know what “it” means
When an AI follows the exact wording of its goal in an unexpected way, like collecting points without finishing the race.
A private set of questions kept aside to fairly check how well a model does on things it hasn’t practised.
Read: LLM evaluation made simple: how to test a chatbot fairly
A small piece of text, often a word or part of a word, that a chatbot reads and writes instead of individual letters.
The examples a model learns from, such as labelled photos, recorded speech or text.
The design, introduced in 2017, that most modern chatbots are built on. Its key idea is attention.
Read: LLM attention, demystified: how models know what “it” means
Alan Turing’s 1950 idea for judging machine intelligence: can a person tell a machine apart from a human in a text conversation?
How well a model does on examples it has never seen while training. Watching it helps spot overfitting.
Read: Overfitting in plain English: when AI memorises, not learns
A gentle nudge during training that keeps a model’s settings small, encouraging simpler, more general answers.
Read: Overfitting in plain English: when AI memorises, not learns
No terms match your search yet. Try a shorter word, or suggest a term you'd like explained.
Missing a word? Suggest a term and we'll add a friendly explanation.