AI Glossary: Simple Definitions of Tricky Words
An AI glossary for kids and grown-ups. Every word gets a short definition first, then a little more. Click a word's # to copy a link straight to it.
- #AGI (artificial general intelligence)
-
An AI that could learn almost any task a person can, and move between tasks the way people do. Nobody has built one yet.
The 'general' is the key: not one skill, but the ability to pick up new ones.
- #AI (artificial intelligence)
-
A computer program that can do things we usually think need a brain — like recognising a face, translating a language or playing a game.
Most AI today learns from examples rather than following rules written by people.
- #Algorithm
-
A list of steps for solving a problem, like a recipe. Computers follow algorithms.
'Turn left, walk ten steps, turn right' is an algorithm for getting to the park.
- #Alignment
-
Making sure an AI wants what we really mean, not just what we literally said.
The Wish Machine granting 'make everyone smile' with paper smiles is an alignment failure.
- #ASI (artificial superintelligence)
-
Another name for superintelligence: a mind far smarter than the smartest humans at almost everything.
People say ASI to make clear they mean 'beyond human', not 'human-level' (AGI).
- #Benchmark
-
A test used to measure how good an AI is at something.
Passing a benchmark is not the same as doing the real job.
- #Benefit
-
A good result — like a new medicine or a cleaner ocean.
- #Chatbot
-
An AI you can talk to by typing or speaking. It predicts good replies based on the text it was trained on.
Chatbots are impressive but not superintelligent — they still make simple mistakes.
- #Compute
-
The computer power an AI runs on — chips, electricity and time.
Bigger models need more compute, which is expensive and physical.
- #Data
-
The examples an AI learns from — pictures, words, sounds, numbers.
If the data is wrong or unfair, the AI learns wrong or unfair things.
- #Goal
-
What a system is trying to achieve or make bigger.
An AI's written goal is never exactly the goal in our heads — that gap is the alignment problem.
- #Guardrails
-
Rules and checks that stop an AI from doing harmful things.
Like the barriers on a mountain road: they don't drive the car, they stop it going off the edge.
- #Hallucination
-
When an AI confidently says something that isn't true — like inventing a book that doesn't exist.
It happens because the AI learned patterns of words, not facts about the world.
- #Intelligence
-
The ability to learn, understand and solve new problems. It is a toolbox of many skills, not one number.
Noticing, remembering, reasoning, planning, imagining and understanding people are all tools in the box.
- #Intelligence explosion
-
The idea that an AI smart enough to improve itself could build a smarter AI, which builds a smarter one, faster each time.
Some experts think this could happen fast; others think progress would stay gradual.
- #Memory
-
Storing what you learned so you can use it later.
Today's AI mostly remembers what it saw in training, not what happened yesterday.
- #Milestone
-
An important moment along the way to something — like the first computer to beat a chess champion.
Our timeline is a list of milestones.
- #Model
-
The finished, trained AI — the thing you actually talk to or use.
People say 'a new model' the way you might say 'a new version'.
- #Narrow AI
-
AI that is brilliant at one job — like chess or spotting cats in photos — and useless at everything else.
Almost every AI you have used is narrow AI.
- #Neural network
-
A web of connected 'dots' inside a computer, loosely inspired by the brain, that learns patterns from examples.
The connections get stronger or weaker during training until the network gives good answers.
- #Prediction
-
A guess about the future. Not a fact.
AI predictions have been wrong in both directions before.
- #Prompt
-
What you type or say to an AI to tell it what you want.
Writing a good prompt is a bit like making a good wish: say what you really mean.
- #Reasoning
-
Working things out step by step.
'If it's raining and I have no coat, I'll get wet' is reasoning.
- #Recursive self-improvement
-
A system that upgrades its own design, then uses the upgrade to make the next upgrade.
The engine behind the intelligence explosion idea.
- #Risk
-
A chance that something goes wrong.
Being careful means lowering risks without giving up benefits.
- #Superintelligence
-
A mind much smarter than the smartest humans at almost everything. Still an idea, not a real thing.
It might think thousands of times faster than us and run as many copies at once.
- #Training
-
Teaching an AI by showing it lots of examples and correcting its guesses.
A chatbot is trained on huge amounts of text; a fruit-spotter on thousands of fruit pictures.