AI WORLD HQDictionary
The little AI dictionary
Every AI word you keep seeing — explained in one breath, no jargon.
AI (artificial intelligence)
Computer programs that can do tasks which normally need human thinking — writing, answering questions, recognising pictures. The AI you meet online today is clever software, not a thinking robot.
Chatbot / AI assistant
A program you talk to by typing or speaking — like ChatGPT, Gemini or Claude. You ask in normal sentences, it answers in normal sentences. No commands or code needed.
LLM (large language model)
The engine inside a chatbot. It learned from enormous amounts of text how words follow each other, which is how it writes and summarises so well. It predicts likely words — it doesn't "know" facts the way you do.
Prompt
The instruction or question you type to an AI. A better prompt gets a better answer: say what you want, who it's for and in what style — like briefing a helpful colleague.
Prompt engineering
The craft of phrasing your requests so the AI gives you what you actually need. No coding involved — just clear instructions, examples and context. Anyone can learn the basics in an afternoon.
Token
The small chunks of text an AI reads and writes — a word is usually one or two tokens. Prices and limits of AI services are often measured in tokens, which is why the word pops up everywhere.
Context window
How much text the AI can keep "in mind" within one conversation. If a chat gets very long, the oldest parts slip out — which is why the AI sometimes seems to forget what you said earlier.
Hallucination
When an AI states something false with full confidence — an invented fact, source or date. It's not lying on purpose; it's predicting plausible-sounding text. Always double-check facts that matter.
Generative AI
AI that creates new content — text, pictures, music, video — rather than just sorting or finding existing things. ChatGPT writing a poem and an app drawing a picture from your description are both generative AI.
Image generator
An AI that draws a picture from your written description — "a koala on a surfboard at sunset" becomes an actual image in seconds. Built into tools like ChatGPT, Gemini and Canva.
Multimodal
An AI that understands more than text: you can show it a photo, a document or a voice message, and it can work with those too — describe the photo, summarise the document, answer the voice note.
AI agent
An AI that doesn't just answer but carries out multi-step tasks with some independence — searching, comparing, filling things in, booking. You give the goal; it works through the steps.
Fine-tuning
Extra training that specialises a general AI for a specific job — say, medical texts or one company's writing style. Like sending a good all-round student to trade school.
Open-source model
An AI model whose "recipe" is public: anyone can download it, run it and build on it, often for free. Meta's Llama family is a well-known example.
API
The plug through which programs talk to each other. Developers use AI companies' APIs to build AI into their own apps and websites — that's how AI shows up inside so many tools you already use.
Deepfake
An AI-made fake video, photo or voice that looks or sounds real. If a shocking clip of a famous person appears out of nowhere, check a trusted news source before believing or sharing it.
Voice cloning
AI copying a real person's voice from a short sample. Useful for audiobooks and accessibility — and abused in phone scams. Agree on a family password for "urgent money" calls.
Training data
The huge pile of text and images an AI learned from. Whatever wasn't in it — very recent events, your private files — the AI can't know by itself, though many assistants can now search the web to catch up.
AGI (artificial general intelligence)
A still-hypothetical AI that could think as flexibly as a human across every area. Today's AI is not this — it's very capable at specific tasks, which is exactly why it's useful and predictable.
Machine learning
A way of building AI: instead of being given step-by-step rules, the computer learns patterns from lots of examples — like learning to recognise cats by seeing thousands of cat photos.
Neural network
The kind of software design most modern AI uses, loosely inspired by how brain cells connect. It's not a real brain — just a maths structure that's very good at spotting patterns.
GPT
The letters stand for ‘Generative Pre-trained Transformer’ — the technology behind ChatGPT and many other assistants. In plain terms: an AI trained on huge amounts of text that generates new text word by word.
Reasoning model
A newer kind of AI that ‘thinks’ step by step before answering, instead of replying instantly. It's slower but often better at maths, logic and tricky problems.
Knowledge cut-off
The date after which an AI hasn't learned anything new. Ask about events after that date and it may not know — unless it can search the web. Always check recent facts.
Bias
When an AI's answers unfairly favour or disadvantage certain groups, because the data it learned from was unbalanced. It's a known problem — treat AI output as a helpful draft, not the final word.
AI watermark
A hidden or visible mark added to AI-made images, video or text so it can be recognised as AI-generated. Still patchy — its absence doesn't prove something is real.
Copilot
A common name for an AI helper built into a tool you already use — writing code, documents or emails alongside you. Microsoft and GitHub use the name, but the idea is general: AI as your assistant, not the pilot.
Jailbreak
Tricking an AI into ignoring its own safety rules — for example to get it to say something it normally wouldn't. Companies keep patching these, and trying them can break a service's terms of use.
Benchmark
A standard test used to compare AI models — like exams for AI. When you read ‘model X beats Y on a benchmark’, it means X scored higher on that particular test, not that it's better at everything.
AI chip (GPU)
The powerful computer chips that run AI. Graphics chips (GPUs), first made for video games, turned out to be brilliant at AI maths — which is why companies like Nvidia are in the news so much.
Guardrails
The built-in limits that stop an AI from doing harmful things — like refusing dangerous instructions. They're not perfect, but they're why an assistant will politely decline some requests.
AI model
The 'brain' behind an AI tool — the trained program that actually does the work, like writing text or making pictures. When you hear names like GPT, Gemini or Claude, those are AI models.
Algorithm
A set of step-by-step rules a computer follows to decide or sort things — like what appears first in your social media feed. 'The algorithm' usually means the hidden rules that decide what you see online.
Machine translation
When software automatically translates text or speech from one language to another, like Google Translate. It's fast and handy, but can still miss jokes, slang or fine nuance.
Voice assistant
An AI you talk to out loud that answers or does small tasks — like Siri, Alexa or Google Assistant. You speak, it listens, then replies, sets a timer or plays music.
Facial recognition
Technology that identifies or verifies a person from their face, like unlocking your phone just by looking at it. It's convenient, but also raises privacy questions when used in public places.
Recommendation system
The AI that suggests what you might like next — the 'Recommended for you' on Netflix, YouTube or online shops. It learns from what you (and others) watched or bought to guess your next pick.
AI avatar
A digital character or lifelike video version of a person, created by AI — used for videos, presentations or virtual hosts. It can look and speak like a real person, with no camera crew.
Foundation model
A big, general-purpose AI trained on huge amounts of data that many tools are then built on top of. 'Frontier model' means the newest, most capable ones at the cutting edge.
RAG (retrieval-augmented generation)
A method that lets a chatbot look things up in a trusted source before answering, instead of relying only on memory. It's why some AI assistants can cite documents or give up-to-date facts.
NLP (natural language processing)
The part of AI that deals with human language — understanding, writing and translating text or speech. It's what lets you type or talk to a computer in plain words instead of code.
Computer vision
The part of AI that 'sees' — making sense of photos and video, like spotting a face, reading a sign or telling a cat from a dog. It powers things like photo search and self-driving car cameras.
Superintelligence
A hypothetical future AI that would be far smarter than humans at almost everything. It doesn't exist today — it's a topic of research and debate about where AI might eventually lead.