Words this site uses

Every word here appears somewhere else on this site. Each definition has a fixed address you can send to a colleague: /en/glossary/#wer, /ar/glossary/#wer and /ckb/glossary/#wer are the same term in three languages.

Model: A computer program that has learned patterns from an enormous amount of text or audio. It is not a person and it does not know facts. It predicts what usually comes next.
Prompt: What you type or paste into the tool. Everything in the prompt is sent to the tool, including anything you pasted without reading it first.
Token: Tools count your text in tokens, not words. One token is roughly a few letters, and free limits and prices are measured in them.
Hallucination: When the tool states something false with complete confidence, an invented quote, law, date or source. It happens often enough that every fact has to be checked against the original.
Training data: The text, images and audio a model learned from. If your text becomes training data, parts of it can shape what the model later says to other people.
Open source: The instructions the software is made of are published, so anyone can read them, change them and run their own copy. Open source does not automatically mean private or safe.
Self-hosting: Running a tool on a computer your organisation controls instead of a company’s. You keep the data; you also take on the updating and the security.
Local and cloud: Local means the work happens on your own device and nothing leaves it. Cloud means your text travels over the internet to someone else’s computer.
API: A way for one program to talk to another with no person clicking anything. Programs use a key instead of a password, and a tool’s API terms are often different from the terms on its website.
MCP: Short for Model Context Protocol: a standard way to plug an AI tool into your files, mail or database. It hands the tool everything you connect it to, so connect as little as possible.
Agent skill: A saved set of instructions telling an AI tool how to do one job your way, every time. It adds a capability, not a safeguard: where your text goes does not change.
OCR: Reading printed or handwritten text inside a photo or a scan and turning it into text you can search and copy.
Transcription: Turning recorded speech into written text. Accuracy drops sharply with dialects and with noisy recordings.
RTL (right to left): The direction Arabic and Kurdish are written in. A tool with no right-to-left support will scramble how your text looks even when the translation itself is fine.
WER (word error rate): Out of every 100 words, how many the tool got wrong. A WER of 20 means one word in five is wrong. Lower is better.