Watch an assistant invent a telephone number
Understand what a chat assistant is doing when it answers — and why it can be wrong with total confidence.
You ask an assistant for a pharmacy’s phone number. It gives you one at once, with complete confidence. There is no such number.
ChatGPT, Copilot, Gemini — whatever your phone has started offering — all work the same way underneath: choosing a likely next word, over and over. Many of them can now also search the web while they answer, and when they do they usually show you the links they used. That is a real improvement and it is not the same as knowing: the sentence you are handed is still assembled a word at a time, and it can still describe the page it just read incorrectly. Where there are links, the answer is the links.
Key points
- It works by choosing a likely next word, over and over
- Plausible and true are made of the same material — it cannot tell them apart
- Hours, telephone numbers and recent events get invented, confidently
- It is not a search engine, not a person, and not accountable
What to remember
- It chooses what usually comes next. Everything else follows from that.
- Strong on language, weak on the world.
Try this
Ask an assistant about something you know perfectly well — your own street, or a job you did for thirty years.
Practice: Watch an assistant invent a telephone number. Your first lesson is free, with no account or card; the rest are part of a Lantern plan. Interactive lessons require JavaScript.