Why ChatGPT and other AI chatbots make things up
Why AI chatbots get famous facts right and rare facts wrong
Ask an AI chatbot who painted the Mona Lisa, and it is right every time. Ask it about a small shop or a person few people know, and it can invent a whole life. Same program, same way of writing, very different results.
-
1
Rono wants to test the chatbot. He asks who painted the Mona Lisa. "Leonardo da Vinci," it says, and that is right.
-
2
The words "Mona Lisa" and "Leonardo" sit together on a huge number of pages: books, school lessons, museum guides. In training, the model met them again and again. So after "painted by", "Leonardo" is by far the most likely word.
-
3
Now think of Rono's shop. Almost nothing was ever written about it, and maybe nothing the model read. There is no strong next word about it.
-
4
So the model uses what texts say about small shops in general. Many of those shops are family shops, and many of those texts name a year and a prize. Those words fit, so the model writes them.
-
5
An answer that reads well and is not true is called a hallucination. Nothing broke inside the model. It came from the same guessing that named Leonardo.
-
6
Rono is still puzzled by one thing. The story of his shop did not sound like a guess at all. It sounded just as sure as Leonardo.
-
Quick check
Two kinds of fact.
Show the answer
The coach. The capital of Japan is written in millions of texts, so the right word is by far the likeliest guess. A small club is on a few pages at most, so the model fills the gap with a name that fits.
-
Go further
One name, and one trap.