Why ChatGPT and other AI chatbots make things up
How to catch an AI chatbot making things up
There is a quick test you can run on any answer from an AI chatbot like ChatGPT. It needs no tools and takes less than a minute. It does not prove that an answer is true, but it shows you where the chatbot is guessing.
-
1
Time for a test. You open a new chat in temporary mode, so it does not remember your other chats. Then you ask the very same question about Rono's shop.
-
2
This time the chatbot says that the shop opened in 2005, that two brothers started it, and that it is known for its café corner. New dates, new founders, and still not a word of it is true.
-
3
Then you ask who painted the Mona Lisa, in one new chat after another. The answer is Leonardo da Vinci, every time.
-
4
That is the test. When a model really learned a fact, it gives the same fact again and again. When it is guessing, the details often move around. Researchers use this very idea to catch many hallucinations.
-
5
A fact that stays the same can still be wrong. A popular myth, a false idea that many texts repeat, comes back every time too. So for a name, a date or a number that matters, look it up yourself, in a place you trust.
-
6
But look at what just happened. The same chatbot got the same question twice, and different words came out. What happens inside, between your question and the first word of the answer? Follow your question all the way in, in "How ChatGPT and other AI chatbots answer your question".
-
Quick check
One band, two years.
Show the answer
That it is guessing: a year it really learned would come back the same. Neither year can be trusted, so look for the band's own page.
-
Go further
Where to go next.