Why ChatGPT and other AI chatbots make things up
Why a confident tone from an AI chatbot proves nothing
A wrong answer from an AI chatbot almost never starts with "maybe". Even Google's own chatbot, in an ad for itself, said something false in a perfectly calm voice, and it cost Google a lot of money.
-
1
Rono is not the only one to be fooled. In February 2023, Google showed its new chatbot, Bard, in an ad. Someone asked what to tell a 9-year-old about the James Webb Space Telescope. One line of the answer said that the telescope took the very first pictures of a planet outside our solar system.
-
2
Astronomers saw the mistake at once. The first picture of a planet outside our solar system was taken in 2004, by another telescope, in Chile.
-
3
On the day the mistake made the news, the shares of Alphabet, the company that owns Google, fell. In one day, the company lost about $100 billion of its value. The mistake was one of the reasons.
-
4
Look at both answers again: Bard's and the one about Rono's shop. No "maybe", no "I think", no "please check".
-
5
The calm voice is made of words too, and the model guesses them like all the others. When the most likely words are plain and sure, a plain, sure sentence comes out, true or not. Researchers who tested chatbots found the same: they often sound sure when they are wrong.
-
6
So the tone tells you almost nothing about the facts. To catch a guess, you need something else, and there is a test you can try in a minute, right here in the shop.
-
Quick check
Two answers, two tones.
Show the answer
You cannot tell from the tone. Both sentences are guessed the same way. If the date matters, check it somewhere else.
-
Go further
What became of them.