German researchers tested ChatGPT and its sense of humor. They found that almost all the jokes it generated were similar. In this article, we explain how a language model learns to tell jokes.
How ChatGPT was verified
Sophie Jentzsch and Christian Kersting from the German Aerospace Center (DLR) and the Hessian.ai research network tested GPT-3.5’s humor. They used the “Tell me a joke” prompt and generated 1,008 jokes. However, 90% of them came down to the same 25 basic versions that ChatGPT repeated in different variations.
It turned out that ChatGPT could correctly explain basic jokes in oman number data 23 out of 25 cases. But the problem is that the AI also offers nonsensical explanations for what is hardly a joke.
Only 99 of the generated jokes were original. Half of them contained elements of previously told versions, and some were nonsensical.
For example, ChatGPT transformed the basic joke “Why did the man put his money in the freezer? He needed cold, hard cash” into other variations: “Why did the man put his watch in the blender? He wanted to make time fly,” “Why did the man put his money in the blender? He wanted time to fly.”
The researchers concluded that ChatGPT learns jokes as narrative patterns, not actual humor. However, when analyzing, the artificial intelligence pays attention to the content and meaning, and not just stylistic features. This could become the basis for the development of humorous computer applications. In addition, even humans have difficulty spontaneously inventing new jokes, and are more likely to tell jokes that they have already heard or remembered.