ChatGPT is Fun, but Not Funny: Humor Remains Challenging for Large Language Models

"ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models" explores ChatGPT's ability to understand and generate humor. The article argues that, while the model may be 'amusing', it is not truly 'funny'. To evaluate the language model's capabilities, the authors conducted a series of experiments involving the generation, explanation, and detection of jokes using prompts.

They discovered that ChatGPT frequently repeats a limited number of existing jokes rather than generating new ones — over 90% of 1008 generated jokes were variations of the same 25 jokes. The system can explain puns and stylistic elements in familiar jokes, but invents absurd explanations for invalid jokes, revealing a limited understanding of humor that focuses on certain structures, puns and themes.

The central conclusion is that ChatGPT has learned a specific joke pattern instead of being able to be actually funny. Its generation, explanation, and identification of jokes focuses on meaning and content rather than deeper comedic understanding. The study demonstrates that, although ChatGPT has made significant progress in computational humor, it has not yet achieved a human-level understanding of humor — it's pattern matching, not true comedic insight.

ChatGPT is fun, but it is not funny! Humor is still challenging Large Language Models

Sophie Jentzsch, Kristian Kersting