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

**Authors:** Sophie Jentzsch, Kristian Kersting  
**arXiv:** [2306.04563](https://arxiv.org/abs/2306.04563)  
**Date:** 2023-06-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.04563v1/

## Key points

- Humor is a complex aspect of human communication that is not fully understood by artificial agents
- Large language models like OpenAI's ChatGPT can capture implicit and contextual information, leading some to wonder if they can tell jokes on a human level
- Researchers conducted experiments around joke generation, explanation, and detection to test this hypothesis
- ChatGPT accurately explained valid jokes but came up with fictional explanations for invalid ones in the joke generation task
- Certain characteristics commonly found in jokes (structure, topic, and wordplay) were central to ChatGPT's conception of humor but could also mislead it in classifying jokes
- ChatGPT represents a significant step towards creating "funny" machines, but more research is needed to fully understand how artificial agents comprehend and generate humor compared to humans.

## Summary

Humor is a complex aspect of human communication that has yet to be fully understood and replicated by artificial agents. Despite this, large language models (LLMs) such as OpenAI's ChatGPT have gained attention for their ability to capture implicit and contextual information, leading some to wonder if they can also tell jokes on a human level. To test this hypothesis, researchers conducted a series of exploratory experiments around joke generation, explanation, and detection. In the joke generation task, ChatGPT was tasked with generating valid jokes. However, this alone does not necessarily reflect the system's ability to understand humor from a human perspective. The results showed that while ChatGPT accurately explained valid jokes, it also came up with fictional explanations for invalid ones. To examine how closely ChatGPT's understanding of humor is connected to certain characteristics commonly found in jokes, researchers manually modified the top 25 jokes to eliminate one or multiple criteria: structure, topic, and wordplay. The results revealed that these criteria were central characteristics for ChatGPT's conception of humor but could also mislead it in classifying jokes. Overall, while ChatGPT has not yet solved computational humor entirely, it represents a significant step towards creating "funny" machines. However, more research is needed to fully understand how artificial agents comprehend and generate humor compared to humans.