A recent study has tested the sense of humor of an AI language model and found that over 90% of the jokes generated were identical. Humor is known to be an important factor in human-computer interaction, and a computational sense of humor has the potential to greatly improve this interaction. However, despite ongoing improvements and recent machine learning discoveries, creating a genuinely funny AI remains an elusive goal.
The study focused on OpenAI’s ChatGPT, an AI model that is capable of engaging in conversation with users through a public chat API. While the AI produced nearly flawless English, the study found that it did occasionally make grammar and content-related errors. One of the most notable shortcomings was that ChatGPT often repeated the same jokes, with researchers finding that 90% of the jokes generated were the same 25 jokes, likely copied from the training data or hard-coded into a list.
Although ChatGPT is capable of reflecting and capturing implicit knowledge, including humor and stereotypes, the study found that the AI model did not actually create the jokes it produced. To further investigate the system’s behavior, researchers conducted prompt-based experiments to gain insights into the model’s generation process.
The study’s authors caution that while AI models strive to emulate human behavior, they are not on a human level and are at best simulations of the human mind. Nonetheless, they remain hopeful that recent machine learning discoveries will enable the creation of more accurate and nuanced AI humor.
The study was conducted by researchers from the German Aerospace Center, Technical University Darmstadt, and Hessian Center for AI, whose goal was to determine how well ChatGPT could capture human humor. The three experimental conditions of joke invention, joke explanation, and joke detection were compiled as the major contribution.
In conclusion, while AI has come a long way, creating a genuinely funny AI remains a distant goal. However, recent machine learning discoveries offer hope that the creation of more accurate and nuanced AI humor is possible. Whether AI can ever think or understand consciously remains a philosophical question, but for now, we can enjoy ChatGPT’s attempts at humor.