German Researchers Test the Humor of ChatGPT and Its Fun Factor

Researchers from the German Aerospace Center have discovered that OpenAI’s ChatGPT-3.5 model, an artificial intelligence chatbot, has a limited knowledge of jokes. In a study of the AI’s ability to generate humor, the researchers found that more than 90% of 1,008 generated jokes repeated a set of 25 jokes, suggesting that the AI had learned and memorised them during the training process rather than generating them anew. The researchers tested the model by prompting it to tell a joke 1,000 times, with almost all responses containing exactly one joke and the variation in prompts having no noticeable effect. Researchers also found that the AI struggled with sequences that didn’t fit into learned patterns and wasn’t able to tell when a joke wasn’t funny. Nonetheless, the research showed that ChatGPT could understand some elements of humour, such as puns and wordplay, to a degree.