While artificial intelligence can now generate jokes, a new study suggests it lacks understanding of what makes them funny. In a recent experiment, AI models and humans were tested on tasks involving New Yorker magazine’s Cartoon Caption Contest entries.
Large neural networks, a form of artificial intelligence, were put to the test alongside humans on tasks involving jokes and humor from the New Yorker’s Cartoon Caption Contest. Despite impressive advances in AI, it became evident that humans outperformed machines in all the given tasks. The lead author, Jack Hessel, acknowledged the difficulty in truly assessing “understanding” in machines while still admiring their noteworthy performance.
The study showcased a considerable disparity between AI and human-level comprehension of humor. In a multiple-choice test, AI’s ability to match cartoons with captions achieved only 62% accuracy, lagging far behind the human performance of 94% in the same setting. Furthermore, when comparing human-generated explanations with AI-generated ones, the former was preferred by a ratio of roughly 2-to-1. While AI may not fully grasp humor yet, the researchers believe it could serve as a valuable collaborative tool for humorists to stimulate creativity and brainstorm innovative ideas.