Humor is a complex human skill. But can AI ever truly understand what makes something funny? New research from AI teams at the Allen Institute, Cornell, OpenAI and elsewhere explores this question using cartoons and captions from The New Yorker magazine's weekly contest.
In the contest, readers submit funny captions for a given cartoon image. The researchers developed three AI tasks using this data:
- Matching a caption to the correct cartoon image.
- Identifying the funniest caption among options.
- Explaining why a caption is funny in relation to the image.
The results showed current AI still struggles with core aspects of humor:
- The best AI scored only 62% on matching captions to images, versus 94% for humans.
- For explaining humor, human explanations were preferred over AI's in 68% of comparisons.
- AI makes mistakes in interpreting images that lead to inaccurate explanations.
- Subtleties like cultural allusions often go over AI's head.
While AI can now generate passable jokes, truly understanding humor remains difficult. Key challenges include perceiving incongruous imagery, resolving indirect connections to captions, and mastering the complexity of human culture/experience.
For businesses employing AI, the research highlights the need for oversight where humor or broader cognition is required. Progress is steady but human collaboration is still crucial in creative domains.
The contest data provides a benchmark to stretch AI's interpretive abilities. And for comedic creators, AI assistance may someday enhance ideation and feedback. But for now, the deepest comedy insights remain uniquely human.
Sources:
