AI Humor Generation: Cognitive, Social and Creative Skills for Effective Humor

TL;DR

This work investigates whether AI can generate humor by adopting human-like cognitive, social, and creative skills, focusing on Gen Z Instagram caption humor. It introduces HumorSkills, a multi-stage pipeline combining visual detail extraction, humor ideation, narrative extrapolation, caption generation, and a Gen-Z ranking agent, all fine-tuned on Gen-Z humor data. In rating studies, HumorSkills outperforms GPT-4o and rivals top Instagram captions in funniness, with significant gains on multiple datasets and robust cross-domain performance to camera-roll and museum-art images. The findings suggest that endowing AI with targeted human-like skills can enhance its ability to produce relatable, engaging humor, with implications for human-AI social bonding and potential ethical considerations.

Abstract

Humor is a social binding agent. It is an act of creativity that can provoke emotional reactions on a broad range of topics. Humor has long been thought to be "too human" for AI to generate. However, humans are complex, and humor requires our complex set of skills: cognitive reasoning, social understanding, a broad base of knowledge, creative thinking, and audience understanding. We explore whether giving AI such skills enables it to write humor. We target one audience: Gen Z humor fans. We ask people to rate meme caption humor from three sources: highly upvoted human captions, 2) basic LLMs, and 3) LLMs captions with humor skills. We find that users like LLMs captions with humor skills more than basic LLMs and almost on par with top-rated humor written by people. We discuss how giving AI human-like skills can help it generate communication that resonates with people.

Figures (6)

• Figure 1: HumorSkills System Diagram. Given an image, the system first extracts visual details with a visual language model, then performs visual humor ideation to analyze the image and propose humorous angles. It then generates ten potential conflicts that could be used to extrapolate the image into a relatable experience. The system then generates humor with and without the narratives, for diversity. A separate instance of the LLM trained to rank gen-Z humor ranks all the captions and returns the top five.
• Figure 2: A diagram for how narrative extrapolation works
• Figure 3: Top Rated Image Captions (marked in green) for Instagram, GPT-4o, and HumorSkills and corresponding top scorers for each image. From left to right, the images contain 1) a guitar next to a machine gun hanging on a wall, 2) a man running a race while smoking a cigarette and, 3) a man with a long beard and his head cropped to a trapezoid shape.
• Figure 4: Images across all 3 datasets with Gen Z slang. Instagram (left), Flickr (center), Museum Art (right)
• Figure 5: Images across all 3 datasets with narrative generation. Instagram (left), Flickr (center), Museum Art (right)