Can AI crack jokes? Human-like skills help machines resonate with audiences

Traditional AI-generated humor has often been perceived as awkward, relying heavily on wordplay or nonsensical associations. However, the study explores a more sophisticated approach using HumorSkills, an AI model fine-tuned to generate humorous captions for Instagram images. By incorporating visual observations, social analogies, and narrative extrapolation, the AI system creates captions that attempt to mirror the humor styles of Gen Z audiences.

Humor has long been considered one of the most complex human traits, requiring a deep understanding of social norms, cognitive reasoning, and creativity. While AI has made significant strides in natural language processing and creative generation, humor remains a difficult frontier.

A recent study, "AI Humor Generation: Cognitive, Social, and Creative Skills for Effective Humor," by Sean Kim and Lydia B. Chilton of Columbia University, published on arXiv, investigates whether AI can be trained to generate humor that resonates with human audiences, particularly Gen Z humor fans.

The challenges of AI-generated humor

Humor is difficult for AI because it is not governed by a single formula. Different people find different things funny depending on timing, context, and cultural background. AI must navigate cognitive reasoning, social awareness, and linguistic creativity to generate jokes that feel natural rather than robotic. This study highlights three key elements of humor: cognitive skills for understanding logical inconsistencies, social skills for audience adaptation, and creative skills for generating unexpected yet relevant punchlines.

Traditional AI-generated humor has often been perceived as awkward, relying heavily on wordplay or nonsensical associations. However, the study explores a more sophisticated approach using HumorSkills, an AI model fine-tuned to generate humorous captions for Instagram images. By incorporating visual observations, social analogies, and narrative extrapolation, the AI system creates captions that attempt to mirror the humor styles of Gen Z audiences.

Evaluating AI humor: A comparison with humans

To test whether AI-generated humor could match human standards, the study compared three sources of humor:

• Highly upvoted human-written captions from Instagram.
• Captions generated by a basic large language model (LLM).
• Captions generated by HumorSkills, an AI system trained with cognitive and social humor strategies.

The results showed that HumorSkills outperformed basic AI models in humor ratings and came surprisingly close to the top-rated human captions. In a survey where users rated captions for 20 images, HumorSkills' captions were only 0.08 points lower on a 5-point humor scale than the best human-written captions. This suggests that AI, when properly trained, can create humor that resonates with a specific audience, though it still struggles with nuance and unpredictability compared to human humorists.

The role of narrative and social understanding in AI humor

One of the key advancements in HumorSkills was its ability to generate narrative-driven humor. Instead of simply describing an image, the model extrapolated broader social themes and common Gen Z experiences. For example, when given an image of a small dog standing defiantly in an urban street, the AI captioned it as “Why does this look like a GTA loading screen?” - a reference to the popular video game Grand Theft Auto. This ability to connect images to broader cultural references helped AI-generated humor feel more relatable and natural.

Additionally, the model incorporated social conflict and exaggeration - two critical elements of humor writing. It reframed mundane situations into exaggerated struggles, such as describing a demolition worker with a small hose as “Me trying to fix my life with self-care Sundays”, adding an emotional layer that made the caption more engaging.

The future of AI-generated humor

While this study demonstrates that AI can generate humor with a reasonable degree of success, significant challenges remain. AI still struggles with irony, sarcasm, and emotional depth, which require a deeper understanding of human experiences. Moreover, humor is highly subjective, and while HumorSkills performed well with Gen Z audiences, it may not translate effectively across different age groups or cultures.

The future of AI humor lies in improving AI’s ability to understand and mimic human social interactions. Enhancing AI’s capacity for self-reflection, long-term narrative building, and deeper audience customization could lead to humor that is more dynamic and adaptable. The study concludes that while AI-generated humor is not yet indistinguishable from human wit, it is rapidly advancing. With the right cognitive, social, and creative frameworks, AI may one day tell jokes that are not just amusing but truly insightful and universally funny.