AI's Missteps in the Game of Humor: When Models Laugh Alone

Humor, a cornerstone of human interaction, remains an elusive target for AI. In a recent exploration, five leading language models were put to the test in the social game Cards Against Humanity. The task? Select the funniest response from a set of ten options, across a staggering 9,894 rounds.

Models vs. Humans

While these models managed to perform better than random chance, they still struggle to align with human humor preferences. Yet, the real eyebrow-raiser is their tendency to agree with each other more than with human players. This unusual camaraderie among models begs the question: are they genuinely funny or merely mirroring systemic biases?

Slapping a model on a GPU rental isn't a convergence thesis. The models’ humor choices seem driven by predictable patterns in position and content. This mechanistic approach stands at odds with the nuanced and often unpredictable nature of human humor, raising doubts about whether AI can ever truly 'get' what makes us laugh.

Structural Artifacts or Genuine Preferences?

As these models navigate the murky waters of humor, one can't help but ask: Is their judgment reflecting a true understanding of humor, or are these just structural artifacts of model inference and alignment? If the AI can hold a wallet, who writes the risk model? The underlying biases suggest that AI’s comedic sensibilities might be more artifice than authenticity.

Decentralized compute sounds great until you benchmark the latency. In the same vein, AI's foray into humor illustrates a broader challenge, AI's ability to genuinely grasp and replicate human emotion. Ninety percent of these AI projects won't see the light of day, but when they do, they could revolutionize the way we interact with technology.

Why It Matters

The intersection is real, but humor highlights the chasm that still exists between human and machine. As industry giants pour resources into developing models with human-like alignment, humor tests such as these serve as a sobering reminder of the complexities involved. After all, what use is an AI partner who can't share in a good laugh?