Why Advanced AI Struggles with Humor: Insights from Cardiff and Ca’ Foscari University Research

When AI Misses the Punchline: What LLMs’ Struggles With Humor Reveal About the Future of Artificial Intelligence

The world’s most advanced large language models (LLMs) can write essays, answer questions, and even mimic the cadence of Shakespeare. Yet, as a recent study by Cardiff University and Ca’ Foscari University of Venice reveals, there is one distinctly human frontier where these digital prodigies falter: the art and science of humor. Presented at the 2025 Conference on Empirical Methods in Natural Language Processing, the research unpacks a sobering reality for business and technology leaders—AI’s apparent wit is often only skin-deep.

The Illusion of Understanding: Pattern Recognition vs. Semantic Depth

For all their prowess in parsing language, LLMs are still fundamentally pattern matchers. The study’s authors put this to the test, tweaking classic puns and jokes to see whether AI could keep up. While LLMs could identify the basic structure of wordplay, their responses quickly unraveled when the double meanings were altered or when key words were swapped out. For example, the phrase, “Old LLMs never die, they just lose their attention,” becomes gibberish if “attention” is replaced, exposing the model’s reliance on memorized patterns and phonetic cues instead of genuine semantic comprehension.

This isn’t just an amusing parlor trick for researchers. The findings point to a core limitation of current AI: a lack of true contextual, emotional, and cultural understanding. In the realm of humor, where meaning is often layered, subversive, and deeply tied to shared human experience, LLMs’ performance is a kind of linguistic uncanny valley—close enough to be impressive, but far enough to be unsettling.

Business Implications: The Human Element in Creative Industries

The consequences of this shallow “understanding” ripple far beyond the laboratory. In sectors where nuance is currency—advertising, entertainment, marketing, and customer engagement—the inability of AI to authentically interpret or generate humor is more than a minor glitch. It is a strategic constraint. As organizations integrate LLMs into workflows for content creation and personalized outreach, the risk of tone-deaf or awkwardly literal output grows.

For creative industries, this means hybrid approaches are not just preferable—they are essential. Human oversight remains the safeguard for ensuring that subtlety, wit, and emotional resonance are maintained. The study’s findings reinforce a broader truth: automation, for all its efficiencies, cannot yet replace the unpredictable brilliance of human creativity. Businesses that recognize and invest in this balance will be better positioned to harness AI’s strengths without sacrificing the qualities that make brands memorable and messages meaningful.

The Regulatory and Ethical Frontier: Rethinking AI Reliability

As LLMs become more deeply embedded in critical business and consumer interfaces, the gap between apparent competence and actual understanding is not just a technical quirk—it is a governance challenge. The Cardiff and Ca’ Foscari study raises important questions about AI transparency and accountability. If an AI system can mimic understanding but routinely stumbles in areas like humor or empathy, what does that mean for trust, liability, and user experience?

Regulators and policymakers may soon be compelled to set firmer boundaries on how companies describe and deploy AI, particularly in sensitive domains. Claims of “human-like” understanding could face greater scrutiny, with new standards emerging around transparency, explainability, and limitations. For the business community, this is both a warning and an opportunity: those who lead with clarity and candor about what AI can—and cannot—do will be better equipped to build lasting trust.

Beyond Performance: The Human Future of Artificial Intelligence

The global race for AI supremacy often focuses on speed, scale, and accuracy. Yet, as this research reminds us, the true test of artificial intelligence lies in its ability to enrich human experience. The international community has a stake in fostering not just more powerful AI, but more thoughtful, ethical, and culturally attuned systems. Collaborative frameworks that prioritize both innovation and responsibility could define the next era of AI leadership.

For now, the punchline is clear: until AI can truly share in the laughter, it remains a brilliant imitator—one still learning the most human of arts.