Study Reveals: Large Language Models Are Systematically Producing 'Kitsch'

When AI Works 'Win the Scores but Lose the Soul'

A puzzling contradiction continues to ferment in the field of AI-generated creativity: in controlled experiments, texts, images, music, and videos generated by large language models (LLMs) frequently score higher than human-created works. Yet at the same time, a growing number of people are complaining that these works are "cookie-cutter" and "soulless." A recent paper published on arXiv (arXiv:2604.25929v1) offers a precise diagnosis — LLMs are systematically producing "Kitsch."

What Is 'Kitsch'? And Why Are LLMs Naturally Inclined Toward It?

The concept of "Kitsch" originates from aesthetics and art criticism, widely discussed by thinkers such as Milan Kundera. It refers to works that deliberately cater to popular emotions, appearing polished on the surface but lacking genuine depth and originality. The core characteristic of kitsch is this: it doesn't challenge the audience, doesn't create discomfort, but instead precisely delivers what people "expect to see."

The paper's authors point out that this aligns remarkably well with how LLMs generate content. The essence of large language models is learning statistical patterns from massive amounts of human data, with their generation process being a probabilistic optimization toward "the output most likely to be accepted." This means LLMs are naturally inclined to produce content that conforms to mainstream aesthetic preferences, is emotionally safe, and structurally well-organized — which is, by definition, kitsch.

The Paradox of High Scores and Hollowness, Explained

This theoretical framework elegantly explains the aforementioned contradiction:

Why do LLM works score high? Because kitsch is essentially "optimized flattery." In evaluation scenarios that lack deep scrutiny, works that appear flawless on the surface naturally score higher. Evaluators tend to rate based on dimensions such as fluency, completeness, and emotional resonance — precisely the qualities LLMs are best at simulating.

Why do they still feel hollow? Because truly moving works of art often contain a certain "imperfection" — personal obsessions, provocative edges, unpredictable leaps. These qualities are precisely what gets "averaged out" under the framework of statistical learning. Poetry generated by LLMs always rhymes neatly and features beautiful imagery, but you'll rarely find in it the struggles of a real living being.

Beyond Text: The Comprehensive 'Kitschification' of Multimodal Creation

Notably, the paper's scope is not limited to text generation. As LLM capabilities expand into multimodal domains such as images, music, and video, the trend toward kitsch is equally pronounced:

• AI Painting: Generated works feature perfect composition and harmonious colors, yet a large number exhibit a highly similar "dreamlike" aesthetic
• AI Music: Melodies are pleasant and arrangements professional, yet they rarely produce truly groundbreaking, awe-inspiring musical phrases
• AI Writing: Wording is precise and logic is clear, but readers frequently report it "feels like reading a template article"

This phenomenon has already sparked widespread discussion on social media. Many creators point out that AI-generated content possesses an "uncanny mediocrity" — every individual detail is good, but the whole is entirely unmemorable.

Deeper Implications: How Should We Redefine 'Good Creation'?

The value of this paper lies not only in providing a diagnosis but also in forcing us to re-examine several fundamental questions:

First, the limitations of evaluation systems. If kitsch works generated by LLMs can win in evaluations, it suggests that our existing evaluation criteria themselves carry an inherent bias — they are better at measuring "technical perfection" than "artistic breakthroughs."

Second, the irreducibility of creation. True creativity may not be a goal that can be approximated through big data statistics. It encompasses intention, experience, risk-taking, and a certain "human noise" that probabilistic models cannot capture.

Third, the boundaries of human-AI collaboration. If the natural output of LLMs is kitsch, then the core value of human creators in human-AI collaboration lies in injecting "anti-kitsch" elements — personal style, controversial viewpoints, and the courage to break conventions.

Looking Ahead: Beware of 'Aesthetic Entropy'

As AI-generated content continues to grow as a proportion of internet content, a more far-reaching concern emerges: when LLMs continue to train on data generated by LLMs, will kitsch keep reinforcing itself? This "aesthetic entropy" could lead to a continuous decline in the richness and diversity of the entire digital content ecosystem.

The paper sounds an alarm — improvements in AI's creative capabilities do not equate to progress in creativity itself. While celebrating LLMs' productivity, we may need even more to safeguard those "imperfect" but truly vibrant forms of human expression. After all, the essence of art has never been about pleasing everyone, but about profoundly moving someone.