LLMs Generate Kitsch

Large language models (LLMs) and other forms of Generative AI are increasingly being used to create content that traditionally required human creativity. GPT 3.5 generates poetry that is rated by humans as highly as works by famous poets (Porter & Machery 2024); over half of all photos available on Adobe Stock are self-reported as AI-generated; and GPT-5 is being used to generate proof ideas in research on complexity theory.

At the same time, AI-generated art often leaves behind a sense of emptiness or deception in its viewers and readers; it feels “safe” and formulaic. For instance, the AI-generated poem below recognizably emulates the style of Lord Byron, but it consists of a sequence of shallow tropes. In this paper, I want to offer a useful concept for thinking about the difference between art that is produced by humans and by LLMs: Generative AI generates kitsch.

Below, I summarize the main points of the paper. Please have a look at the full paper (linked above) for my more nuanced position. I would be happy to receive your feedback.

What is Kitsch?

This postcard from 1901 competently evokes a stock emotion. Image from Wikipedia Commons.

Kitsch is hard to define; the most specific definition I could find is from Kulka (1996). Instead of focusing just on a “kitschy surface form”, Kulka argues that kitsch is designed for mass appeal, at the expense of artistic originality. He identifies three core characteristics of kitsch: It appeals to stock emotions (in the picture above, the happiness of an innocent childhood); it depicts a subject that evokes these emotions in a way that is instantly recognizable; and the work of kitsch does not substantially enrich the viewer’s inner life with respect to the subject. These definitions can apply to any artform, ranging from the visual arts over texts and poetry to music and architecture.

LLMs generate kitsch

An AI-generated poem in the style of Lord Byron, from Porter & Machery (2024).

One can measure LLM-generated art by Kulka’s standards. I assume as given and trivial that AI-generated art is not driven by an artistic intention; it is generated by a model without an inner life of its own that it could convey through the artwork. I argue further that AI-generated art has competent, but conventional surface forms; in fact, LLMs are trained to give high probability to the conventional forms of expression that make up their training data. Furthermore, the mass appeal of AI-generated art is not an accident; it is the successful result of RLHF training.

Perhaps the most interesting question to me, because it affects my own profession, is how to think about creative acts that are too utility-driven to be considered art, such as programming and doing research. The example of the proof idea I mentioned above shows that there can be value in including LLMs in such activities. But if fully automated research is given into the hands of an LLM, we should expect the equivalent of kitsch: safe and incremental research. I would personally prefer to reserve peer-reviewed publications for papers that involve a spark of human creativity.

Final note

Why is this paper here and not on Arxiv? I submitted this paper to ArXiV on 15 October 2025, not realizing that ArXiV had just decided to enforce their policy of not accepting opinion papers. The paper got stuck in a loop for a few weeks; now I finally had the time to create a blog page for the paper under my own control (12 November 2025).

References

Kulka, T. (1996). Kitsch and Art. Penn State University Press.

Porter, B., & Machery, E. (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports, 14(26133).

BibTeX

@misc{koller2025kitsch,
      title={LLMs Generate Kitsch},
      author={Alexander Koller},
      year={2025},
      howpublished={Manuscript},
      url={https://coli-saar.github.io/kitsch}, 
}