LLMs Generate Kitsch

1. Problem Statement

The paper addresses a growing tension in the field of Generative AI:

• The Phenomenon: Large Language Models (LLMs) are increasingly used to generate creative artifacts (texts, images, code) that, in controlled studies, are frequently rated higher by human users than human-created works.
• The Criticism: Despite their technical competence and popularity, these artifacts are often perceived as "generic," "content-empty," or "AI junk"—superficially beautiful but lacking depth.
• The Gap: Existing literature lacks a unified theoretical framework to explain why LLMs produce content that users simultaneously "like" and perceive as artistically empty. The authors propose that this is not a flaw but a systematic feature: LLMs systematically generate kitsch.

2. Theoretical Framework: Defining Kitsch

The authors ground their argument in art philosophy and psychology, defining kitsch through three specific criteria (based on Kulka, 1996; Dorfles, 1969):

1. Appeal to Standard Emotions: Kitsch evokes emotions in the "grand style" (e.g., love, nostalgia, patriotism) that are universally understandable but unreflected. It relies on "standard emotions" rather than complex, nuanced, or original emotional experiences.
2. Competent but Conventional Surface Form: Kitsch is technically proficient (grammatically correct, coherent) but adheres strictly to established, conventional canons of representation. It avoids stylistic innovations that might confuse the audience, prioritizing "hedonic fluency" (ease of processing) over cognitive challenge.
3. Lack of Artistic Value: Although aesthetically appealing, kitsch is a "counterfeit of art." It reflects the viewer's existing associations back to them rather than offering a new perspective or transforming the viewer's experience. It lacks genuine artistic intent.

3. Methodology

The paper combines a theoretical argument with a novel empirical study.

A. Theoretical Argument (Section 3)

The authors argue that the training objectives of modern LLMs inevitably lead to kitsch:

• No Artistic Intent: LLMs sample from next-token distributions without inner life or original intent.
• Conventional Surface: The goal of next-token prediction rewards frequent continuations in the training data, making stylistic innovations statistically unlikely.
• Optimization for Mass Appeal: Reinforcement Learning from Human Feedback (RLHF) explicitly aligns models to maximize human "likes," which correlates with mass appeal and encourages the avoidance of alienating or challenging content.

B. Empirical Study (Section 4)

To test whether humans perceive LLM outputs as kitsch, the authors conducted a two-phase study:

• Creation of Stimuli:
  • Human Source: 10 short stories (~500 words) were curated from a flash fiction competition.
  • LLM Generation: GPT-4o was instructed to extract the narrative premise from each human story and then generate a new story of similar length based solely on that premise (without seeing the original text). This resulted in 10 pairs of (Human, LLM) stories with similar themes/characters.
• Pilot Study:
  • Participants: 60 distinct participants (485 votes).
  • Task: Rate which story in a pair was "more kitschy" and provide a free-text definition of "kitsch."
  • Result: Definitions varied widely. Participants were clustered into four groups based on their definitions: Emotionally Charged, Hedonic Fluency, Bad Art, and Niche Taste.
• Main Study:
  • Participants: 101 distinct participants (505 votes).
  • Task: Participants were presented with the four definitions, selected the one they agreed with most, and then rated the story pairs regarding kitschiness and enjoyment.
  • Analysis: Mixed logistic regression was used to analyze the relationship between kitsch judgment, enjoyment, and definition groups.

4. Main Results

1. Perception of Kitsch: Controlling for the definition of kitsch, human raters consistently identified LLM-generated stories as more kitschy than human-written ones.
   • Under the "Emotionally Charged" definition (closest to the academic viewpoint), 85.1% of ratings identified the AI story as more kitschy.
   • Under "Hedonic Fluency," 70.6% identified the AI story as more kitschy.
   • Overall, 59.5% of all ratings found the AI story more kitschy.
2. The "Like" Paradox: Despite identifying AI stories as more kitschy, participants liked them more.
   • 67% of all preference ratings favored the LLM story.
   • This preference held across all definition groups.
3. Divergent Effects of Definition:
   • For the "Bad Art" group, identifying a story as kitsch significantly reduced enjoyment (from 90.7% to 28.6%).
   • For the "Hedonic Fluency" and "Emotionally Charged" groups, identifying a story as kitsch increased enjoyment (e.g., +28 percentage points for the fluency group).
   • Conclusion: For many users, "kitsch" (easy to process, emotionally resonant) is synonymous with "enjoyable."

5. Main Contributions

• Theoretical Synthesis: The paper provides the first rigorous argument linking the technical training mechanisms of LLMs (next-token prediction, RLHF) directly to the philosophical concept of kitsch.
• Empirical Validation: It offers the first empirical evidence that humans perceive LLM outputs as kitsch, even when they prefer them.
• Methodological Insight: The study shows that "enjoyment" is a flawed metric for evaluating AI art. High enjoyment scores may actually indicate high kitsch levels (mass appeal) rather than artistic merit.
• Broad Applicability: The authors extend the argument beyond text to other modalities (images, music) and non-artistic domains (programming, scientific research), suggesting that fully automated generation in these areas will also lead to "incremental" or "kitschy" outputs.

6. Significance and Implications

• For AI Research: The results suggest that current RLHF optimization strategies are inherently biased toward generating kitsch. To create "real" art or innovative research, human intervention (inventiveness) is required to break the cycle of conventional, mass-oriented generation.
• For Future Studies: Researchers evaluating AI creativity must go beyond simple preference surveys. They must account for how participants define aesthetic concepts and recognize that "liking" an AI output is not equivalent to artistic value.
• Cultural Impact: The paper warns of a future where "AI junk" dominates cultural consumption because it is optimized to be immediately consumable and emotionally safe, potentially drowning out authentic, challenging, or ambivalent human creativity.

In summary, the paper argues that LLMs generate kitsch by design. Their training to maximize probability and human consent leads to content that is technically competent and emotionally resonant but artistically hollow. The empirical data confirms that while humans enjoy this "kitsch," they can also distinguish it from genuine human creativity when the definition of kitsch is made explicit.