LLMs Generate Kitsch

2026-04-30 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

A paper submitted to EMNLP 26 by Xenia Klinge, Stefan Ortlieb, and Alexander Koller, titled "LLMs Generate Kitsch," proposes that Large Language Models (LLMs) systematically produce kitsch due to their training methodology. While LLM-generated content, including pictures, texts, music, and videos, often receives higher ratings than human-generated works in controlled studies, it can also appear generic and hollow. The authors argue that this tension is resolved by classifying LLM output as kitsch. Empirical evidence is presented showing that readers perceive LLM-generated stories as kitschier, even when controlling for their individual definitions of "kitsch." The paper discusses the implications of this finding for the design of future studies and for creative applications such as research and coding.

Key takeaway

For AI Product Managers evaluating creative content generation, you should consider the inherent "kitsch" factor in LLM outputs. This suggests that while LLMs may produce technically proficient or highly-rated content, its perceived originality or depth might be compromised. Incorporate metrics beyond simple preference ratings to assess the nuanced artistic value and avoid generic results in your product offerings.

Key insights

LLMs systematically generate kitsch due to their training, despite often outperforming human-generated works in ratings.

Principles

• LLM training leads to generic outputs.
• Perceived quality differs from perceived kitsch.

Method

The study empirically demonstrated that readers perceive LLM-generated stories as kitschier, controlling for individual definitions of kitsch, to support the hypothesis.

In practice

• Re-evaluate LLM output quality metrics.
• Consider "kitsch" in creative AI applications.

Topics

• Large Language Models
• Kitsch Generation
• Creative AI
• Empirical Evaluation
• Model Training Bias

Best for: Research Scientist, AI Product Manager, AI Scientist, NLP Engineer, Creative Technologist

Related on AIssential

• Exploring conversational AI and poetry but not as we know it — LLMs challenge traditional notions of language, creativity, and authenticity, particularly in fields like poetry.
• The TechBeat: Can LLMs Generate Quality Code? A 40,000-Line Experiment (1/6/2026) — LLMs generate sloppy code over time, requiring formal analysis and multi-model reviews for quality assurance.
• Granite 4.1 3B SVG Pelican Gallery — Quantization level does not guarantee improved output quality for specific creative tasks.
• On the Limits of LLM-as-Judge for Scientific Novelty Assessment — LLMs are unreliable judges of scientific novelty for research questions, often creating a "novelty mirage."
• Yes, AI, There is a Santa Claus — LLMs model user age and culture, adjusting responses in ways that can be inconsistent or culturally inaccurate.
• I created an LLM trained solely on Jeffrey Epsteins emails to see how messed up it becomes :) — Fine-tuning an LLM on a controversial, unannotated dataset can reveal unique behavioral shifts and spark technical debate.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.