LLMs Generate Kitsch
LLMs beat humans in quality tests, but a new paper labels them 'kitsch'...
Deep Dive
The paper "LLMs Generate Kitsch" by Xenia Klinge, Stefan Ortlieb, and Alexander Koller argues that LLMs systematically generate kitsch as a consequence of their training. The authors show empirically that readers perceive LLM-generated stories as kitschier when controlling for their definition of "kitsch." The paper discusses implications for future studies and creative tasks like research and coding.
Key Points
- Saarland University researchers argue LLMs generate kitsch due to training on averaged data, producing generic outputs.
- In a controlled study, readers rated LLM-generated stories as kitschier than human-written ones.
- The paper, submitted to EMNLP 2026, warns of implications for creative tasks like research and coding.
Why It Matters
LLMs may be efficient but lack creativity, risking homogenized outputs in professional and creative work.