Skip to content
conftrace_
2026 EMNLP EMNLP 2026

How Do Self-Referential LLM Loops Contract Political Diversity?

AbstractAs Large Language Models (LLMs) increasingly consume their own outputs through retrieval-augmented generation and synthetic data pipelines, self-referential loops are becoming prevalent in modern information ecosystems. While prior work has studied variance collapse during recursive training, the dynamic consequences for political diversity during inference remain underexplored. We investigate this by conducting iterative paraphrasing experiments across eight widely used LLMs, and formalize self-referential generation as a discrete-time Markov process to characterize its long-term behavior. Our results reveal a systematic contraction of political diversity, with several models converging toward center-dominant equilibria. Notably, this homogenization is strongly front-loaded: the majority of diversity loss occurs within the earliest iterative rounds, indicating that even a few AI-mediated transformations can disproportionately narrow the political spectrum. We further examine two interventions: Stance-Conditioned Generation (SCG) substantially reduces marginal stance-distribution drift even without ground-truth labels, leaving no statistically detectable shift for three of the four homogenizing models, whereas Self-Identified Stance Preservation (SISP) overshoots into over-polarization; even SCG improves but does not guarantee instance-level stance retention. These findings highlight a structural risk associated with self-referential generation in AI-driven information pipelines. Our code and prompts are publicly available at https://github.com/timothy-coshin/SelfReferentialPoliticalDiversity.

More from EMNLP 2026

All EMNLP 2026 papers →