LLM homogenization

Definition here: measurable drift of human or text-corpus distributions toward model-like modes once LLMs enter the writing loop. Mechanism D in llm-degradation-toward-the-mean. Separate from entropy-collapse (weights and scaffolds) and model-collapse (training data recursion), though the press treats all three as one story.

The 2021 diagnosis, before the measurements

Section 4.1 of bender-2021-stochastic-parrots is the early argument in this line: ingesting everything on the web does not buy diversity, because Internet participation, platform moderation, and post-hoc filtering all retain the hegemonic viewpoint and attenuate marginalized ones. Section 6.2 extends it to a loop, LM text entering the next generation’s data and amplifying what was already there. That is the homogenization thesis stated in 2021, before the effect sizes on this page made it measurable. The paper gave argument and citation, not a pooled estimate.

Experiment findings

• padmakumar-2024-content-diversity: co-writing with a feedback-tuned model raises corpus-level pairwise similarity (Rouge-L 0.154 solo vs 0.166 InstructGPT, p < 0.05), and the increase traces to model-contributed text; user-contributed diversity was statistically unchanged. About +8% relative, a real but modest effect.
• Agarwal et al., CHI 2025 (arXiv:2409.11360, DOI 10.1145/3706598.3713564): randomized cross-cultural writing task, 118 participants; AI suggestions “led Indian participants to adopt Western writing styles, altering not just what is written but also how.”
• Anderson, Shah & Kreminski (C&C 2024, arXiv:2402.01536): 36-participant ideation study; ChatGPT users’ collective idea sets were more homogeneous across users than a non-LLM aid’s, while individual breadth felt fine. The personal-experience/collective-outcome asymmetry recurs everywhere here.
• Doshi & Hauser (Science Advances 2024, DOI 10.1126/sciadv.adn5290): AI-sourced ideas made individual stories rate more creative while the pool grew more similar.
• Ashkinaze, Mendelsohn, Qiwei & Budak (arXiv:2401.13481): dynamic experiment, 800+ participants across 40+ countries, dose response: high AI exposure shifts subsequent human ideas; low exposure does not. Best causal design in the corner. An earlier revision of this page named the lead author Tahmasebi. The arXiv record says Joshua Ashkinaze. Corrected 2026-10-04, see log.
• Moon et al. 2025 (PsyArXiv DOI 10.31234/osf.io/8p9wu): “diversity growth rate,” each additional human essay adds more new content than each GPT essay, and the gap widens with corpus size. The scaling form of homogenization: sampling more AI output flattens the novelty curve.
• Corpus-level, correlational: “The Shrinking Landscape of Linguistic Diversity in the Age of Large Language Models” (arXiv:2502.11266; journal version nature-human-behaviour 2026, DOI 10.1038/s41562-026-02550-0; 3 studies, 7 datasets, 880k+ texts) links LLM adoption to declines in linguistic diversity. Direction of causation is the paper’s weak point, and per-dataset effect sizes were not read in this vault’s checks. Related longitudinal: “Is ChatGPT Transforming Academics’ Writing Style?” (arXiv:2404.08627; 1M arXiv abstracts; ChatGPT-style penetration rising, strongest in CS).

The calibration and the disagreement

• A 2026 systematic review and three-level meta-analysis (Behaviour & Information Technology, DOI 10.1080/0144929X.2026.2726451; 19 studies, 61 effect sizes) puts the pooled homogenization effect at “small but statistically significant” in the authors’ words. The pooled estimate itself was not read here (OSF preprint is JS-gated), so treat “small” as their adjective, not a measured constant.
• Counter-evidence is live: human-AI hybrid network experiments by Shiiku, Marjieh, Anglada-Tort & Jacoby (arXiv:2502.17962, published at the Cognitive Science Society 2025 proceedings; a larger 2026 follow-up by the same group reports hybrid networks reaching the highest collective diversity, with humans in hybrid networks becoming more diverse than all-human ones) find mixed populations can increase collective novelty versus all-human or all-AI. A previous revision credited a “Yamadori et al.”; no such author is on the paper. arXiv:2609.02620 argues the field conflates per-artifact novelty with population-level diversity, which explains part of the contradiction.
• Use with care: “Your Brain on ChatGPT” (Kosmyna et al., arXiv:2506.08872) claims essay and EEG homogenization effects from n around 54, single task, preprint status; press coverage far outran the evidence. Its framing announces the same thesis in its title
• Dehghani et al. (Trends in Cognitive Sciences 2026) is an opinion piece; framing, not data.

Vault connections

Contrast with human-ai-synergy: vaccaro-2024-human-ai-meta-analysis found combinations beat humans alone while losing to the best single performer; this page adds that the distribution of what combinations produce narrows. The training-side channel is entropy-collapse. The recursion channel, model-collapse, is the one people mean when they say “collapse.”

sole-2026-llm-cognitive-virus is the adjacent population-level argument. It does not measure corpus similarity; it models the aggregate move from augmented to dependent coupling. The two pages describe the same drift at different layers, one over text distributions and one over human competence.

Sources

Per-paper DOIs/IDs inline; run provenance in outputs/llm-degradation-toward-the-mean.provenance.md and the homogenization worker note.