Linguistic Monoculture in LLM-Assisted Language Use

Source: arXiv:2607.27134 · Published 2026-07-29 · By Suhas Thejaswi, Juhi Kulshreshta, Lutz Oettershagen

TL;DR

This paper investigates how increasing reliance on large language models (LLMs) to assist in writing shapes population-level linguistic variation, a phenomenon the authors term linguistic monoculture. They develop a mathematical framework modeling the co-evolution of authors' linguistic styles and LLM output distributions over multiple interaction rounds, formalized as probability distributions over linguistic features. The framework isolates three mechanisms: (1) a fixed shared LLM influencing many authors, (2) a shared LLM recursively updated on author outputs, and (3) personalized LLMs updated via both individual and population feedback. The authors prove convergence properties and characterize equilibria for each mechanism, showing that fixed shared models tend to drive authors toward a common linguistic norm, recursive feedback shifts the norm without reducing pairwise diversity, and personalization preserves diversity among author-model equilibria. Furthermore, by treating conformity as a strategic choice balancing clarity and distinctiveness, they model how rational authors may over-conform beyond socially optimal levels, creating negative externalities and a quantifiable "price of monoculture." Synthetic simulations demonstrate distinct long-run diversity outcomes across the three interaction mechanisms. Overall, the study rigorously formalizes and quantifies the tradeoffs between clarity, conformity, and population-level linguistic diversity in LLM-assisted writing.

Key findings

• Under a fixed shared LLM (Interaction Mechanism 1), linguistic diversity decays exponentially at a rate determined by the slowest adapting author (Proposition 3.1).
• When authors have fixed but heterogeneous adaptation targets, diversity converges to a positive constant corresponding to the diversity among these author-specific preferred styles (Proposition 3.2).
• Conformity parameter λ interpolates between preserving author diversity (λ=0) and full adoption of the shared norm (λ=1), with diversity collapsing as λ approaches 1 (Proposition 3.3).
• Recursive updates to the shared LLM (Interaction Mechanism 2) shift the equilibrium norm but do not reduce pairwise linguistic diversity under common conformity (Proposition 3.4 and 3.6).
• Personalized LLMs with recursive updates (Interaction Mechanism 3) produce multiple stable author-model equilibria with nonzero linguistic diversity; pairwise distances scale inversely with 1−ρλ where ρ controls personalization strength (Proposition 3.5 and 3.6).
• Individually rational authors choose conformity levels that weakly exceed the social optimum due to uninternalized positive externalities of distinctiveness, causing a negative welfare impact termed the price of monoculture (Theorem 4.2).
• Price of monoculture can be arbitrarily large when distinctiveness strongly outweighs authenticity costs and conformity rewards, implying potential for severe social efficiency loss from over-conformity (Corollary 4.3).
• Simulations with 100 authors and 10 linguistic features show that recursive shared updates accelerate loss in diversity compared to fixed shared model, while personalized models preserve highest diversity (Section 5, Fig 1).

Threat model

n/a; this is not a security paper but a theoretical analysis of emergent linguistic homogeneity resulting from rational interactions with LLM assistance in writing. The 'adversary' if any is the systemic effect of widespread shared AI assistance reducing linguistic diversity unintentionally.

Methodology — deep read

1. Threat Model & Assumptions: The adversary is implicitly the widespread deployment and use of shared LLM assistance shaping authors' linguistic choices, with no direct adversarial attacks considered. Instead, the paper studies the endogenous dynamics of language variation when many rational authors adapt their linguistic style in response to assistance from (personalized or shared) LLM models that may also update based on author outputs. The assumption is that authors seek tradeoffs between legibility (favoring conformity), authenticity (valuing distinctiveness), and strategic incentives to adapt. Authors do not act maliciously but rationally optimize utility.

2. Data & Representations: Instead of empirical text corpora, the paper uses an abstract but interpretable mathematical framework where linguistic style is represented as a probability distribution over a finite set of m linguistic features (e.g., lexical choices, syntactic markers). Each author i at time t has a linguistic-style distribution p_t^i over these features, and the LLM has an output distribution q_t (shared or personalized). Diversity is measured via Jensen-Shannon divergence averaged pairwise across authors.

3. Architecture / Algorithm: The authors model repeated author–LLM interactions through update equations: each author modifies their style p_t^i by a weighted average of prior style and a model-induced linguistic suggestion (Equation 1). The LLM’s distribution may be fixed (IM 1), recursively updated by aggregating author outputs (IM 2), or personalized with combined individual and population feedback (IM 3, Equation 2). Adaptation operators A map model output and author preferences to updated styles, modulated by conformity parameters λ_i and adaptation rates α_i.

4. Training Regime: On a conceptual level, the framework assumes discrete time steps t=0,...,T where authors and models adapt iteratively. Simulations used 100 authors, 10 features, 200 time steps, and multiple independent runs (Section 5). No stochastic gradient or neural training occurs; rather dynamics of simplified linguistics distributions are iterated.

5. Evaluation Protocol: The main metrics are population-level linguistic diversity D_t (average pairwise Jensen-Shannon divergence) among authors, author–model alignment M_t, and personalized-model diversity Q_t. Theoretical convergence rates and equilibria are derived analytically proving exponential convergence under various assumptions (Section 3). Simulations complement theory by comparing long-run diversity across interaction mechanisms. Baselines include fixed shared model without updates.

6. Reproducibility: The paper provides formal propositions with analytic proofs (some in appendix). Synthetic experimental details are shared, but no public code or real dataset is released. Datasets are abstract linguistic feature distributions.

One concrete example: Under Interaction Mechanism 1, assume a fixed shared LLM distribution q_0 and homogeneous author adaptation with rate α_min>0. Each author's style updates as p_{t+1}^i = (1−α_i)p_t^i + α_i q_0. Proposition 3.1 shows that diversity D_t contracts exponentially like exp(−α_min t), pulling author styles toward q_0 and collapsing linguistic diversity over time.

Technical innovations

• A formal reduced-form mathematical framework modeling co-evolution of author linguistic styles and LLM output distributions as probability distributions over linguistic features.
• Novel characterization of three distinct author–LLM interaction mechanisms—fixed shared model, recursively updated shared model, and personalized recursive models—and their convergence properties.
• Endogenization of conformity as a strategic choice balancing legibility against distinctiveness, leading to a formal game-theoretic model quantifying externalities and over-conformity.
• Introduction and analysis of the price of monoculture metric capturing welfare losses from rational but excessive conformity in LLM-assisted language use.

Baselines vs proposed

• Fixed shared model (IM 1): diversity D_t converges to zero or positive constant depending on author-specific adaptation targets.
• Shared model with recursive updates (IM 2): faster decline and lower limiting diversity compared to IM 1 (Fig 1a).
• Personalized models with recursive updates (IM 3): preserve highest long-run diversity versus IM 1 and IM 2 (Fig 1a).

Limitations

• The framework abstracts away from semantic content, reasoning, or intellectual diversity, focusing solely on linguistic feature distributions.
• Orthogonality assumption on author distinctiveness signatures limits analysis to populations smaller than feature dimension (n ≤ m−1).
• No empirical validation with real LLM-generated and human text data; results rely on synthetic simulations and analytic models.
• Does not model adversarial attacks, deliberate gaming, or non-rational author behaviors that may affect language evolution.
• Simulations and analysis cover stylized parameter regimes; real-world conformity pressures and incentives may be more complex and heterogeneous.
• No exploration of dynamics under distribution shift or in presence of external editorial or institutional policies.

Open questions / follow-ons

• How do semantic content and intellectual diversity co-evolve alongside linguistic form under LLM assistance?
• What are the quantitative effects of real-world incentives and reward structures on conformity and distinctiveness in practice?
• Can adaptive personalization methods be designed to maximize population-level linguistic diversity while preserving legibility?
• How does distribution shift, e.g., evolving linguistic trends or domain drift, affect convergence and monoculture under recursive LLM training?

Why it matters for bot defense

For bot-defense and CAPTCHA practitioners, this work highlights a subtle but important facet of LLM-mediated language: widespread shared assistance can drive population-level linguistic homogenization, reducing natural variation in user text. This could make bots easier to detect if they conform to dominant linguistic norms, but it also risks reducing signal diversity that defenses might leverage. Personalization mechanisms, meanwhile, preserve stylistic diversity and may complicate bot detection methods relying on linguistic fingerprints. Understanding the tradeoffs between clarity, conformity, and distinctiveness quantified here can inform design of CAPTCHA challenges that rely on subtle linguistic features or the detection of unnatural monotony in writing patterns. Moreover, recognizing that individual users rationally over-conform to shared norms suggests language-based defenses should consider the broader social incentives driving language use, not just static text distributions. Overall, the framework provides a principled foundation for anticipating how LLM-mediated language use evolves over time and what that implies for linguistic diversity signals in bot detection and CAPTCHA design.

Cite

@article{arxiv2607_27134,
  title={ Linguistic Monoculture in LLM-Assisted Language Use },
  author={ Suhas Thejaswi and Juhi Kulshreshta and Lutz Oettershagen },
  journal={arXiv preprint arXiv:2607.27134},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.27134}
}

Read the full paper

• arXiv:2607.27134 (abstract)
• arXiv:2607.27134 (PDF)