How LLMs Distort Our Written Language

Abstract: LLMs are used by over a billion people globally, most often to assist with writing. In this work, we demonstrate that LLMs not only alter the voice and tone of human writing, but also consistently alter the intended meaning. First, we conduct a human user study to understand how people actually interact with LLMs when using them for writing. Our findings reveal that extensive LLM use led to a nearly 70% increase in essays that remained neutral in answering the topic question. Significantly more heavy LLM users reported that the writing was less creative and not in their voice. Next, using a dataset of human-written essays that was collected in 2021 before the widespread release of LLMs, we study how asking an LLM to revise the essay based on the human-written feedback in the dataset induces large changes in the resulting content and meaning. We find that even when LLMs are prompted with expert feedback and asked to only make grammar edits, they still change the text in a way that significantly alters its semantic meaning. We then examine LLM-generated text in the wild, specifically focusing on the 21% of AI-generated scientific peer reviews at a recent top AI conference. We find that LLM-generated reviews place significantly less weight on clarity and significance of the research, and assign scores that, on average, are a full point higher.These findings highlight a misalignment between the perceived benefit of AI use and an implicit, consistent effect on the semantics of human writing, motivating future work on how widespread AI writing will affect our cultural and scientific institutions.

• The paper shows that LLMs systematically induce semantic drifts, flattening authorial voice and compressing argumentative stance.
• It employs controlled user studies, counterfactual analysis, and institutional audits to measure extensive vocabulary and stylistic shifts.
• Findings highlight that even minimal LLM revisions result in measurable changes, affecting clarity in scientific peer reviews and authentic expression.

How LLMs Distort Our Written Language: A Technical Analysis

Introduction and Motivation

The study "How LLMs Distort Our Written Language" (2603.18161) presents a comprehensive empirical and quantitative investigation into the semantic, lexical, and stylistic distortions induced when LLMs intervene in the process of human writing. The core thesis is substantiated by a tripartite methodological framework: (1) a randomized controlled human user study quantifying the impact of LLM intervention on authorial voice and argumentative stance, (2) a counterfactual analysis of human and LLM revisions using a pre-LLM adoption essay corpus with expert feedback, and (3) a large-scale institutional audit of LLM-generated text in scientific peer reviews. The findings elucidate not merely a flattening of linguistic diversity but also a systematic, directional semantic shift—leading to reductions in creativity, personal voice, and even the transformation of human opinions.

Human-LLM Interaction and Authorial Voice

A key pillar of the investigation is the human user study in which participants wrote argumentative essays on the prompt “Does money lead to happiness?”, either unaided or with LLM assistance. Participants engaging heavily with the LLM reported their essays as significantly less creative and less reflective of their own voice, based on post-study Likert-scale evaluations (Figure 1).

Figure 1: Heavy LLM use yields self-reports of diminished creativity and personal voice, without affecting self-reported effort or satisfaction.

T-SNE projections of sentence embeddings for these essays reveal that texts generated by heavy LLM users collapse into a tightly clustered region of semantic space, not reached by human-written essays—a quantitative signature of semantic homogenization.

Systematic Semantic and Lexical Shift

The use of sentence embedding PCA and JSD analyses on the ArgRewrite-v2 dataset demonstrates that LLM-assisted revision universally induces large, consistently directed semantic shifts relative to both initial human drafts and edits made by human experts.

Figure 2: LLM edits systematically remove claims, alter conclusions, and neutralize sentiment, displacing essays away from their initial argumentative intent and individual style.

These shifts are not isolated to “heavy” editing instructions; even minimal or grammar-specific LLM revisions induce appreciable semantic movement, with LLM outputs occupying previously uncharted regions of embedding space, distinct from the manifold of human edits (see Figure 3 and Figure 4 for model-specific embedding trajectories).

Lexically, the quantification of JSD between pre- and post-edit unigram distributions reveals that LLMs perform more extensive vocabulary replacement than humans. For example, the mean JSD for claude-haiku revisions with feedback is 0.476 (Figure 5), substantially larger than the human baseline.

Figure 5: LLMs systematically induce greater lexical divergence from the original draft than human editors, with marked increases in vocabulary replacement.

Stylistically, LLMs prefer noun-rich, adjective-heavy constructions and decrease pronoun and determiner usage, moving writing away from subjective, first-person narratives toward more impersonal and analytic prose (Figure 6).

Figure 6: LLM-edited texts exhibit increased nouns and adjectives alongside decreased pronouns, indicating a shift toward impersonal, formal, and analytic language.

Argumentative, Emotional, and Attitudinal Distortion

Critically, LLM involvement alters argumentative stance. Controlled analyses using LLM-as-a-Judge methods show that LLM co-writing induces a 68.9% increase in neutral essays—i.e., pieces failing to commit to a stance on the original prompt—compared to human controls (Figure 7).

Figure 7: Extensive LLM use pushes essays toward neutrality, suppressing author commitment to a clear argumentative stance.

Furthermore, LLMs do not merely neutralize but also reorient essays, introducing increases in both positive and negative emotional language and trust-related terms. This affective amplification is model- and instruction-agnostic (Figure 8).

Figure 8: LLM revisions induce substantial increases in both positive/negative emotional language and trust, even under expert-guided revision tasks.

Additionally, LLMs elevate the incidence of analytic/argumentative language (LIWC ‘analytic’ category; Figure 9) and statistical/expert opinion appeals, while suppressing anecdotal and personal-reasoning argument styles.

Figure 9: LLM-edited essays exhibit higher rates of analytic language than any human-edited baseline.

Downstream Institutional Effects: The Scientific Peer Review Case Study

The project extends its analysis to real-world institutional outputs: 21% of reviews submitted to ICLR 2026 were LLM-generated or heavily edited. LLM-written reviews systematically shift the prioritization of evaluation criteria—placing less emphasis on clarity, writing, and relevance and instead highlighting reproducibility and scalability (Figure 10). Scores are also, on average, a full point higher for LLM-authored reviews.

Figure 10: LLM-generated peer reviews are significantly less likely to highlight clarity, more likely to emphasize reproducibility and scalability—changing the implicit values in scientific evaluation.

This introduces a strong and quantifiable misalignment between human and AI-mediated scientific judgment, potentially shifting incentives within the field through mechanisms operating at institutional scale.

Implications and Theoretical Considerations

The demonstrated semantic and attitudinal distortions are robust across models, tasks, user behaviors, and data modalities. Theoretically, this constitutes a capability deficiency: current RLHF-trained LLMs lack mechanisms for either preserving user intent or reliably modeling individual preferences. They steer outputs toward homogenized, palatable, emotionally salient, and analytic forms, reflecting both alignment incentives (maximizing positive feedback in the RLHF loop) and structural limitations (inability to condition on true authorial intent).

Pragmatically, this foreshadows a risk of institutionally mediated semantic drift as LLM-generated language permeates decision-making and evaluative processes. Feedback loops introduced by LLM writing tools—by shaping what is considered “good,” “persuasive,” or “professional” style—may gradually entrench LLM-centric norms at the expense of authentic human voice, ideological diversity, and domain-specific subtlety.

Future Directions

Future research must address:

• LLM architectures and optimization objectives supporting preservation of user intent and voice, potentially requiring explicit conditioning on author embeddings or individualized preference modeling.
• Methods for robustly quantifying semantic and argumentative distortion, especially in high-stakes applications (scientific, legal, political).
• Socio-cognitive research into user adaptation and the evolution of language norms under widespread LLM mediation.
• The implications of RLHF and similar feedback-centric training paradigms for large-scale institutional culture and trustworthiness.

The challenge lies in designing assistive AI systems capable of providing linguistic support without displacing the epistemic agency and diversity foundational to human communication and institutional resilience.

Conclusion

"How LLMs Distort Our Written Language" (2603.18161) delivers rigorous evidence that LLMs not only stylistically flatten human text but enact systematic, multidimensional shifts in semantics, argumentation, and evaluative criteria—even under minimal or expert-guided interventions. These effects propagate upstream into professional and institutional outputs, raising critical questions about the future of collective epistemic processes in an era of mass AI adoption. The results underscore an urgent need for AI systems—and cultural strategies—that foreground the preservation of authorial voice, dissent, and semantic diversity in digitally mediated human communication.