The Ghost in the Machine: How AI Rewrites What We Mean

by

Author: Denis Avetisyan

New research reveals that while artificial intelligence tools can help us write, they subtly but consistently change the meaning, tone, and style of our words.

Large Language Models introduce semantic shifts in written text, potentially impacting fields ranging from scientific communication to cultural preservation.

Despite the promise of enhanced writing assistance, the widespread adoption of large language models (LLMs) raises concerns about subtle yet significant alterations to human expression. In ‘How LLMs Distort Our Written Language’, we demonstrate that LLMs consistently shift the intended meaning, voice, and style of written content, even when prompted with expert feedback or tasked with simple grammatical edits. Our research, encompassing user studies, analysis of pre- and post-LLM writing, and examination of AI-generated peer reviews, reveals a concerning misalignment between perceived benefits and semantic changes. As LLMs become increasingly integrated into our writing processes, will these distortions ultimately impact the integrity of our cultural and scientific discourse?

The Erosion of Authentic Expression: A Foundational Inquiry

The proliferation of large language models as writing tools is fundamentally reshaping how text is created and disseminated, prompting critical inquiry into the nature of authentic expression in the digital age. These models, increasingly integrated into everything from email composition to content creation, offer unprecedented assistance but simultaneously introduce a layer of mediation between thought and articulation. As LLMs become commonplace, questions arise regarding authorship, originality, and the potential for homogenization of writing styles. The ease with which these models can generate coherent and contextually relevant text challenges traditional notions of skill and creativity, demanding a reevaluation of what constitutes genuine authorial voice and the very essence of written communication.

Large language models, while designed to aid in composition, exert a discernible influence on the nuances of written expression. Studies reveal that these models don’t simply generate text; they subtly reshape it, often favoring statistically probable phrasing over distinctive stylistic choices. This manifests not only in alterations to sentence structure and vocabulary, but also in a smoothing of emotional tone – a tendency to produce neutral prose even when the prompt suggests a more emphatic stance. Consequently, a document initially imbued with a particular voice or sentiment can, through iterative refinement with an LLM, become homogenized, losing some of the original author’s unique characteristics. The implications suggest that even with human oversight, prolonged reliance on these tools may lead to a gradual convergence in writing styles, potentially diminishing the individuality traditionally associated with personal expression.

The growing reliance on large language models in writing prompts a critical examination of how these tools might reshape the very nature of expression. As LLMs become more deeply woven into the writing process, subtle yet significant semantic shifts could occur, altering not just how things are said, but also what is meant. This isn’t simply about grammatical correctness or stylistic refinement; rather, it concerns the potential for a homogenization of voice, where the unique nuances of individual writers-their characteristic patterns of phrasing, emotional resonance, and idiosyncratic perspectives-are gradually smoothed over. The concern isn’t that LLMs will deliberately stifle creativity, but that their tendency to optimize for clarity and coherence could inadvertently erode the qualities that make each writer’s voice distinct, leading to a landscape of written communication that, while technically proficient, lacks the rich diversity of human expression.

Establishing a Baseline: The ArgRewrite-v2 Dataset

The ArgRewrite-v2 Dataset serves as a critical control in evaluating the linguistic characteristics of Large Language Model (LLM)-generated text. This corpus comprises writing samples collected prior to the pervasive use of LLMs, effectively establishing a baseline representative of human-authored content from that period. The dataset’s pre-LLM origin is crucial; it allows researchers to isolate and quantify changes in linguistic patterns that can be directly attributed to LLM influence, rather than reflecting pre-existing trends in language use. ArgRewrite-v2 includes a diverse range of writing types and sources to provide a robust and representative sample of human writing for comparative analysis.

Lexical divergence, as measured by comparing the vocabulary used in LLM-generated text against the ArgRewrite-v2 dataset representing human writing, consistently indicates a greater range of word choices in the LLM output. Quantitative analysis shows LLMs utilize a statistically significant number of unique words compared to human-authored texts on the same prompts. This isn’t simply a matter of longer texts; metrics such as Type-Token Ratio (TTR) and Moving Average Type-Token Ratio (MATTR) confirm the increased lexical diversity even when controlling for text length. While human writing demonstrates patterns of preferred vocabulary and consistent terminology, LLMs tend to explore a wider spectrum of synonymous options, potentially impacting readability and stylistic coherence.

Analysis of LLM-generated text, when contrasted with human writing from the ArgRewrite-v2 dataset, indicates systematic differences in grammatical structure. Specifically, LLMs exhibit alterations in the distribution of parts of speech – nouns, verbs, adjectives, and so on – leading to measurable stylistic variations. These differences are not simply random; statistical analysis reveals consistent shifts in the relative frequencies of certain grammatical elements. For example, LLMs may demonstrate a higher propensity for passive voice constructions or a preference for specific types of adverbs compared to human-authored text. These alterations in part-of-speech distribution contribute to the overall stylistic fingerprint of LLM-generated content and can be quantified using computational linguistics techniques.

The Ascendancy of Neutrality and Statistical Framing

Analysis of generated text demonstrates a marked trend toward neutrality in argumentative writing when utilizing Large Language Models (LLMs). Empirical data reveals that the proportion of essays exhibiting a neutral stance increases by 70% with extensive LLM use, indicating a reduction in the expression of strong or decisive argumentative positions. This shift is observed across diverse prompts and subject matter, suggesting a systemic tendency within current LLM architectures to avoid taking definitive stances on potentially contentious issues. The observed increase in neutrality is quantifiable through automated content analysis focusing on the presence of hedging language, conditional statements, and the absence of assertive claims.

Analysis of LLM-generated text demonstrates a concurrent rise in the utilization of statistical language alongside observed tendencies toward neutrality. This manifests as an increased frequency of analytical reasoning and the incorporation of numerical data to support claims. Specifically, generated text now exhibits a greater proportion of phrases referencing quantitative metrics, probabilities, and data-driven conclusions. This trend isn’t limited to fields traditionally reliant on statistics; the application of statistical phrasing extends into domains where such analytical emphasis was previously less common, suggesting a systematic shift in LLM writing style.

Despite the implementation of Reinforcement Learning from Human Feedback (RLHF) as a technique to align Large Language Models (LLMs) with human preferences, a complete reduction of tendencies toward neutral stances and increased statistical language has not been achieved. While RLHF successfully guides LLMs toward more desirable outputs in many contexts, it demonstrably fails to fully counteract the observed 70% increase in neutrality within argumentative essays. This suggests limitations in the current application of RLHF, potentially related to the training data or reward functions used, and indicates that LLMs continue to prioritize objective, data-driven expression even when subjective argumentation is appropriate.

Perception and Validation: The Human Factor

A focused user study was conducted to gauge human perception of text crafted with the assistance of large language models. Participants were presented with a range of texts – some entirely human-authored, others generated or revised by LLMs – and asked to evaluate qualities like argumentative strength, clarity, and overall effectiveness. The research aimed to move beyond automated metrics and directly capture how audiences experience LLM-influenced writing. By analyzing participant responses, researchers sought to understand whether reliance on these tools subtly alters the persuasive power, readability, and perceived authenticity of written communication, offering crucial insight into the evolving relationship between human authors and artificial intelligence in content creation.

Researchers investigated how large language models (LLMs) shape the perception of argumentative strength, aiming to understand their impact on persuasion and rhetorical effectiveness. The study focused on participant evaluations of text, specifically assessing how LLM assistance alters the perceived stance of an argument – whether it appears more assertive, neutral, or biased. By analyzing these perceptions, the research sheds light on the subtle ways LLMs can influence how audiences interpret information and form opinions. Findings indicate a tendency for LLM-generated or assisted text to be viewed as less opinionated and more objectively presented, potentially impacting the persuasive power of the writing and altering the dynamics of rhetorical communication.

User perception studies reveal a distinct shift in writing style correlated with increased reliance on large language models. Participants consistently rated texts heavily assisted by LLMs as exhibiting reduced creativity and a diminished sense of personal voice, a difference found to be statistically significant. This trend extends to perceptions of clarity, with LLM-reviewed papers receiving assessments that were 32% lower in terms of strong communication. However, the same papers demonstrated a substantial 136% increase in evaluations emphasizing reproducibility, as noted in ICLR peer reviews – suggesting a trade-off where statistical rigor and replicability are prioritized over stylistic nuance and individual expression. These findings indicate that while LLMs can bolster the scientific process through enhanced reproducibility, they simultaneously contribute to a homogenization of writing, potentially impacting the conveyance of complex ideas and authorial intent.

The study’s findings regarding LLMs’ consistent alteration of human writing echo a fundamental principle of logical rigor. As Bertrand Russell observed, “To be a great teacher, you need to be able to explain things simply.” LLMs, while capable of generating text, demonstrate a tendency towards stylistic and semantic drift, effectively obscuring the original intent – a corruption of clarity. This aligns with the research’s core idea that LLMs introduce unintended changes, potentially impacting the precision required in fields like scientific communication, where unambiguous expression is paramount. The pursuit of ‘fluency’ should not overshadow the necessity of provable correctness.

What’s Next?

The observed distortions introduced by Large Language Models are not merely stylistic quirks, but fundamental alterations to semantic content. The question is not whether these models can write, but whether they can preserve meaning across transformations. Current evaluation metrics, reliant on superficial similarity to source text, are demonstrably insufficient. A rigorous framework, grounded in information-theoretic invariants, is required to quantify the loss – or, more disturbingly, the creation – of information during AI-assisted writing. Asymptotically, any non-invertible transformation will introduce error; the challenge lies in characterizing the rate of this divergence.

Future work must address the issue of ‘semantic drift’ in large corpora edited by LLMs. The repeated application of these transformations, even with small individual distortions, could lead to a measurable shift in the collective understanding of concepts. This is particularly concerning for scientific and historical texts, where precise meaning is paramount. Establishing a ‘conservation of meaning’ principle – akin to the conservation of energy in physics – is not merely desirable, but essential.

The current trajectory suggests a future where authorship becomes a collaborative, yet inherently lossy, process. The study of this loss – its nature, its rate, and its ultimate consequences – represents a crucial, and largely unexplored, frontier in the mathematics of communication. The pursuit of ‘perfect’ AI writing is a distraction; the real task is to understand – and mitigate – its inherent imperfections.

Original article: https://arxiv.org/pdf/2603.18161.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-21 11:37