Author: Denis Avetisyan
New research reveals that human writing isn’t simply being replaced by artificial intelligence, but evolving alongside it in surprising ways.
A stylometric analysis demonstrates a ‘Dual-Track Evolution’ of thematic convergence and stylistic differentiation in human writing influenced by large language models.
The increasing prevalence of Large Language Models challenges conventional notions of authorship and creative originality. This study, ‘Writing in Symbiosis: Mapping Human Creative Agency in the AI Era’, investigates how human writers are adapting alongside increasingly capable AI, revealing a complex pattern of coevolution. Our analysis of a large-scale corpus demonstrates not a simple convergence of styles, but a ‘Dual-Track Evolution’ characterized by thematic alignment around AI alongside distinct stylistic differentiation, resulting in identifiable author archetypes. How will these emergent patterns of textual adaptation reshape our understanding of authorship, and what implications do they hold for detecting AI-generated content and preserving creative diversity?
The Shifting Sands of Expression
The rise of Large Language Models (LLMs) is enacting a demonstrable shift in the very fabric of written communication, moving beyond simple grammar checking to influence stylistic choices and overall textual complexity. These models, trained on massive datasets of text, are not merely tools for generating content, but active participants in shaping how individuals and organizations compose written materials. Evidence suggests a growing homogenization of style, as users increasingly rely on LLMs to refine or even author text, potentially leading to a decline in uniquely identifiable authorial voices. Furthermore, the accessibility of these models allows for rapid experimentation with different tones and structures, accelerating linguistic evolution at an unprecedented rate and prompting questions about authenticity and the future of human expression in a digitally mediated world.
Conventional stylometry, long relied upon to identify authorship and trace linguistic evolution, now struggles to keep pace with the accelerating changes driven by Large Language Models. These techniques, typically focused on frequency distributions of words, sentence length, and punctuation, are proving inadequate when applied to texts increasingly influenced-or even generated by-AI. The nuance of authorial voice, once discernible through subtle stylistic choices, is being obscured by the homogenizing effect of LLMs, which often prioritize grammatical correctness and clarity over individual expression. Moreover, the sheer speed at which these models are evolving means that static analytical benchmarks quickly become obsolete, unable to capture the dynamic shifts in complexity and phrasing that characterize contemporary writing. Consequently, researchers are actively developing new methodologies capable of detecting not just what is written, but also how it diverges from-or conforms to-patterns established by artificial intelligence.
The evolving relationship between human authors and increasingly sophisticated language models demands careful examination of adaptive strategies in writing. As AI-generated text becomes more prevalent, observing how individuals modify their stylistic choices – whether embracing new patterns suggested by these models or consciously diverging to maintain a distinct voice – provides critical insight into the future of authorship. This isn’t merely about tracking linguistic trends; it’s about understanding the subtle negotiations occurring as humans and machines co-create content, and how resistance or accommodation to AI influences the authenticity and perceived originality of written expression. Determining the extent to which authors actively shape their prose in response to algorithmic influence is therefore fundamental to gauging the long-term impact of AI on the very nature of human communication and creative output.
Quantifying the Echo: Perplexity-Gap Analysis
Perplexity-Gap Analysis is a quantitative method for assessing shifts in an author’s writing style over time. The technique establishes a baseline stylistic profile using texts authored before 2022. This profile is then compared to the author’s more recent writing, with divergence measured using perplexity scores generated by two large language models: GPT-2 Medium and Llama-3-8B-base. Lower perplexity indicates a greater similarity between the author’s current writing and the patterns learned by the LLMs, while higher perplexity suggests a greater stylistic distance. The resulting “perplexity gap” provides a numerical value representing the degree to which an author’s writing has shifted towards or away from LLM-generated text.
Stylometric profiling in this analysis utilizes quantifiable linguistic characteristics to establish a baseline for authorial style. Specifically, Mean Sentence Length is calculated as the average number of words per sentence, providing insight into syntactic complexity. Punctuation Density is determined by dividing the total number of punctuation marks by the total word count, indicating the author’s rhythmic and emphasis patterns. Finally, Passive Voice Ratio is computed by dividing the number of passive voice constructions by the total number of clauses, revealing tendencies toward agentless or indirect phrasing. These features, when combined, generate a multi-dimensional stylistic profile for each author and LLM model, enabling comparative analysis of textual characteristics.
Perplexity-Gap Analysis quantifies stylistic alignment between an author’s writing and large language model (LLM) outputs by comparing stylometric profiles. This is achieved through the calculation of a “gap” score based on the divergence of features – including Mean Sentence Length, Punctuation Density, and Passive Voice Ratio – between pre-2022 writing samples and current work, relative to established baselines from models like GPT-2 Medium and Llama-3-8B-base. A smaller gap indicates convergence with LLM stylistic patterns, while a larger gap suggests the author maintains a distinct, differentiated style. The resulting score provides a measurable metric for assessing the influence of LLMs on an author’s writing and identifying potential stylistic drift.
Decoding Authorial Responses: Resistance, Pragmatism, and Adoption
Application of HDBSCAN clustering to Stylistic Change Vectors – calculated from linguistic analysis of both arXiv preprints and the Discord Unveiled Dataset – revealed the presence of three distinct author archetypes. Stylistic Change Vectors quantify the shift in an author’s writing style before and after the widespread availability of large language models. The HDBSCAN algorithm was chosen for its ability to identify clusters of varying densities without requiring a pre-defined number of clusters. This methodology allowed for the empirical identification of groups exhibiting similar patterns of stylistic evolution, forming the basis for categorizing authors based on their response to LLM technology.
Analysis of author stylistic change, using HDBSCAN clustering, identified two distinct approaches to language use following the emergence of large language models. ‘Resistors’ are characterized by the maintenance of pre-LLM linguistic complexity, indicating a prioritization of uniquely human stylistic signatures. Conversely, ‘Adopters’ exhibit a high degree of perplexity gap – a measure of divergence from pre-LLM language patterns – aligning their writing with patterns increasingly common in LLM-generated text. This suggests an embrace of, or adaptation to, the stylistic characteristics of contemporary language models, as opposed to the preservation of prior linguistic habits.
The ‘Pragmatist’ author archetype, comprising 41% of the studied sample, is characterized by a moderate increase in perplexity gaps when compared to pre-LLM linguistic patterns. This suggests a measured adoption of language features associated with large language models, falling between the full adoption of ‘Adopters’ and the stylistic preservation of ‘Resistors’. Furthermore, Pragmatists demonstrate a statistically significant increase in their engagement with themes related to artificial intelligence, as evidenced by topic modeling of their written content. This combined linguistic and thematic profile indicates a strategic adaptation of language, potentially to enhance clarity or relevance in the context of evolving AI technologies and discourse.
Analysis of the author cohort revealed a distribution of archetypes where 18% of authors are classified as ‘Adopters’, 21% as ‘Resistors’, and the largest group, 41%, are ‘Pragmatists’. This proportional representation indicates that strategic adaptation – evidenced by the ‘Pragmatist’ archetype – is the most common response to the emergence of large language models within the studied author population. The relatively lower proportions of both full adoption and resistance suggest a nuanced linguistic landscape, where a majority of authors are actively modulating their style rather than fully embracing or rejecting LLM-influenced patterns.
The validity of the identified author archetypes was quantitatively assessed using two distinct clustering evaluation metrics. The silhouette score, calculated as 0.426 with a 95% confidence interval of 0.419-0.433, indicates moderate separation between clusters, suggesting reasonably well-defined groupings. Furthermore, the Adjusted Rand Index (ARI) of 0.891 (95% CI: 0.884-0.898) demonstrates a high degree of similarity between the cluster assignments generated by the HDBSCAN algorithm and a theoretical, ideal clustering, indicating strong agreement and reliability of the archetype identification.
The Dual-Track Evolution of Human Writing
Recent analyses of written communication reveal a compelling pattern – a ‘Dual-Track Evolution’ in human writing driven by the rise of artificial intelligence. This evolution isn’t simply about mirroring AI-generated text; instead, it’s characterized by a simultaneous narrowing of topical focus and a broadening of stylistic approaches. While authors increasingly address themes related to AI itself – indicating thematic convergence – their methods of expression are demonstrably diversifying. This divergence isn’t random; it suggests a deliberate negotiation of authorial voice, as writers subtly – and sometimes overtly – distinguish their work from the predictable patterns of large language models. The result is a complex interplay where shared subject matter coexists with increasingly unique linguistic fingerprints, hinting at a future where human and artificial writing occupy distinct, yet interconnected, communicative spaces.
Current analyses reveal a striking phenomenon: as discussions surrounding artificial intelligence permeate both scholarly articles and everyday online conversations, human writing isn’t simply becoming more uniform. Instead, a paradoxical trend emerges – thematic convergence is coupled with stylistic divergence. While the subjects of writing increasingly center on AI – its capabilities, implications, and societal impact – the ways in which people write are becoming more varied. This isn’t a homogenization of voice, but rather an expansion of authorial expression, suggesting individuals are actively shaping their linguistic identities in response to the rise of machine-generated text.
Analysis of recent writing samples reveals a significant alteration in linguistic patterns, evidenced by a measurable increase in the “Perplexity Gap” across both social and formal communication. This gap, a metric reflecting the difference between the predictability of language and its actual complexity, rose by 23% in social discourse and 15% in formal writing. These shifts suggest that authors are increasingly employing more varied and nuanced language structures, moving beyond the predictable patterns often favored by current large language models. The observed divergence isn’t simply about increased randomness; rather, it indicates a deliberate stylistic negotiation, as writers appear to be actively differentiating their work from AI-generated text through deliberate complexity and originality in expression.
Analysis of writing samples reveals that human authors aren’t passively adopting the stylistic traits of large language models, but are instead engaged in a complex process of self-differentiation. Researchers identified distinct archetypes – stylistic responses ranging from deliberate ornamentation to stark minimalism – demonstrating a conscious negotiation of authorial voice in relation to AI-generated text. This isn’t simple imitation; instead, authors appear to be actively defining their own linguistic identities by contrasting or complementing the patterns commonly associated with artificial intelligence. The observed patterns suggest a dynamic interplay, where human writers are leveraging stylistic choices to reaffirm their unique communicative strengths and signal their distinct presence in an increasingly AI-mediated world.
The trajectory of writing suggests a future not of replacement, but of coexistence, where human and artificial texts occupy distinct communicative niches. As artificial intelligence increasingly automates certain forms of writing – tasks prioritizing efficiency and clarity – human expression appears poised to emphasize qualities like nuanced perspective, emotional resonance, and stylistic innovation. This divergence isn’t simply about avoiding imitation; rather, it indicates an active recalibration of authorial voice, a deliberate assertion of uniquely human characteristics in response to the rise of machine-generated content. The anticipated result is a landscape populated by both pragmatic, AI-driven communication and richly textured, deliberately crafted human prose, each serving different, yet equally valuable, purposes in the evolving information ecosystem.
The study’s observation of ‘Dual-Track Evolution’-where thematic convergence occurs alongside stylistic divergence-feels less like a prediction and more like a post-mortem. It recalls Donald Knuth’s assertion that “Premature optimization is the root of all evil.” One might initially strive for seamless integration of human and AI writing, attempting to close the ‘Perplexity-Gap’ entirely. Yet, this research suggests that such homogenization isn’t the natural outcome; instead, a divergence arises, a kind of optimized differentiation. The architecture isn’t a diagram of perfect alignment, but a compromise that survived deployment-a system where thematic unity coexists with the persistent echoes of individual style. Everything optimized will, one day, be optimized back, and the study quietly documents that process unfolding.
The Road Ahead (And It’s Probably Paved With Corner Cases)
The notion of ‘archetypes’ emerging from human-AI coevolution is… predictable. Any taxonomy is a temporary respite from chaos. The real question isn’t whether these stylistic classifications will hold, but when the inevitable outliers will appear, forcing a re-evaluation. The paper correctly identifies a divergence beyond simple convergence, but assumes this differentiation will remain legible. It won’t. Production always finds a way to make elegant theory resemble a hastily patched system.
Perplexity-Gap analysis, while useful, is ultimately a measurement of current failure modes. It’s a snapshot, not a forecast. Future work must grapple with the fact that ‘textual adaptation’ isn’t linear. Humans will likely overcorrect, developing styles specifically designed to avoid AI mimicry – creating a constant arms race measured in fleeting stylistic trends. Any attempt to predict these shifts will, of course, be instantly outdated.
Ultimately, the value isn’t in mapping the current landscape, but in accepting its impermanence. Documentation of these ‘archetypes’ is a charming fiction; collective self-delusion masquerading as insight. If a bug is reproducible, it has a stable system. And if a stylistic trend is easily categorized, it’s already on its way out. The next step isn’t refinement, but embracing the inherent messiness of language.
Original article: https://arxiv.org/pdf/2512.13697.pdf
Contact the author: https://www.linkedin.com/in/avetisyan/
See also:
- Brawl Stars December 2025 Brawl Talk: Two New Brawlers, Buffie, Vault, New Skins, Game Modes, and more
- Clash Royale Best Boss Bandit Champion decks
- Best Hero Card Decks in Clash Royale
- Call of Duty Mobile: DMZ Recon Guide: Overview, How to Play, Progression, and more
- Clash Royale December 2025: Events, Challenges, Tournaments, and Rewards
- Best Arena 9 Decks in Clast Royale
- Clash Royale Best Arena 14 Decks
- Clash Royale Witch Evolution best decks guide
- Brawl Stars December 2025 Brawl Talk: Two New Brawlers, Buffie, Vault, New Skins, Game Modes, and more
- Decoding Judicial Reasoning: A New Dataset for Studying Legal Formalism
2025-12-17 13:36