Algorogenic Thumbprints in Literary Translation: An Epistemic Network Analysis of LLM Translations of Classical Chinese Fantasy

Algorogenic Thumbprints in Literary Translation: An Epistemic Network Analysis of LLM Translations of Classical Chinese Fantasy | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Algorogenic Thumbprints in Literary Translation: An Epistemic Network Analysis of LLM Translations of Classical Chinese Fantasy Kan Wu, Siqi Jiang, Defeng Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9323835/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 13 You are reading this latest preprint version Abstract This study investigates algorogenic thumbprints, defined as consistent and reproducible stylistic patterns generated by large language models (LLMs) in literary translation. Using epistemic network analysis (ENA) across seven decision-making domains, lexical choice, syntactic structure, cultural adaptation, stylistic register, semantic fidelity, narrative flow, and intertextual reference, we examine translations of two classical Chinese fantasy novels produced by Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro, alongside published human translations and a Western fantasy reference corpus. The findings show that each LLM exhibits a stable, model-specific epistemic architecture oriented toward narrative coherence, whereas human translators structure decisions more strongly around semantic fidelity. This architectural divergence persists across texts and models. The study therefore suggests that Baker’s framework of translator thumbprints can be meaningfully extended to algorithmic translation and demonstrates the usefulness of ENA for modeling the relational structures underlying translation decision-making. Humanities/Cultural and media studies Social science/Cultural and media studies Humanities/Language and linguistics Social science/Language and linguistics Humanities/Literature algorogenic thumbprints epistemic network analysis large language models translator style classical Chinese fantasy Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction When classical Chinese fantasy novels traverse linguistic boundaries in the digital age, they raise not merely translation challenges but deeper questions about how artificial intelligence organize linguistic and stylistic choices when rendering culturally embedded narratives. Journey to the West and Creation of the Gods , canonical works of the Chinese Gods and Demons genre, have long fascinated Western readers through their mythological systems, supernatural battles, and philosophical depth (Hsia, 2016 ). As large language models (LLMs) increasingly take on literary translation tasks, these texts become productive testing grounds for a focused inquiry: how do LLMs make decisions in translation when rendering culturally embedded narratives, and do these decision-making patterns differ meaningfully from those of human translators? This question gains urgency as LLMs transform translation workflows. Models such as Claude, ChatGPT, and Gemini now produce translations at unprecedented speed and scale, prompting both enthusiasm about democratized access to world literature and concern about the erosion of translator craft (Kenny, 2022 ). Yet, amid ongoing debates about quality and authenticity, the epistemic architecture underlying AI translation decisions remains underexplored, particularly whether different LLMs develop distinctive signatures in coordinating competing demands such as lexical precision, cultural adaptation, and stylistic register. Translation quality assessment has traditionally focused on error analysis and fluency metrics (Läubli et al., 2018 ), while computational approaches tend to emphasize neural network architectures rather than decision-making patterns (Elmoazen et al., 2022 ). Recent computational stylistics has begun addressing this gap through stylometric studies of LLM translations (Yao et al., 2025 ; Ping and Wang, 2024 ) and AI creative writing (O’Sullivan, 2025 ; Elkins, 2024 ; Mikros, 2025 ), suggesting that LLMs may develop identifiable stylistic signatures. Baker’s ( 2000 ) influential framework of translator thumbprints, which posits that translators leave varying styles through systematic linguistic choices, has inspired substantial research on human translators but remains largely unapplied to AI systems. Process-oriented methods such as keystroke logging and eye-tracking (Jakobsen and Jensen, 2008 ) are not readily transferable to LLM cognition. What is needed, then, is a method capable of modeling not isolated translation choices, but the relational structures through which meaning is constructed. Epistemic network analysis (ENA) offers one such approach. Originally developed to model complex thinking in collaborative learning environments, ENA quantifies and visualizes co-occurrence patterns among coded epistemic activities as weighted network graphs (Shaffer, 2018 ). Unlike traditional content analysis, ENA captures how elements function relationally within a system, making it particularly suited to translation, where decisions are inherently interdependent (Csanadi et al., 2018 ). While ENA has been applied productively in studies of professional practice (Ruis et al., 2018 ) and educational discourse (Siebert-Evenstone et al., 2017 ), it remains largely underexplored in translation studies. This study applies ENA to compare three LLMs, Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro, translating Creation of the Gods and Journey to the West against published human translations and a Western heroic fantasy reference corpus. Systematic coding across seven decision-making domains enables construction of epistemic networks revealing how each translator coordinates translational choices. Specifically, three research questions guide the analysis: How do the epistemic networks of the three LLMs, Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro, differ in their translations of classical Chinese fantasy? How do the epistemic networks of LLM translations differ from those of human translations of classical Chinese fantasy? How do the epistemic networks of LLM translations differ from the network structure of the Western fantasy reference corpus? 2. Related Theories 2.1 Baker’s Thumbprint Framework and Beyond Baker ( 2000 : 245) conceptualizes translator style as a thumbprint , that is, a translator’s characteristic use of language and individual profile of linguistic habits in comparison with other translators. This target-oriented perspective treats stylistic idiosyncrasies as operating independently of any particular source text (Saldanha, 2011 ), in contrast to source-oriented approaches advocated by Malmkjær ( 2003 ) and Boase-Beier ( 2006 ), which view translator style as responsive to source text features. These divergent perspectives have shaped distinct methodological trajectories in translator style research, yet both presuppose human consciousness, with its biographical accumulation, socio-cultural embedding, and creative intentionality, as the generative force behind stylistic choices. This presupposition is one this study interrogates rather than inherits uncritically. The framework’s human-centric foundations become particularly complicated when extended to LLMs. Even within human translation, distinguishing “deliberate” from “unconscious” stylistic choices remains methodologically elusive (Saldanha, 2011 ; Winters, 2004 ). This difficulty is compounded further when the translator is algorithmic, as LLM patterns emerge not from accumulated biographical experience, but from statistical regularities encoded across training data (Ping and Wang, 2024 ). Yet, the consistency and cross-textual stability of such patterns, as suggested by recent computational stylistics research (Yao et al., 2025 ; Mikros, 2025 ), suggests that the absence of human consciousness may not preclude identifiable stylistic signatures. This study therefore extends Baker’s (ibid.) framework to examine whether LLMs develop consistent stylistic signatures and, if so, whether the mechanisms generating such patterns are structurally comparable to those underlying human translation choices or constitute a categorically distinct phenomenon. Traditional translator style studies employ two primary comparative modes: the “one-to-many mode,” examining multiple translations of a single source text, and the “many-to-many mode,” comparing translations across varied texts (Wu and Li, 2024 , p. 273). Winters ( 2004 ) exemplifies the one-to-many approach through his analysis of loanwords and modal particles in translations of The Beautiful and Damned , while Mastropierro ( 2018 ) identifies distinctive stylistic markers across Italian translations of Lovecraft . Baker ( 2000 ) herself employs the many-to-many mode, tracing patterns in type/token ratio and reporting verbs; Olohan ( 2003 ) and Saldanha ( 2011 ) similarly examine contraction preferences and italicization practices. Source-oriented researchers such as Bosseaux ( 2004 ; 2007 ) and Huang ( 2015 ) focus instead on how translators reproduce narrative voice through deixis, modality, transitivity, and free indirect discourse. The existing literature thus leaves a clear gap: although Baker’s framework and its later extensions have been widely applied to human translators, their relevance to algorithmic translation remains largely untested. It is still unclear whether LLMs exhibit consistent stylistic signatures comparable to human translator thumbprints or a categorically different kind of patterning. To address this issue, the present study distinguishes between anthropogenic thumbprints and algorogenic thumbprints . Anthropogenic thumbprints arise from human translators’ agency, experience, and socio-cultural embeddedness, while algorogenic thumbprints from statistical training, probabilistic prediction, and optimization in large language models. This heuristic distinction is proposed to evaluate whether Baker’s framework can be meaningfully extended to algorithmic translation. With this distinction in place, the next section turns to the role of LLMs in literary translation and their current translational capabilities. 2.2 Large Language Models as Literary Translators Machine translation has been substantially transformed by the emergence of AI-powered large language models (Han and Lu, 2025 ). ChatGPT exemplifies this shift. Though not specifically designed for translation, its contextual understanding and coherent text generation capabilities have shown competitive performance relative to conventional MT systems in some translation settings, positioning such models as potential translators’ prostheses (Lee, 2024 ). These developments raise questions about whether AI systems can function as literary translators, a role traditionally associated with human expertise in what Landers ( 2001 : 7) termed “the most demanding type of translation.” Research on LLMs as literary translators has broadly developed along two trajectories: evaluative studies of translation quality and descriptive studies of translational behaviour. Evaluative research has moved beyond narrow measures of accuracy toward multidimensional frameworks that incorporate stylistic considerations, including document-level performance, literary fluency, coherence, and sensitivity to tone, rhythm, and culturally specific expression (Karpinska and Iyyer, 2023 ; Gao et al., 2024 ; Zhang et al., 2025 ). At the same time, these studies suggest that although LLMs have made clear gains in literary translation, human translators continue to retain an advantage in overall quality, particularly when stylistic and terminological factors are taken into account (Zhang et al., 2025 ). While this quality-oriented line of research provides important evidence of LLMs’ literary translation capabilities, its normative orientation leaves a more basic question insufficiently addressed: whether LLMs exhibit distinctive translational patterns comparable to Baker’s ( 2000 ) translator “thumbprints.” Descriptive research offers a complementary perspective by examining recurrent linguistic features in LLM-generated translations without reducing analysis to predefined standards of adequacy. Emerging work (Lee, 2024 ; Li, 2024 ; Yao et al., 2025 ; Mikros, 2025 ) in this area points to the possibility that LLMs can display consistent tendencies in the management of cohesion, coherence, source-text restructuring, and stylistic patterning, raising the broader question of whether such regularities may be understood as a form of translational voice or style. Nonetheless, descriptive research on LLMs as literary translators remains limited, particularly with regard to consistency across texts, genres, and models. While Lee ( 2024 ) offers qualitative insight into the linguistic behaviour of a single model, the field still lacks systematic empirical investigation into whether different LLMs develop model-specific styles in translation and how such styles compare with those of human translators. This gap is especially significant because Baker’s ( 2000 ) framework is grounded in assumptions of human consciousness, agency, and socio-cultural positioning, assumptions that cannot simply be transferred to LLMs. If stylistic analysis in human translation often relies, at least implicitly, on the idea of an intentional and socially situated subject, then the study of LLM translation requires a different analytical basis: one that does not depend on access to inner consciousness or inferred motivation, but instead focuses on observable patterns in how translational choices are coordinated. For this reason, what is needed is not only the identification of surface-level stylistic features, but also a method for modelling the relational structures through which such features co-occur across translations. Epistemic network analysis offers precisely this kind of relational account, and the following section considers its relevance for translation research. 2.3 Epistemic Networks in Translator Style Analysis Traditional approaches to translator style analysis, including foundational work by Baker ( 2000 ) and subsequent studies (Li, 2016 ; Wu and Li, 2024 ), have relied primarily on frequency-based methods to identify stylistic patterns. While such methods have proven productive, they face notable limitations. Conventional coding-and-counting strategies tend to treat translational features in isolation, ignoring temporality in data and the broader patterns connecting translational activities (Csanadi et al., 2018 ). Counting how often a translator employs explicitation, for instance, or how frequently they adopt a formal register, reveals little about whether these choices are epistemically linked in the translator’s decision-making. What such approaches largely cannot capture is the relational structure of translation, not merely which strategies occur, but how they co-occur and interconnect. Epistemic Network Analysis (ENA) was developed to address precisely this kind of structural question by modelling patterns of association among elements in coded data (Shaffer et al., 2009 ; Shaffer, 2018 ). Originally designed to examine epistemic networks, the patterns of association among knowledge, skills, values, and habits of mind that characterize complex thinking, ENA quantifies, visualizes, and enables statistical comparison of such structures. A key assumption of ENA aligns well with conceptualizing translation as an epistemic process: that the structure of connections among elements may be more analytically meaningful than the mere presence or absence of those elements in isolation (Cacioppo and Cacioppo, 2012 ). Applied to translator style, ENA thus enables researchers to model not simply which translational strategies appear, but how these strategies are relationally organized across a translation. Methodologically, ENA identifies co-occurrences of coded elements within defined data segments, called stanzas , and represents them as weighted connections in network models (Shaffer et al., 2016 ). Each translator’s network thereby captures both the frequency and strength of associations among different translational choices. Dimensional reduction and optimization techniques then position network nodes such that the centroid of each network corresponds to its location in a projected space, facilitating both visual and statistical comparison across translators (Shaffer et al., 2016 ). This co-registration ensures that differences in network visualizations correspond directly to quantifiable structural differences, lending interpretive clarity to statistical findings. Thus, ENA provides several methodological advantages for translator style analysis. It captures the relational nature of translation decision-making by modelling how choices in one domain, for example lexical selection, connect systematically with choices in another, including syntactic restructuring. Network visualizations also provide more interpretable representations of complex associative patterns than conventional statistical outputs (Csanadi et al., 2018 ). Furthermore, unlike sequential analysis methods that typically require large datasets and can produce difficult-to-interpret transition probabilities, ENA is well-suited to relatively small, fixed sets of elements with complex patterns of association (Csanadi et al., 2018 ). Individual translators’ networks can each be represented as a point in projected space, where proximity indicates similarity in translational patterning, and standard statistical tests can then determine whether different translator groups exhibit meaningfully different epistemic architectures. These properties make ENA particularly well-suited to investigating LLMs as literary translators. If LLMs do develop distinctive “algorogenic thumbprints,” such patterns would likely be expressed not in isolated feature frequencies alone, but in systematic structures of association among translational choices. ENA offers a principled means of operationalizing Baker’s ( 2000 ) insight that translator thumbprints emerge from habitual patterns of choice-making, while extending this framework to encompass AI algorithmic processing alongside human consciousness. 3. Methodology To examine whether LLMs develop consistent stylistic signatures comparable to, or categorically different from, human thumbprints, the study performs epistemic network analysis (ENA) across three comparative dimensions: variations and similarities among LLM models, between LLMs and human translators, and between LLM translations and non-translated original English texts. The following details the research design and procedures. 3.1 Data and Design The corpus design (Fig. 1 ) includes both translated and non-translated fantasy literature, creating a comparative framework for the analysis of epistemic structures in literary translation. The translated component features English translations of two Chinese classical novels (see Table 1 ): 封神演義 ( Creation of the Gods , CG) and 西遊記 ( Journey to the West , JTW), representing the Gods and Demons subgenre of Chinese fantasy. These translations include human translations (Gu Zhizhong’s 1992 CG and W.J.F. Jenner’s 1993 JTW) and LLM translations of these works from three models: Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro. The selected chapters from both Chinese novels represent diverse thematic and stylistic elements. For CG: Chap. 1 introduces mythological themes; Chap. 12 features Nezha’s birth with Taoist terminology; Chap. 21 focuses on political intrigue; Chap. 70 presents battles with religious elements; Chap. 79 combines supernatural and military aspects; Chap. 100 concludes with ceremonial language. For JTW: Chap. 1 establishes the stone monkey’s origin; Chap. 7 details Sun Wukong’s celestial confrontation; Chap. 15 introduces pilgrimage’s spiritual dimensions; Chap. 27 demonstrates humour and action; Chap. 47 presents Buddhist concepts; Chap. 68 showcases allegorical supernatural encounters. Table 1 Translations by human and LLMs and their token size Translation Translator Token size Creation of the Gods (CG) Chap. 1, 12, 21, 70, 79, 100 Human by Gu Zhizhong 21,420 LLM by ChatGPT 5.2 Pro 17,648 LLM by Claude Opus 4.5 19,021 LLM by Gemini 3 Pro 15,487 Journey to the West (JTW) Chapter 1, 7, 15, 27, 47, 68 Human by W.J.F. Jenner 27,552 LLM by ChatGPT 5.2 Pro 25,704 LLM by Claude Opus 4.5 28,211 LLM by Gemini 3 Pro 18,266 Table 2 Details of the reference corpus Western Heroic Fantasies Year Author Token Size The Lord of the Rings 1–2 1954–1955 J. R. R. Tolkien 188,361 A Song of Ice and Fire 1 1996 G. R. R. Martin 293,856 Conan the Barbarian 1954 R. E. Howard 104,972 Wheel of Time 1 1990 R. Jordan 319,124 The Chronicles of Amber 1–4 1970–1976 R. Zelazny 239,940 The Chronicles of Narnia 1–5 1950–1954 C. S. Lewis 233,369 Total Size 1,379,622 The non-translated component includes six influential works of Western heroic fantasy (Table 2 ), totalling approximately 1.38 million words. This reference corpus is used as a broad stylistic comparator for fantasy writing in English. Because it consists of original English texts rather than translations, and because it is much larger than the translated dataset, it is not treated as a directly equivalent baseline for assessing translation quality, translator expertise, or translationese. The function of the reference corpus is limited to providing a contextual point of comparison for interpreting whether certain network tendencies in the translated corpus appear more translation-specific or more broadly genre-compatible. 3.2 Framework and Procedures The analytical framework defines seven decision-making domains for describing salient translational features in the English outputs: lexical choice (word selection and terminology decisions), syntactic structure (sentence construction and clause arrangement), cultural adaptation (handling culture-specific items), stylistic register (tone, formality, and genre conventions), semantic fidelity (preserving source text meaning), narrative flow (coherence and readability), and intertextual reference (allusions and literary echoes). Informed by House’s ( 2015 ) model of translation quality assessment and Vandepitte and Hartsuiker’s ( 2011 ) categorization of translation problems, these categories describe observable translation solutions to multidimensional demands (Pym, 2015 ) and broadly align with Chesterman’s ( 2016 ) account of translation universals. They are used as descriptive analytic constructs rather than direct indicators of internal decision-making, enabling recurrent co-occurrence patterns to be modeled through epistemic network analysis. The research unfolds in four systematic steps. Step 1: Translation Generation and Segmentation. All AI translations were generated via each model’s official API rather than web-based interfaces, ensuring controlled parameter settings and minimizing platform-side variability. A standardized, stylistically neutral prompt was employed across all three models: “ You are a translation assistant. Translate the following Chinese text into English. Output only the translated text, preserving the original paragraph structure and line breaks exactly as they appear .” The prompt imposed no stylistic preferences, register directives, or target-audience specifications, ensuring that observed differences reflect each model’s default translational tendencies rather than prompt-induced behavior. Temperature was set to 0 across all API calls, minimizing stochastic sampling variation and producing deterministic, reproducible outputs, thereby ensuring stable and consistent translation versions across the corpus. It should be noted, however, that zero-temperature sampling may not fully represent each model’s characteristic stylistic range; the thumbprints identified here thus reflect deterministic rather than fully naturalistic translational behaviour. All translations were subsequently segmented into 200-word passages, balancing contextual sufficiency with analytical granularity. Step 2: Manual Coding and Reliability Assessment. All segments were independently coded by four PhD-trained coders in translation studies using the predefined codebook (see Appendix A). Coding was conducted at the segment level, and multiple codes could be assigned when more than one domain was salient. Following iterative calibration, the coders re-annotated a shared subset comprising 20% of all segments, sampled proportionally across texts and translator/model outputs. Because four coders were involved, inter-rater reliability for each code was calculated as the mean of the six pairwise Cohen’s kappa values. Acceptable agreement was defined a priori as κ ≥ 0.80. Final kappa values are reported in Appendix C. Remaining disagreements were then resolved through discussion to produce the final coding dataset. Step 3: Data Preparation and Stanza Configuration. Following ENA methodology, coded segments were organized into stanza-based interaction data (Shaffer, 2018 ). Stanzas, defined as moving windows of 5 consecutive lines, establish which translation decisions co-occur within meaningful units of context (Siebert-Evenstone et al., 2017 ). This window size was selected through pilot testing as a balance between local context and analytical tractability, producing stable and interpretable networks for the present dataset. For each stanza, ENA creates an adjacency matrix representing co-occurrences among the seven domains, then accumulates these matrices into a cumulative adjacency matrix for each translator or model, quantifying the total connections made across all stanzas. Python scripts (version 3.14) formatted coded transcripts for ENA input. Step 4: Network Generation and Visualization. Python scripts utilizing the pyENA library quantified code co-occurrences and accumulated connection strengths across stanzas. ENA employs singular value decomposition (SVD) for dimensional reduction while optimizing node positions such that each network’s centroid corresponds to its location in the reduced space, enabling meaningful visualization interpretation (Shaffer et al., 2016 ). The analysis generated individual and mean network visualizations for comparing network density, domain centrality, and connection weights. Goodness-of-fit statistics (Pearson and Spearman correlations > 0.90) indicated acceptable model fit for the ENA projections used in this study. 4. Results 4.1 Epistemic Networks between Different LLMs The epistemic network analysis reveals that while all three models, Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3.0 Pro, demonstrate competence in handling classical Chinese texts, they exhibit markedly different epistemic strategies in processing and rendering these works into English, with distinct patterns emerging across the two literary corpora. For Creation of the Gods (CG), the three models display different network architectures visible in the left panels of Figs. 2 – 4 . Claude Opus 4.5 (Fig. 2 , left) generated 96 coded units with 15 active connections, achieving a mean connection strength of 0.1796. This relatively high connection density suggests Opus activates multiple translation codes simultaneously when processing CG passages, creating a richly interconnected epistemic landscape. The model’s strongest connections appear between Narrative Flow and Lexical Choice (weight: 0.2847), followed by Narrative Flow and Intertextual Reference (weight: 0.2753). Notably, Opus displays prominent connections between Cultural Adaptation and both Lexical Choice (weight: 0.2478) and Intertextual Reference (weight: 0.2478), suggesting this model more actively considers target culture accessibility when processing culturally-embedded passages compared to its counterparts. ChatGPT 5.2 Pro (Fig. 3 , left) analysed 64 units across 21 connections with a mean strength of 0.1452, displaying the most distributed network pattern among the three models for CG. Its strongest link connects Narrative Flow and Lexical Choice (weight: 0.3331), though the relatively lower mean connection strength compared to Opus suggests a more exploratory strategy that considers multiple translation dimensions with less concentrated intensity on specific code combinations. This distributed architecture indicates ChatGPT engages with diverse translation challenges simultaneously rather than prioritizing hierarchical coordination patterns. Gemini 3.0 Pro (Fig. 4 , left) generated 53 coded units across 10 connections, creating the most sparse but focused network structure among the three models. Gemini’s strongest connection links Narrative Flow and Lexical Choice (weight: 0.3966), with Intertextual Reference receiving secondary attention (weight: 0.3589). This pattern reveals Gemini’s tendency to anchor translation decisions in narrative continuity, using this dimension as an organizational principle for coordinating other translation codes. The model’s relatively compact network with fewer but stronger connections implies a more selective, focused translation approach that concentrates epistemic effort on specific translation challenges rather than distributing attention broadly. The Journey to the West (JTW) networks, shown in the right panels of Figs. 2 – 4 , reveal both consistency and adaptation in the models’ translation strategies when confronting a different classical text. Claude Opus 4.5 (Fig. 2 , right) demonstrates the most dramatic scaling behaviour, expanding from 96 units in CG to 167 units, a 74% increase that makes it the most prolific model in the JTW corpus. Despite this expansion, the model maintains 21 connections with a mean strength of 0.1451, slightly lower than its CG mean. The network reveals strong coordination between Narrative Flow and Intertextual Reference (weight: 0.2391) and between Lexical Choice and Intertextual Reference (weight: 0.2133), demonstrating consistent attention to cultural context. Cultural Adaptation assumes even greater prominence in JTW (weight: 0.2922) compared to CG, possibly reflecting the text’s more fantastical elements requiring enhanced domestication strategies. This adaptive scaling suggests Opus shows greater variation in network size and connection distribution across the two texts, engaging more extensively with longer, more complex narratives while maintaining reasonable coordination intensity. ChatGPT 5.2 Pro (Fig. 3 , right) expands to 129 units with 21 connections but shows a decreased mean strength of 0.1196, indicating the model processes substantially more material with less intensive code coordination per unit. Significantly, ChatGPT’s strongest connection in JTW shifts to Cultural Adaptation-Lexical Choice (weight: 0.2803), diverging from the Narrative Flow emphasis observed in CG. This shift suggests adaptive capacity based on text type, demonstrating responsiveness to different generic characteristics between the historical-mythological CG and the picaresque adventure narrative of JTW. However, the lower connection weights in JTW may indicate challenges in maintaining coherent strategy application across the episodic structure characteristic of this narrative form. Gemini 3.0 Pro (Fig. 4 , right) demonstrates remarkable consistency with 47 coded units forming 14 connections at a mean strength of 0.1596, maintaining its relatively sparse but focused network pattern observed in CG. The model’s primary connection strength lies between Narrative Flow and Lexical Choice (weight: 0.4144), with Intertextual Reference maintaining equal weight (0.4144) and secondary emphasis on Syntactic Structure and Lexical Choice (weight: 0.2464). This striking structural stability across both texts, 53 units and 10 connections in CG versus 47 units and 14 connections in JTW, suggests Gemini exhibits a relatively stable and narrowly distributed network structure across both texts. The comparative analysis across both corpora reveals that despite these differences, all three models share the centrality of Narrative Flow connections, indicating that maintaining story coherence remains paramount regardless of the AI system or source text. However, their divergent approaches, Gemini’s systematic consistency, ChatGPT’s exploratory breadth, and Opus’s adaptive scaling, suggest different operational philosophies in navigating the multidimensional challenges of classical Chinese literary translation. 4.2 Epistemic Networks between AI and Human Translations The epistemic network analysis reveals differences in how human translators and LLMs construct meaning when rendering classical Chinese texts into English. The comparison between Gu Zhizhong’s translation of CG and W.J.F. Jenner’s translation of JTW with the three LLMs demonstrates qualitatively distinct epistemic architectures that persist across both literary corpora. For Creation of the Gods , the human translator Gu Zhizhong (Fig. 5 , left) demonstrates a distinctive epistemic architecture centred on Semantic Fidelity. Strong connections radiate from Semantic Fidelity to multiple translation codes, particularly Lexical Choice, Syntactic Structure, and Intertextual Reference. This pattern indicates that Gu consistently anchors decision-making in preserving the semantic integrity of the source text while simultaneously coordinating multiple other translation dimensions, creating what might be termed a “semantic-first” translation architecture. Claude Opus 4.5 (Fig. 2 , left) positions Narrative Flow as the primary hub from which other connections emanate, diverging from Gu’s semantic-first approach. With 96 coded units across 15 connections and a mean strength of 0.1796, Opus’s strongest connection links Narrative Flow and Lexical Choice (weight: 0.2847), with Narrative Flow and Intertextual Reference receiving equal attention (weight: 0.2753). While Opus displays engagement with Cultural Adaptation through connections to both Lexical Choice (weight: 0.2478) and Intertextual Reference (weight: 0.2478), it lacks the semantic anchoring that characterizes Gu’s network, prioritizing story coherence over precise meaning preservation. ChatGPT 5.2 Pro (Fig. 3 , left) exhibits a more distributed network with 64 units across 21 connections and a mean strength of 0.1452, contrasting sharply with Gu’s concentrated semantic-centred structure. ChatGPT’s strongest connection links Narrative Flow and Lexical Choice (weight: 0.3331), following a hierarchical pattern similar to Opus but with broader, less intense code co-activations throughout. This distributed architecture reveals an exploratory strategy that considers multiple translation dimensions without Gu’s clear hierarchical prioritization around meaning preservation, suggesting a different organizational principle in translation. Gemini 3.0 Pro (Fig. 4 , left) generated 53 coded units across 10 connections, creating the sparsest yet focused network. Gemini’s strongest connection links Narrative Flow and Lexical Choice (weight: 0.3966), with Intertextual Reference receiving secondary attention (weight: 0.3589). This configuration reveals systematic prioritization of narrative continuity, diverging from Gu’s semantic-anchored network structure. Where Gu integrates multiple codes through semantic fidelity as the organizing hub, Gemini employs a selective approach that concentrates on narrative coherence and literary allusion management, treating semantic correspondence as secondary. For Journey to the West , W.J.F. Jenner’s translation (Fig. 5 , right) maintains the characteristic Semantic Fidelity centrality observed in Gu’s CG translation. Robust connections from Semantic Fidelity to Lexical Choice, Syntactic Structure, and Intertextual Reference form a coherent cluster, demonstrating that human translators coordinate these codes simultaneously while anchoring decisions in meaning preservation regardless of source text characteristics. This consistency suggests a stable human epistemic approach to classical Chinese translation. Claude Opus 4.5 (Fig. 2 , right) demonstrates dramatic scaling in JTW, expanding to 167 units with 21 connections while maintaining its narrative-centred orientation distinct from Jenner’s approach. The network reveals strong coordination between Narrative Flow and Intertextual Reference (weight: 0.2391) and between Lexical Choice and Intertextual Reference (weight: 0.2133). Cultural Adaptation assumes greater prominence (weight: 0.2922) compared to CG, possibly reflecting JTW’s fantastical elements. Despite this adaptive scaling, Opus continues organizing translation patterns around narrative coherence rather than adopting Jenner’s semantic-first architecture, revealing different priorities between human and artificial intelligence in managing literary translation. ChatGPT 5.2 Pro (Fig. 3 , right) expands to 129 units with 21 connections but shows decreased mean strength of 0.1196, contrasting with Jenner’s more concentrated code coordination. ChatGPT’s strongest connection shifts to Cultural Adaptation-Lexical Choice (weight: 0.2803), diverging from both its own CG pattern and Jenner’s semantic emphasis. This shift suggests some adaptive capacity based on text type, yet the model’s prioritization of cultural accessibility over meaning preservation represents a qualitative departure from Jenner’s translation philosophy, where semantic fidelity remains the gravitational center regardless of the text’s fantastical or cultural content. Gemini 3.0 Pro (Fig. 4 , right) maintains remarkable consistency with 47 units and 14 connections, with strongest connections continuing to link Narrative Flow with Lexical Choice (weight: 0.4144) and Intertextual Reference (weight: 0.4144). This structural stability across both texts contrasts sharply with Jenner’s flexible yet consistently semantic-centred approach. Where Jenner adapts specific translation choices while maintaining semantic fidelity as the organizing principle, Gemini applies a systematic algorithm that prioritizes narrative coherence regardless of source material characteristics, suggesting inflexibility in its translation philosophy. The comparative analysis across both corpora illuminates an architectural divergence: human translators construct epistemic networks where semantic preservation serves as the gravitational centre coordinating all other translation decisions, while LLMs organize their translation around narrative coherence, treating semantic fidelity as peripheral. This pattern persists across both classical texts and all three LLMs, suggesting these network differences reflect distinctions in human versus artificial intelligence approaches to cross-linguistic meaning-making. 4.3 Epistemic Networks between AI and Reference Corpus The epistemic network analysis reveals structural differences between human and AI translation processes across both CG and JTW. Understanding these differences requires first examining the reference corpus network architecture, which provides a structural point of reference for contextualizing the AI network architectures examined below, not as a directly equivalent baseline, but as an illustration of the kind of integrated epistemic organization characteristic of proficient original writing in the target language. The reference corpus demonstrates a densely interconnected epistemic network (see Fig. 6 ) with mean connection strength of 0.1147 across 11,955 coded segments. Authors of these original writings establish robust co-occurrences among all code pairs, with the five strongest connections being Syntactic Structure with Cultural Adaptation (0.2751), Syntactic Structure with Intertextual Reference (0.2677), Lexical Choice with Cultural Adaptation (0.2436), Cultural Adaptation with Intertextual Reference (0.2068), and Syntactic Structure with Semantic Fidelity (0.1989). This connectivity pattern indicates that human writers engage in integrated decision-making where choices in one dimension systematically influence and coordinate with others, creating a holistic writing process. The network positioning shows balanced distribution across both structural-cultural dimensions (SVD1) and narrative-semantic dimensions (SVD2), enabling flexible responses to varied textual demands. Critically, the network reveals no isolated nodes or peripheral codes; instead, all writing considerations maintain substantial mutual engagement, suggesting that expert writers simultaneously manage multiple constraints rather than processing them sequentially or independently. When examined alongside the reference corpus, Opus 4.5 exhibits markedly different network characteristics for CG translation. With mean connection strength of 0.1451 and maximum strength of 0.3474, the model shows higher mean density than some AI comparators in this study. Its strongest connections centre on Cultural Adaptation, linking it robustly to Lexical Choice (0.2922) and Intertextual Reference (0.2922). However, structural gaps emerge: Syntactic Structure connections remain weaker, with the Syntactic Structure to Intertextual Reference link reaching 0.2079, compared with 0.2677 in the reference corpus. This indicates reduced capacity for coordinating structural manipulation with allusion handling, essential for rendering the mythologically-embedded and allusion-dense passages characteristic of Creation of the Gods . For JTW translation, Opus 4.5 shifts emphasis toward Narrative Flow connections, achieving mean strength of 0.1811 with maximum of 0.3880. The strongest connection becomes Narrative Flow to Intertextual Reference (0.3272), suggesting recognition of JTW’s episodic-allusive structure. Yet the Narrative Flow to Cultural Adaptation link appears weaker than in the reference corpus, potentially limiting coordinated cultural transfer within narrative progression. Gemini 3 Pro exhibits a more fragmented network architecture across both texts. For CG, the model achieves mean connection strength of 0.1596, but maximum strength of 0.4743 concentrated narrowly in Narrative Flow to Intertextual Reference. While establishing moderately strong links between Narrative Flow and Lexical Choice (0.4144), the model fails to connect Cultural Adaptation with Syntactic Structure in meaningful ways. This fragmentation creates more isolated processing modules rather than the integrated architecture visible in the reference corpus, with the network centroid positioned peripherally in epistemic space. For JTW, Gemini improves substantially with mean strength of 0.1765 and maximum of 0.4431, establishing strong connections from Narrative Flow to Intertextual Reference (0.3589) and moderate links to Cultural Adaptation (0.1474). However, structural weaknesses persist: Syntactic Structure remains poorly integrated, and high connection standard deviation (0.1600) indicates inconsistent strategy application across different passages, contrasting with the more evenly distributed connectivity of the reference corpus. ChatGPT 5.2 Pro occupies an intermediate structural position between Opus and Gemini. For CG translation, the model achieves mean strength of 0.1196 with maximum of 0.3785, prioritizing Cultural Adaptation connections at 0.2803 while maintaining moderate links to Intertextual Reference (0.2252) and Lexical Choice (0.2100). Unlike the balanced topology found in the reference corpus, ChatGPT 5.2 Pro exhibits Cultural Adaptation as an over-weighted hub while peripheralizing structural considerations, explaining its tendency toward surface-level cultural substitution without corresponding structural adaptation. The Syntactic Structure to Intertextual Reference connection falls significantly below reference corpus figures, indicating fragmented rather than integrated processing. For JTW, ChatGPT 5.2 Pro shows better balance than on CG through Narrative Flow emphasis (0.3331) and dual Cultural Adaptation connections to both Intertextual Reference and Lexical Choice (0.2100 each), with mean strength of 0.1452 and maximum of 0.3578. Nevertheless, connections remain structurally divergent from reference corpus figures by 15–40% across critical code pairs. Across all comparisons, a consistent structural pattern emerges: LLMs demonstrate lower overall network density, unbalanced topologies with over-reliance on specific code pairs, and peripheral rather than central positioning in epistemic space. Where the reference corpus shows mean connection strengths above 0.11 with standard deviations around 0.10, LLMs exhibit either comparable means with higher variance (indicating inconsistency) or lower means with concentrated maximums (indicating fragmentation). These structural differences suggest that AI translation organizes meaning construction in more atomistic, hub-dependent ways, producing renderings that may lack the dimensional coordination visible in the reference corpus network. The balanced connectivity across all dimensions in the reference corpus contrasts with the AI tendency toward hub-based architectures, where certain codes dominate while others remain underutilized, a pattern that may bear on translation quality in ways quantitative metrics alone cannot capture, though the reference corpus’s different functional context means this comparison illuminates structural tendencies rather than establishes a performance standard. 5. Discussion 5.1 Between Anthropogenic and Algorogenic Thumbprints The epistemic network analysis reveals a complex relationship between AI and human translator style that resists simple categorization. Through the distinction between anthropogenic and algorogenic thumbprints (see Section 2.1), both human translators and LLMs can be understood as exhibiting identifiable recurrent patterns in Baker’s ( 2000 ) sense. At this level, the findings indicate that translator-like stylistic regularity is not exclusive to human agents. Both human and AI translators in this study satisfy Baker’s criterion of translator style as a “characteristic use of language” that persists across texts. Gu Zhizhong and W.J.F. Jenner display stable tendencies across Creation of the Gods and Journey to the West , while Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro likewise maintain relatively distinctive network configurations across both novels. This suggests that LLMs can produce recognizable and reproducible stylistic signatures, rather than merely generating isolated local solutions. In this respect, the findings are consistent with Saldanha’s ( 2011 ) observation that translator style emerges through recurring patterns of choice. The more revealing difference lies in how these recurrent patterns are organized. Human translators in this study consistently structure their epistemic networks around Semantic Fidelity as the central coordinating node, with strong connections extending to Lexical Choice, Syntactic Structure, and Intertextual Reference (see Section 4). This semantic-first organization suggests that meaning preservation functions as the principle through which other translation decisions are negotiated. By contrast, all three LLMs show stronger centrality around Narrative Flow, while Semantic Fidelity occupies a less dominant position. Because this pattern recurs across both source texts and across all three models, it appears to indicate a broader tendency in LLM literary translation rather than a text-specific effect. The contrast therefore concerns not simply different stylistic preferences, but different ways of organizing translational attention. Comparison with the Western fantasy reference corpus sharpens this interpretation. The reference corpus exhibits denser and more balanced epistemic networks, suggesting a relatively integrated coordination of form, meaning, cultural framing, and stylistic convention. The LLM networks, by contrast, show lower density and greater dependence on a narrower range of code pairings. Particularly revealing is the weak coordination between Syntactic Structure and Cultural Adaptation in the AI networks, a relation that is much stronger in the reference corpus. This suggests that although LLM outputs may be fluent and narratively coherent, they are less effective at integrating multiple translational dimensions into a unified stylistic structure. The comparison should not be read as a direct quality ranking, since the reference corpus consists of original writing rather than translation, but it does help contextualize the structural tendencies observed in the LLM outputs. Taken together, these findings suggest that human and AI translators may display comparable stylistic regularities while organizing translational choices differently. The distinction between anthropogenic and algorogenic thumbprints is useful here because it captures this contrast at the level of pattern formation rather than evaluative judgment. The findings therefore support a reading of translator style (thumbprint) in which recurrent textual signatures can emerge from more than one kind of underlying process, even if the resulting forms of stylistic organization are not equivalent. 5.2 Reconceptualizing the “Thumbprints” in AI translation The distinction between anthropogenic and algorogenic thumbprints has important implications for Baker’s ( 2000 ) “thumbprints” framework. If comparable stylistic regularities can emerge from different generative processes, then translator style cannot be understood only in relation to human agency, intentionality, and socio-cultural situatedness. Baker’s framework remains valuable in identifying translation as a site of recurrent stylistic patterning, but its explanatory assumptions appear less adequate when extended to LLM translation. In this context, the distinction between anthropogenic and algorogenic thumbprints provides a more flexible theoretical vocabulary. It allows translator style to be treated as an observable pattern of translational organization without presupposing a single source for that pattern. In human translation, such regularities may plausibly be related to biography, training, interpretation, and socio-cultural positioning. In LLM translation, comparable regularities are more plausibly associated with model architecture, training distributions, and optimization dynamics. The point is not to deny the existence of regularity in either case, but to avoid treating formally similar outcomes as evidence of the same underlying process. From this perspective, Baker’s framework needs not be rejected, but reformulated. Its central insight, that translation leaves recurrent stylistic traces, remains productive. What requires revision is the assumption that such traces necessarily derive from human subjectivity. The concept of algorogenic thumbprints may therefore serve as a provisional analytical category for describing translator-like regularities produced by non-human systems, without attributing to LLMs the forms of consciousness or agency associated with human translators. In this sense, the term is intended as a heuristic, rather than a definitive theoretical claim, and it remains open to revision as future research examines other model families, language pairs, literary genres, and prompting conditions. This reconceptualization also has methodological implications for translator style research. Traditional approaches have often treated stylistic patterning as evidence of human interpretation and intentionality, making agency both the object and, implicitly, the explanation of analysis. Such assumptions are less applicable in the case of LLM translation, where intentionality is not in itself an adequate explanatory framework. Variables such as training data distributions, architectural constraints, decoding settings, and prompting conditions may instead be more relevant. In this regard, epistemic network analysis is valuable because it models how different translator types organize translational choices relationally without presupposing identical cognitive mechanisms. It thus provides a useful framework for comparing human and algorithmic translators while remaining sensitive to the distinction between anthropogenic and algorogenic forms of stylistic patterning. 5.3 Architecturally Aware Human–AI Collaboration in Literary Translation The epistemic network analysis reveals preliminary implications for how literary translation practice might be organized in an era of increasingly capable AI systems. The structural contrasts identified here, particularly the complementary strengths and limitations of semantic-anchored human networks and narrative-prioritized AI architectures, suggest a theoretical basis for exploring collaborative models. The workflow proposals that follow should, however, be understood as theoretically motivated hypotheses requiring empirical validation rather than evidence-based prescriptions ready for immediate adoption. The network comparisons suggest that LLMs demonstrate comparative strengths in dimensions where systematic pattern application proves advantageous. In this study, Claude Opus 4.5’s consistent narrative-centred architecture and strong Narrative Flow connections (weights: 0.2847–0.3272) indicate reliable capacity for maintaining story coherence across extended passages. Similarly, all three models demonstrate robust Lexical Choice coordination, suggesting potential effectiveness in terminology consistency management. These observations align with Kenny’s ( 2022 ) identification of AI’s advantage in applying linguistic patterns systematically at scale, a capacity potentially valuable for managing the elaborate mythological systems and supernatural terminology that classical Chinese fantasy translation demands, though direct empirical confirmation within this specific literary domain awaits further investigation. The networks simultaneously expose limitations that resonate with findings in existing translation technology research. The systematic absence of Semantic Fidelity as a central organizing principle across all three LLMs represents a weakness in literary translation, where meaning preservation constitutes what Landers ( 2001 ) terms the translator’s primary obligation. Human translators’ semantic-anchored networks demonstrate integrated decision-making that the models examined here do not replicate. This architectural gap finds corroboration in Li’s ( 2024 ) empirical study of AI literary translation of classical Chinese poetry, which found that AI systems face significant limitations in cultural conveyance and translator subjectivity, with human intervention remaining indispensable for preserving the authenticity and depth of literary renderings. More broadly, post-editing research demonstrates that AI-generated literary output consistently requires substantial human revision, particularly where cultural density and syntactic complexity coincide, precisely the conditions characteristic of classical Chinese narrative texts (Jiang et al., 2025 ). These patterns suggest a possible epistemic division of labour worth testing in future studies. In principle, the narrative-centred processing of LLMs could inform initial draft generation, maintaining story coherence and lexical consistency across extended passages, while human translators apply semantic-anchored revision, strengthening meaning preservation, cultural adaptation, and the coordinated decision-making that AI networks systematically underutilize. However, it is worth stressing that this specific task allocation has not been empirically tested in classical Chinese literary translation. Whether the quality improvements documented in adjacent post-editing domains (Jiang et al., 2025 ) extend to the demands of classical literary texts requires controlled experimental investigation before any practical implementation can be responsibly recommended. Realizing such collaboration would also require addressing challenges the analysis brings into relief. The architectural divergence between semantic-anchored human networks and narrative-prioritized AI networks may create integration difficulties: AI-generated passages organized around narrative coherence could resist revision aimed at reorienting them around semantic fidelity without near-complete retranslation. As Lee ( 2024 ) cautions from a posthumanist perspective, over-reliance on AI augmentation risks eroding translators’ epistemic engagement with source texts, potentially weakening the very semantic-anchored architecture that distinguishes expert human translation. These challenges reinforce that the collaborative framework sketched here is a research agenda, not a deployable workflow model. The future of literary translation may therefore lie in developing what might be conceptualized as “architecturally aware collaboration”, an approach that recognizes human and AI systems as constructing meaning through different epistemic structures, each offering distinct capabilities for mediating classical Chinese literature across cultural boundaries. This study provides a conceptual framework and preliminary evidence for such an approach; the empirical validation of the collaborative model itself constitutes the necessary next step for translating these theoretical insights into practice. 6. Conclusion This study set out to examine whether LLMs develop consistent stylistic signatures when translating classical Chinese fantasy, and whether such signatures are structurally comparable to those of human translators. The ENA findings confirm that they are not, but the more consequential contribution lies in specifying how they differ and what that difference demands theoretically. The three research questions yield a coherent interpretive picture: LLMs exhibit distinct, cross-textually stable epistemic architectures organized around narrative coherence; human translators systematically anchor meaning-making in semantic fidelity; and both diverge structurally from the balanced, densely interconnected networks characteristic of original English fantasy writing. Taken together, these findings do not simply indicate that AI translates differently, they reveal that the epistemic network structures observed in this study differ systematically between the LLM and human translations. This distinction carries a consequence Baker’s ( 2000 ) original framework was not designed to accommodate: that stylistic consistency can be generated without consciousness, intentionality, or socio-cultural situatedness. The anthropogenic/algorogenic distinction proposed here is therefore not a terminological refinement but a conceptual necessity, one that repositions translator style theory from a human-centred account toward a broader framework capable of modelling meaning-making across hybrid human-AI translation ecologies increasingly characteristic of contemporary literary practice. Future work should empirically test whether algorogenic architectures are prompt-sensitive, examine whether the semantic/narrative divergence identified here persists across non-English-centric language pairs, and investigate whether architecturally aware human-AI collaboration can produce networks that neither humans nor models achieve independently. Declarations Author Contribution Author contributions: Conceptualisation: Kan Wu, Siqi Jiang & Defeng Li.Supervision: Defeng Li. Data collection: Kan Wu & Siqi Jiang. Methodology: Kan Wu. Data analysis and interpretation: Kan Wu. Writing—original draft: Kan Wu & Siqi Jiang. Writing—revising and proofreading: Kan Wu & Siqi Jiang. Acknowledgement This study is supported by the Zhejiang Provincial Philosophy and Social Science Planning Annual Project (26NDJC277YB). Data Availability Dataset related to this study is available at: 10.6084/m9.figshare.31915422 References Baker M (2000) Towards a methodology for investigating the style of a literary translator. Target 12(2):241–266 Baker M (2018) In other words: a coursebook on translation. Routledge, London Boase-Beier J (2006) Stylistic approaches to translation. Routledge, London Bosseaux C (2004) Point of view in translation: a corpus-based study of French translations of Virginia Woolf’s To the Lighthouse. Lang Cult 5(1):107–122 Bosseaux C (2007) How does it feel? Point of view in translation: the case of Virginia Woolf into French. Rodopi, Amsterdam Cacioppo S, Cacioppo JT (2012) Decoding the invisible forces of social connections. Front Integr Neurosci 6:51 Chesterman A (2016) Memes of translation: the spread of ideas in translation theory, revised edn. John Benjamins, Amsterdam Csanadi A, Eagan B, Kollar I, Shaffer DW, Fischer F (2018) When coding-and-counting is not enough: using epistemic network analysis (ENA) to analyze verbal data in CSCL research. Int J Comput-Support Collab Learn 13(4):419–438 Elkins K (2024) In search of a translator: using AI to evaluate what’s lost in translation. Front Comput Sci 6:1444021 Elmoazen R, Saqr M, Tedre M, Hirsto L (2022) A systematic literature review of empirical research on epistemic network analysis in education. IEEE Access 10:17330–17348 Gao R, Lin Y, Zhao N, Cai ZG (2024) Machine translation of Chinese classical poetry: a comparison among ChatGPT, Google Translate, and DeepL Translator. Humanit Soc Sci Commun 11(1):1–10 Han C, Lu X (2025) Beyond BLEU: repurposing neural-based metrics to assess interlingual interpreting in tertiary-level language learning settings. Res Methods Appl Linguist 4(1):100184 House J (2015) Translation quality assessment: past and present. Routledge, London Hsia CT (2016) The classic Chinese novel: a critical introduction. Chinese University of Hong Kong, Hong Kong Huang L (2015) Style in translation: a corpus-based perspective. Springer, Berlin Jakobsen AL, Jensen KTH (2008) Eye movement behaviour across four different types of reading task. Cph Stud Lang 36:103–124 Jiang L, Wei B, Al-Shaibani GKS (2025) Effective neural machine translation with human post-editing of Chinese intangible cultural heritage corpus into English. SAGE Open 15(4):21582440251386954 Karpinska M, Iyyer M (2023) Large language models effectively leverage document-level context for literary translation, but critical errors persist. arXiv:2304.03245 Kenny D (2022) Machine translation for everyone: empowering users in the age of artificial intelligence. Language Science, Berlin Landers CE (2001) Literary translation: a practical guide. Multilingual Matters, Clevedon Läubli S, Sennrich R, Volk M (2018) Has machine translation achieved human parity? a case for document-level evaluation. arXiv:1808.07048 Lee TK (2024) Artificial intelligence and posthumanist translation: ChatGPT versus the translator. Appl Linguist Rev 15(6):2351–2372 Li D (2016) Translator style: a corpus-assisted approach. In: Ji M, Oakes M, Li D, Hareide L (eds) Corpus methodologies explained: an empirical approach to translation studies. Routledge, London, pp 103–136 Li Q (2024) Bridging languages: the potential and limitations of AI in literary translation—a case study of the English translation of ‘A Pair of Peacocks Southeast Fly’. Adv Humanit Res 8(1):1–7 Malmkjær K (2003) What happened to God and the angels: an exercise in translational stylistics. Target 15(1):37–58 Mastropierro L (2018) Key clusters as indicators of translator style. Target 30(2):240–259 Mikros G (2025) Beyond the surface: stylometric analysis of GPT-4o’s capacity for literary style imitation. Digit Scholarsh Humanit 40(2):587–600 Olohan M (2003) How frequent are the contractions? a study of contracted forms in the translational English corpus. Target 15(1):59–89 O’Sullivan J (2025) Stylometric comparisons of human versus AI-generated creative writing. Humanit Soc Sci Commun 12:1708 Ping Y, Wang B (2024) Retranslated Chinese classical canon Journey to the West : a stylometric comparison between Julia Lovell’s retranslation and Arthur Waley’s translation. Digit Scholarsh Humanit 39(1):308–320 Pym A (2015) Translating as risk management. J Pragmat 85:67–80 Ruis AR, Rosser AA, Quandt-Walle C, Nathwani JN, Shaffer DW, Pugh CM (2018) The hands and head of a surgeon: modeling operative competency with multimodal epistemic network analysis. Am J Surg 216(5):835–840 Saldanha G (2011) Translator style: methodological considerations. Translator 17(1):25–50 Shaffer D (2018) Transforming big data into meaningful insights: introducing quantitative ethnography. Scientia, Madison Shaffer DW, Collier W, Ruis AR (2016) A tutorial on epistemic network analysis: analyzing the structure of connections in epistemic, social, and interaction data. J Learn Anal 3(3):9–45 Shaffer DW, Hatfield D, Svarovsky GN, Nash P, Nulty A, Bagley E et al (2009) Epistemic network analysis: a prototype for 21st-century assessment of learning. Int J Learn Media 1(2):33–53 Siebert-Evenstone AL, Irgens GA, Collier W, Swiecki Z, Ruis AR, Shaffer DW (2017) In search of conversational grain size: modeling semantic structure using moving stanza windows. J Learn Anal 4(3):123–139 Vandepitte S, Hartsuiker RJ (2011) Metonymic language use as a student translation problem: towards a controlled psycholinguistic investigation. In: Alvstad C, Hild A, Tiselius E (eds) Methods and strategies of process research. John Benjamins, Amsterdam, pp 67–92 Wang Q, Li D (2011) Looking for translator’s fingerprints: a corpus-based study on Chinese translations of Ulysses . Lit Linguist Comput 27(1):81–93 Winters M (2004) German translations of F. Scott Fitzgerald’s The Beautiful and Damned: a corpus-based study of modal particles as features of translators’ style. In: Kemble I (ed) Using corpora and databases in translation. University of Portsmouth, Portsmouth, pp 71–88 Wu K, Li D (2024) Unraveling Eileen Chang’s stylistic multiverse: insights from multivariate analysis with multifactorial design. Digit Scholarsh Humanit 39(3):1001–1018 Yao X, Kang Y-B, McCosker A (2025) Missing the human touch? a computational stylometric analysis of GPT-4 translations of online Chinese literature. Transl Spaces 14(2):303–330 Zhang R, Zhao W, Eger S (2025) How good are LLMs for literary translation, really? literary translation evaluation with humans and LLMs. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, vol 1. p 10961–10988 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 11 May, 2026 Reviews received at journal 10 May, 2026 Reviews received at journal 04 May, 2026 Reviews received at journal 27 Apr, 2026 Reviewers agreed at journal 27 Apr, 2026 Reviewers agreed at journal 26 Apr, 2026 Reviewers agreed at journal 22 Apr, 2026 Reviewers agreed at journal 21 Apr, 2026 Reviewers invited by journal 20 Apr, 2026 Editor assigned by journal 20 Apr, 2026 Editor invited by journal 17 Apr, 2026 Submission checks completed at journal 17 Apr, 2026 First submitted to journal 17 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9323835","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":630424412,"identity":"ae1f12ec-ad90-41f6-b587-1935e8eccb33","order_by":0,"name":"Kan Wu","email":"","orcid":"","institution":"Zhejiang University of Finance and Economics Dongfang College","correspondingAuthor":false,"prefix":"","firstName":"Kan","middleName":"","lastName":"Wu","suffix":""},{"id":630424413,"identity":"dd1d26fc-f184-43c5-8ce0-0cb8d01dd62c","order_by":1,"name":"Siqi Jiang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAu0lEQVRIie3QsQrCMBCA4QuBTKVzXFof4aTQyYe5LslSQfAF6mP4FnXpHCjYpeAmQibfoC5dLKijm9dNMP92kI/jAhAK/WSye0yokjlEkY62cTaHRKijISkqNsDOIWrMbX1pb4Mo1wzSExGi2dTeZFo0hkGu4Bzh6U0oB9G0HCL2lcOnRW9HLpFSVKgIfcncsuiNkoBqdfDlThecW+LuPEqYVBp7exzuDePHlu5zou8AIK04r0KhUOi/ewEJ+jqPHEuqNgAAAABJRU5ErkJggg==","orcid":"","institution":"University of Macau","correspondingAuthor":true,"prefix":"","firstName":"Siqi","middleName":"","lastName":"Jiang","suffix":""},{"id":630424415,"identity":"af52c2ba-22b8-4e55-b72b-dc1901290f34","order_by":2,"name":"Defeng Li","email":"","orcid":"","institution":"University of Macau","correspondingAuthor":false,"prefix":"","firstName":"Defeng","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2026-04-05 03:38:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9323835/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9323835/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108106437,"identity":"540917dd-d963-461f-8680-031ead4d573e","added_by":"auto","created_at":"2026-04-29 11:55:16","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":506701,"visible":true,"origin":"","legend":"\u003cp\u003eData and corpus design\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/750cfe9ed2bf660839303392.png"},{"id":108106436,"identity":"6621c687-9eef-48d1-a7e8-653d94a94878","added_by":"auto","created_at":"2026-04-29 11:55:16","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1259872,"visible":true,"origin":"","legend":"\u003cp\u003eEpistemic networks of Claude Opus 4.5 for \u003cem\u003eCG \u003c/em\u003e(Left) and \u003cem\u003eJTW \u003c/em\u003e(Right)\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/1f34e743812242e97e6ec17f.png"},{"id":108182781,"identity":"c68bb721-a989-4b7e-aab7-2bd813e8f1d5","added_by":"auto","created_at":"2026-04-30 08:59:32","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1304853,"visible":true,"origin":"","legend":"\u003cp\u003eEpistemic networks of ChatGPT 5.2 Pro for \u003cem\u003eCG \u003c/em\u003e(Left) and \u003cem\u003eJTW \u003c/em\u003e(Right)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/7c2faaf8826b292a8d9c77f9.png"},{"id":108181733,"identity":"6c224c48-1bd6-4ec0-ad8e-acc495a0e0d1","added_by":"auto","created_at":"2026-04-30 08:58:51","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1079261,"visible":true,"origin":"","legend":"\u003cp\u003eEpistemic networks of Gemini 3.0 Pro for \u003cem\u003eCG \u003c/em\u003e(Left) and \u003cem\u003eJTW \u003c/em\u003e(Right)\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/365efcc7eaab7aea143eccf5.png"},{"id":108106441,"identity":"eea77355-27f1-46fc-a5f2-123efdb53f32","added_by":"auto","created_at":"2026-04-29 11:55:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1010395,"visible":true,"origin":"","legend":"\u003cp\u003eEpistemic networks of human translators for \u003cem\u003eCG \u003c/em\u003e(Left) and \u003cem\u003eJTW \u003c/em\u003e(Right)\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/3b13d3e02268f9d914806c3f.png"},{"id":108182447,"identity":"4e577155-1a15-4119-9b73-751c3fe95919","added_by":"auto","created_at":"2026-04-30 08:59:22","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":340636,"visible":true,"origin":"","legend":"\u003cp\u003eEpistemic networks of the reference corpus\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/a1eebad9d2a684479a79ef4a.png"},{"id":108183842,"identity":"ad67a956-503e-4541-a3cc-2e76355c0031","added_by":"auto","created_at":"2026-04-30 09:02:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5839492,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9323835/v1/d1eac87d-40c6-4883-a553-8b4633eae9f2.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Algorogenic Thumbprints in Literary Translation: An Epistemic Network Analysis of LLM Translations of Classical Chinese Fantasy","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eWhen classical Chinese fantasy novels traverse linguistic boundaries in the digital age, they raise not merely translation challenges but deeper questions about how artificial intelligence organize linguistic and stylistic choices when rendering culturally embedded narratives. \u003cem\u003eJourney to the West\u003c/em\u003e and \u003cem\u003eCreation of the Gods\u003c/em\u003e, canonical works of the Chinese Gods and Demons genre, have long fascinated Western readers through their mythological systems, supernatural battles, and philosophical depth (Hsia, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). As large language models (LLMs) increasingly take on literary translation tasks, these texts become productive testing grounds for a focused inquiry: how do LLMs make decisions in translation when rendering culturally embedded narratives, and do these decision-making patterns differ meaningfully from those of human translators?\u003c/p\u003e \u003cp\u003eThis question gains urgency as LLMs transform translation workflows. Models such as Claude, ChatGPT, and Gemini now produce translations at unprecedented speed and scale, prompting both enthusiasm about democratized access to world literature and concern about the erosion of translator craft (Kenny, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Yet, amid ongoing debates about quality and authenticity, the epistemic architecture underlying AI translation decisions remains underexplored, particularly whether different LLMs develop distinctive signatures in coordinating competing demands such as lexical precision, cultural adaptation, and stylistic register.\u003c/p\u003e \u003cp\u003eTranslation quality assessment has traditionally focused on error analysis and fluency metrics (L\u0026auml;ubli et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), while computational approaches tend to emphasize neural network architectures rather than decision-making patterns (Elmoazen et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Recent computational stylistics has begun addressing this gap through stylometric studies of LLM translations (Yao et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Ping and Wang, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and AI creative writing (O\u0026rsquo;Sullivan, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Elkins, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mikros, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), suggesting that LLMs may develop identifiable stylistic signatures. Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) influential framework of translator thumbprints, which posits that translators leave varying styles through systematic linguistic choices, has inspired substantial research on human translators but remains largely unapplied to AI systems. Process-oriented methods such as keystroke logging and eye-tracking (Jakobsen and Jensen, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2008\u003c/span\u003e) are not readily transferable to LLM cognition. What is needed, then, is a method capable of modeling not isolated translation choices, but the relational structures through which meaning is constructed.\u003c/p\u003e \u003cp\u003eEpistemic network analysis (ENA) offers one such approach. Originally developed to model complex thinking in collaborative learning environments, ENA quantifies and visualizes co-occurrence patterns among coded epistemic activities as weighted network graphs (Shaffer, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Unlike traditional content analysis, ENA captures how elements function relationally within a system, making it particularly suited to translation, where decisions are inherently interdependent (Csanadi et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). While ENA has been applied productively in studies of professional practice (Ruis et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) and educational discourse (Siebert-Evenstone et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), it remains largely underexplored in translation studies.\u003c/p\u003e \u003cp\u003eThis study applies ENA to compare three LLMs, Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro, translating \u003cem\u003eCreation of the Gods\u003c/em\u003e and \u003cem\u003eJourney to the West\u003c/em\u003e against published human translations and a Western heroic fantasy reference corpus. Systematic coding across seven decision-making domains enables construction of epistemic networks revealing how each translator coordinates translational choices. Specifically, three research questions guide the analysis:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHow do the epistemic networks of the three LLMs, Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro, differ in their translations of classical Chinese fantasy?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHow do the epistemic networks of LLM translations differ from those of human translations of classical Chinese fantasy?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eHow do the epistemic networks of LLM translations differ from the network structure of the Western fantasy reference corpus?\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e"},{"header":"2. Related Theories","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Baker\u0026rsquo;s Thumbprint Framework and Beyond\u003c/h2\u003e \u003cp\u003eBaker (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e: 245) conceptualizes translator style as a \u003cem\u003ethumbprint\u003c/em\u003e, that is, a translator\u0026rsquo;s characteristic use of language and individual profile of linguistic habits in comparison with other translators. This target-oriented perspective treats stylistic idiosyncrasies as operating independently of any particular source text (Saldanha, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2011\u003c/span\u003e), in contrast to source-oriented approaches advocated by Malmkj\u0026aelig;r (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2003\u003c/span\u003e) and Boase-Beier (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2006\u003c/span\u003e), which view translator style as responsive to source text features. These divergent perspectives have shaped distinct methodological trajectories in translator style research, yet both presuppose human consciousness, with its biographical accumulation, socio-cultural embedding, and creative intentionality, as the generative force behind stylistic choices. This presupposition is one this study interrogates rather than inherits uncritically.\u003c/p\u003e \u003cp\u003eThe framework\u0026rsquo;s human-centric foundations become particularly complicated when extended to LLMs. Even within human translation, distinguishing \u0026ldquo;deliberate\u0026rdquo; from \u0026ldquo;unconscious\u0026rdquo; stylistic choices remains methodologically elusive (Saldanha, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Winters, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). This difficulty is compounded further when the translator is algorithmic, as LLM patterns emerge not from accumulated biographical experience, but from statistical regularities encoded across training data (Ping and Wang, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Yet, the consistency and cross-textual stability of such patterns, as suggested by recent computational stylistics research (Yao et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Mikros, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), suggests that the absence of human consciousness may not preclude identifiable stylistic signatures. This study therefore extends Baker\u0026rsquo;s (ibid.) framework to examine whether LLMs develop consistent stylistic signatures and, if so, whether the mechanisms generating such patterns are structurally comparable to those underlying human translation choices or constitute a categorically distinct phenomenon.\u003c/p\u003e \u003cp\u003eTraditional translator style studies employ two primary comparative modes: the \u0026ldquo;one-to-many mode,\u0026rdquo; examining multiple translations of a single source text, and the \u0026ldquo;many-to-many mode,\u0026rdquo; comparing translations across varied texts (Wu and Li, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, p. 273). Winters (\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2004\u003c/span\u003e) exemplifies the one-to-many approach through his analysis of loanwords and modal particles in translations of \u003cem\u003eThe Beautiful and Damned\u003c/em\u003e, while Mastropierro (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) identifies distinctive stylistic markers across Italian translations of \u003cem\u003eLovecraft\u003c/em\u003e. Baker (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) herself employs the many-to-many mode, tracing patterns in type/token ratio and reporting verbs; Olohan (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2003\u003c/span\u003e) and Saldanha (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) similarly examine contraction preferences and italicization practices. Source-oriented researchers such as Bosseaux (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2004\u003c/span\u003e; \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2007\u003c/span\u003e) and Huang (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) focus instead on how translators reproduce narrative voice through deixis, modality, transitivity, and free indirect discourse.\u003c/p\u003e \u003cp\u003eThe existing literature thus leaves a clear gap: although Baker\u0026rsquo;s framework and its later extensions have been widely applied to human translators, their relevance to algorithmic translation remains largely untested. It is still unclear whether LLMs exhibit consistent stylistic signatures comparable to human translator thumbprints or a categorically different kind of patterning. To address this issue, the present study distinguishes between \u003cem\u003eanthropogenic thumbprints\u003c/em\u003e and \u003cem\u003ealgorogenic thumbprints\u003c/em\u003e. Anthropogenic thumbprints arise from human translators\u0026rsquo; agency, experience, and socio-cultural embeddedness, while algorogenic thumbprints from statistical training, probabilistic prediction, and optimization in large language models. This heuristic distinction is proposed to evaluate whether Baker\u0026rsquo;s framework can be meaningfully extended to algorithmic translation. With this distinction in place, the next section turns to the role of LLMs in literary translation and their current translational capabilities.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Large Language Models as Literary Translators\u003c/h2\u003e \u003cp\u003eMachine translation has been substantially transformed by the emergence of AI-powered large language models (Han and Lu, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). ChatGPT exemplifies this shift. Though not specifically designed for translation, its contextual understanding and coherent text generation capabilities have shown competitive performance relative to conventional MT systems in some translation settings, positioning such models as potential translators\u0026rsquo; prostheses (Lee, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These developments raise questions about whether AI systems can function as literary translators, a role traditionally associated with human expertise in what Landers (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2001\u003c/span\u003e: 7) termed \u0026ldquo;the most demanding type of translation.\u0026rdquo;\u003c/p\u003e \u003cp\u003eResearch on LLMs as literary translators has broadly developed along two trajectories: evaluative studies of translation quality and descriptive studies of translational behaviour. Evaluative research has moved beyond narrow measures of accuracy toward multidimensional frameworks that incorporate stylistic considerations, including document-level performance, literary fluency, coherence, and sensitivity to tone, rhythm, and culturally specific expression (Karpinska and Iyyer, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Gao et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Zhang et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). At the same time, these studies suggest that although LLMs have made clear gains in literary translation, human translators continue to retain an advantage in overall quality, particularly when stylistic and terminological factors are taken into account (Zhang et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWhile this quality-oriented line of research provides important evidence of LLMs\u0026rsquo; literary translation capabilities, its normative orientation leaves a more basic question insufficiently addressed: whether LLMs exhibit distinctive translational patterns comparable to Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) translator \u0026ldquo;thumbprints.\u0026rdquo; Descriptive research offers a complementary perspective by examining recurrent linguistic features in LLM-generated translations without reducing analysis to predefined standards of adequacy. Emerging work (Lee, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Li, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Yao et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Mikros, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) in this area points to the possibility that LLMs can display consistent tendencies in the management of cohesion, coherence, source-text restructuring, and stylistic patterning, raising the broader question of whether such regularities may be understood as a form of translational voice or style.\u003c/p\u003e \u003cp\u003eNonetheless, descriptive research on LLMs as literary translators remains limited, particularly with regard to consistency across texts, genres, and models. While Lee (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) offers qualitative insight into the linguistic behaviour of a single model, the field still lacks systematic empirical investigation into whether different LLMs develop model-specific styles in translation and how such styles compare with those of human translators. This gap is especially significant because Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) framework is grounded in assumptions of human consciousness, agency, and socio-cultural positioning, assumptions that cannot simply be transferred to LLMs. If stylistic analysis in human translation often relies, at least implicitly, on the idea of an intentional and socially situated subject, then the study of LLM translation requires a different analytical basis: one that does not depend on access to inner consciousness or inferred motivation, but instead focuses on observable patterns in how translational choices are coordinated. For this reason, what is needed is not only the identification of surface-level stylistic features, but also a method for modelling the relational structures through which such features co-occur across translations. Epistemic network analysis offers precisely this kind of relational account, and the following section considers its relevance for translation research.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Epistemic Networks in Translator Style Analysis\u003c/h2\u003e \u003cp\u003eTraditional approaches to translator style analysis, including foundational work by Baker (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) and subsequent studies (Li, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu and Li, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), have relied primarily on frequency-based methods to identify stylistic patterns. While such methods have proven productive, they face notable limitations. Conventional coding-and-counting strategies tend to treat translational features in isolation, ignoring temporality in data and the broader patterns connecting translational activities (Csanadi et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Counting how often a translator employs explicitation, for instance, or how frequently they adopt a formal register, reveals little about whether these choices are epistemically linked in the translator\u0026rsquo;s decision-making. What such approaches largely cannot capture is the relational structure of translation, not merely which strategies occur, but how they co-occur and interconnect.\u003c/p\u003e \u003cp\u003eEpistemic Network Analysis (ENA) was developed to address precisely this kind of structural question by modelling patterns of association among elements in coded data (Shaffer et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Shaffer, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Originally designed to examine epistemic networks, the patterns of association among knowledge, skills, values, and habits of mind that characterize complex thinking, ENA quantifies, visualizes, and enables statistical comparison of such structures. A key assumption of ENA aligns well with conceptualizing translation as an epistemic process: that the structure of connections among elements may be more analytically meaningful than the mere presence or absence of those elements in isolation (Cacioppo and Cacioppo, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). Applied to translator style, ENA thus enables researchers to model not simply which translational strategies appear, but how these strategies are relationally organized across a translation.\u003c/p\u003e \u003cp\u003eMethodologically, ENA identifies co-occurrences of coded elements within defined data segments, called \u003cem\u003estanzas\u003c/em\u003e, and represents them as weighted connections in network models (Shaffer et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Each translator\u0026rsquo;s network thereby captures both the frequency and strength of associations among different translational choices. Dimensional reduction and optimization techniques then position network nodes such that the centroid of each network corresponds to its location in a projected space, facilitating both visual and statistical comparison across translators (Shaffer et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). This co-registration ensures that differences in network visualizations correspond directly to quantifiable structural differences, lending interpretive clarity to statistical findings.\u003c/p\u003e \u003cp\u003eThus, ENA provides several methodological advantages for translator style analysis. It captures the relational nature of translation decision-making by modelling how choices in one domain, for example lexical selection, connect systematically with choices in another, including syntactic restructuring. Network visualizations also provide more interpretable representations of complex associative patterns than conventional statistical outputs (Csanadi et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Furthermore, unlike sequential analysis methods that typically require large datasets and can produce difficult-to-interpret transition probabilities, ENA is well-suited to relatively small, fixed sets of elements with complex patterns of association (Csanadi et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Individual translators\u0026rsquo; networks can each be represented as a point in projected space, where proximity indicates similarity in translational patterning, and standard statistical tests can then determine whether different translator groups exhibit meaningfully different epistemic architectures.\u003c/p\u003e \u003cp\u003eThese properties make ENA particularly well-suited to investigating LLMs as literary translators. If LLMs do develop distinctive \u0026ldquo;algorogenic thumbprints,\u0026rdquo; such patterns would likely be expressed not in isolated feature frequencies alone, but in systematic structures of association among translational choices. ENA offers a principled means of operationalizing Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) insight that translator thumbprints emerge from habitual patterns of choice-making, while extending this framework to encompass AI algorithmic processing alongside human consciousness.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Methodology","content":"\u003cp\u003eTo examine whether LLMs develop consistent stylistic signatures comparable to, or categorically different from, human thumbprints, the study performs epistemic network analysis (ENA) across three comparative dimensions: variations and similarities among LLM models, between LLMs and human translators, and between LLM translations and non-translated original English texts. The following details the research design and procedures.\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Data and Design\u003c/h2\u003e \u003cp\u003eThe corpus design (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) includes both translated and non-translated fantasy literature, creating a comparative framework for the analysis of epistemic structures in literary translation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe translated component features English translations of two Chinese classical novels (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e): 封神演義 (\u003cem\u003eCreation of the Gods\u003c/em\u003e, CG) and 西遊記 (\u003cem\u003eJourney to the West\u003c/em\u003e, JTW), representing the Gods and Demons subgenre of Chinese fantasy. These translations include human translations (Gu Zhizhong\u0026rsquo;s 1992 CG and W.J.F. Jenner\u0026rsquo;s 1993 JTW) and LLM translations of these works from three models: Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro.\u003c/p\u003e \u003cp\u003eThe selected chapters from both Chinese novels represent diverse thematic and stylistic elements. For CG: Chap.\u0026nbsp;1 introduces mythological themes; Chap.\u0026nbsp;12 features Nezha\u0026rsquo;s birth with Taoist terminology; Chap.\u0026nbsp;21 focuses on political intrigue; Chap.\u0026nbsp;70 presents battles with religious elements; Chap.\u0026nbsp;79 combines supernatural and military aspects; Chap.\u0026nbsp;100 concludes with ceremonial language. For JTW: Chap.\u0026nbsp;1 establishes the stone monkey\u0026rsquo;s origin; Chap.\u0026nbsp;7 details Sun Wukong\u0026rsquo;s celestial confrontation; Chap.\u0026nbsp;15 introduces pilgrimage\u0026rsquo;s spiritual dimensions; Chap.\u0026nbsp;27 demonstrates humour and action; Chap.\u0026nbsp;47 presents Buddhist concepts; Chap.\u0026nbsp;68 showcases allegorical supernatural encounters.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTranslations by human and LLMs and their token size\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTranslation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTranslator\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eToken size\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cem\u003eCreation of the Gods\u003c/em\u003e (CG) Chap.\u0026nbsp;1, 12, 21, 70, 79, 100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHuman by Gu Zhizhong\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e21,420\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM by ChatGPT 5.2 Pro\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17,648\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM by Claude Opus 4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19,021\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM by Gemini 3 Pro\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15,487\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cem\u003eJourney to the West\u003c/em\u003e (JTW)\u003c/p\u003e \u003cp\u003eChapter\u0026nbsp;1, 7, 15, 27, 47, 68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHuman by W.J.F. Jenner\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e27,552\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM by ChatGPT 5.2 Pro\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e25,704\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM by Claude Opus 4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e28,211\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM by Gemini 3 Pro\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e18,266\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDetails of the reference corpus\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWestern Heroic Fantasies\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYear\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAuthor\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eToken Size\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eThe Lord of the Rings 1\u0026ndash;2\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1954\u0026ndash;1955\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eJ. R. R. Tolkien\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e188,361\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eA Song of Ice and Fire 1\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1996\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eG. R. R. Martin\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e293,856\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eConan the Barbarian\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1954\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR. E. Howard\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e104,972\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eWheel of Time 1\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1990\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR. Jordan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e319,124\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eThe Chronicles of Amber 1\u0026ndash;4\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1970\u0026ndash;1976\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR. Zelazny\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e239,940\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eThe Chronicles of Narnia 1\u0026ndash;5\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1950\u0026ndash;1954\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC. S. Lewis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e233,369\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal Size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003e1,379,622\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe non-translated component includes six influential works of Western heroic fantasy (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), totalling approximately 1.38\u0026nbsp;million words. This reference corpus is used as a broad stylistic comparator for fantasy writing in English. Because it consists of original English texts rather than translations, and because it is much larger than the translated dataset, it is not treated as a directly equivalent baseline for assessing translation quality, translator expertise, or translationese. The function of the reference corpus is limited to providing a contextual point of comparison for interpreting whether certain network tendencies in the translated corpus appear more translation-specific or more broadly genre-compatible.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Framework and Procedures\u003c/h2\u003e \u003cp\u003eThe analytical framework defines seven decision-making domains for describing salient translational features in the English outputs: lexical choice (word selection and terminology decisions), syntactic structure (sentence construction and clause arrangement), cultural adaptation (handling culture-specific items), stylistic register (tone, formality, and genre conventions), semantic fidelity (preserving source text meaning), narrative flow (coherence and readability), and intertextual reference (allusions and literary echoes). Informed by House\u0026rsquo;s (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) model of translation quality assessment and Vandepitte and Hartsuiker\u0026rsquo;s (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) categorization of translation problems, these categories describe observable translation solutions to multidimensional demands (Pym, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) and broadly align with Chesterman\u0026rsquo;s (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) account of translation universals. They are used as descriptive analytic constructs rather than direct indicators of internal decision-making, enabling recurrent co-occurrence patterns to be modeled through epistemic network analysis.\u003c/p\u003e \u003cp\u003eThe research unfolds in four systematic steps. Step 1: Translation Generation and Segmentation. All AI translations were generated via each model\u0026rsquo;s official API rather than web-based interfaces, ensuring controlled parameter settings and minimizing platform-side variability. A standardized, stylistically neutral prompt was employed across all three models: \u0026ldquo;\u003cem\u003eYou are a translation assistant. Translate the following Chinese text into English. Output only the translated text, preserving the original paragraph structure and line breaks exactly as they appear\u003c/em\u003e.\u0026rdquo; The prompt imposed no stylistic preferences, register directives, or target-audience specifications, ensuring that observed differences reflect each model\u0026rsquo;s default translational tendencies rather than prompt-induced behavior. Temperature was set to 0 across all API calls, minimizing stochastic sampling variation and producing deterministic, reproducible outputs, thereby ensuring stable and consistent translation versions across the corpus. It should be noted, however, that zero-temperature sampling may not fully represent each model\u0026rsquo;s characteristic stylistic range; the thumbprints identified here thus reflect deterministic rather than fully naturalistic translational behaviour. All translations were subsequently segmented into 200-word passages, balancing contextual sufficiency with analytical granularity.\u003c/p\u003e \u003cp\u003eStep 2: Manual Coding and Reliability Assessment. All segments were independently coded by four PhD-trained coders in translation studies using the predefined codebook (see Appendix A). Coding was conducted at the segment level, and multiple codes could be assigned when more than one domain was salient. Following iterative calibration, the coders re-annotated a shared subset comprising 20% of all segments, sampled proportionally across texts and translator/model outputs. Because four coders were involved, inter-rater reliability for each code was calculated as the mean of the six pairwise Cohen\u0026rsquo;s kappa values. Acceptable agreement was defined a priori as κ\u0026thinsp;\u0026ge;\u0026thinsp;0.80. Final kappa values are reported in Appendix C. Remaining disagreements were then resolved through discussion to produce the final coding dataset.\u003c/p\u003e \u003cp\u003eStep 3: Data Preparation and Stanza Configuration. Following ENA methodology, coded segments were organized into stanza-based interaction data (Shaffer, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Stanzas, defined as moving windows of 5 consecutive lines, establish which translation decisions co-occur within meaningful units of context (Siebert-Evenstone et al., \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). This window size was selected through pilot testing as a balance between local context and analytical tractability, producing stable and interpretable networks for the present dataset. For each stanza, ENA creates an adjacency matrix representing co-occurrences among the seven domains, then accumulates these matrices into a cumulative adjacency matrix for each translator or model, quantifying the total connections made across all stanzas. Python scripts (version 3.14) formatted coded transcripts for ENA input.\u003c/p\u003e \u003cp\u003eStep 4: Network Generation and Visualization. Python scripts utilizing the pyENA library quantified code co-occurrences and accumulated connection strengths across stanzas. ENA employs singular value decomposition (SVD) for dimensional reduction while optimizing node positions such that each network\u0026rsquo;s centroid corresponds to its location in the reduced space, enabling meaningful visualization interpretation (Shaffer et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). The analysis generated individual and mean network visualizations for comparing network density, domain centrality, and connection weights. Goodness-of-fit statistics (Pearson and Spearman correlations\u0026thinsp;\u0026gt;\u0026thinsp;0.90) indicated acceptable model fit for the ENA projections used in this study.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Epistemic Networks between Different LLMs\u003c/h2\u003e \u003cp\u003eThe epistemic network analysis reveals that while all three models, Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3.0 Pro, demonstrate competence in handling classical Chinese texts, they exhibit markedly different epistemic strategies in processing and rendering these works into English, with distinct patterns emerging across the two literary corpora.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFor \u003cem\u003eCreation of the Gods\u003c/em\u003e (CG), the three models display different network architectures visible in the left panels of Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. Claude Opus 4.5 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, left) generated 96 coded units with 15 active connections, achieving a mean connection strength of 0.1796. This relatively high connection density suggests Opus activates multiple translation codes simultaneously when processing CG passages, creating a richly interconnected epistemic landscape. The model\u0026rsquo;s strongest connections appear between Narrative Flow and Lexical Choice (weight: 0.2847), followed by Narrative Flow and Intertextual Reference (weight: 0.2753). Notably, Opus displays prominent connections between Cultural Adaptation and both Lexical Choice (weight: 0.2478) and Intertextual Reference (weight: 0.2478), suggesting this model more actively considers target culture accessibility when processing culturally-embedded passages compared to its counterparts.\u003c/p\u003e \u003cp\u003eChatGPT 5.2 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, left) analysed 64 units across 21 connections with a mean strength of 0.1452, displaying the most distributed network pattern among the three models for CG. Its strongest link connects Narrative Flow and Lexical Choice (weight: 0.3331), though the relatively lower mean connection strength compared to Opus suggests a more exploratory strategy that considers multiple translation dimensions with less concentrated intensity on specific code combinations. This distributed architecture indicates ChatGPT engages with diverse translation challenges simultaneously rather than prioritizing hierarchical coordination patterns.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eGemini 3.0 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, left) generated 53 coded units across 10 connections, creating the most sparse but focused network structure among the three models. Gemini\u0026rsquo;s strongest connection links Narrative Flow and Lexical Choice (weight: 0.3966), with Intertextual Reference receiving secondary attention (weight: 0.3589). This pattern reveals Gemini\u0026rsquo;s tendency to anchor translation decisions in narrative continuity, using this dimension as an organizational principle for coordinating other translation codes. The model\u0026rsquo;s relatively compact network with fewer but stronger connections implies a more selective, focused translation approach that concentrates epistemic effort on specific translation challenges rather than distributing attention broadly.\u003c/p\u003e \u003cp\u003eThe \u003cem\u003eJourney to the West\u003c/em\u003e (JTW) networks, shown in the right panels of Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, reveal both consistency and adaptation in the models\u0026rsquo; translation strategies when confronting a different classical text. Claude Opus 4.5 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, right) demonstrates the most dramatic scaling behaviour, expanding from 96 units in CG to 167 units, a 74% increase that makes it the most prolific model in the JTW corpus. Despite this expansion, the model maintains 21 connections with a mean strength of 0.1451, slightly lower than its CG mean. The network reveals strong coordination between Narrative Flow and Intertextual Reference (weight: 0.2391) and between Lexical Choice and Intertextual Reference (weight: 0.2133), demonstrating consistent attention to cultural context. Cultural Adaptation assumes even greater prominence in JTW (weight: 0.2922) compared to CG, possibly reflecting the text\u0026rsquo;s more fantastical elements requiring enhanced domestication strategies. This adaptive scaling suggests Opus shows greater variation in network size and connection distribution across the two texts, engaging more extensively with longer, more complex narratives while maintaining reasonable coordination intensity.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eChatGPT 5.2 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, right) expands to 129 units with 21 connections but shows a decreased mean strength of 0.1196, indicating the model processes substantially more material with less intensive code coordination per unit. Significantly, ChatGPT\u0026rsquo;s strongest connection in JTW shifts to Cultural Adaptation-Lexical Choice (weight: 0.2803), diverging from the Narrative Flow emphasis observed in CG. This shift suggests adaptive capacity based on text type, demonstrating responsiveness to different generic characteristics between the historical-mythological CG and the picaresque adventure narrative of JTW. However, the lower connection weights in JTW may indicate challenges in maintaining coherent strategy application across the episodic structure characteristic of this narrative form.\u003c/p\u003e \u003cp\u003eGemini 3.0 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, right) demonstrates remarkable consistency with 47 coded units forming 14 connections at a mean strength of 0.1596, maintaining its relatively sparse but focused network pattern observed in CG. The model\u0026rsquo;s primary connection strength lies between Narrative Flow and Lexical Choice (weight: 0.4144), with Intertextual Reference maintaining equal weight (0.4144) and secondary emphasis on Syntactic Structure and Lexical Choice (weight: 0.2464). This striking structural stability across both texts, 53 units and 10 connections in CG versus 47 units and 14 connections in JTW, suggests Gemini exhibits a relatively stable and narrowly distributed network structure across both texts.\u003c/p\u003e \u003cp\u003eThe comparative analysis across both corpora reveals that despite these differences, all three models share the centrality of Narrative Flow connections, indicating that maintaining story coherence remains paramount regardless of the AI system or source text. However, their divergent approaches, Gemini\u0026rsquo;s systematic consistency, ChatGPT\u0026rsquo;s exploratory breadth, and Opus\u0026rsquo;s adaptive scaling, suggest different operational philosophies in navigating the multidimensional challenges of classical Chinese literary translation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Epistemic Networks between AI and Human Translations\u003c/h2\u003e \u003cp\u003eThe epistemic network analysis reveals differences in how human translators and LLMs construct meaning when rendering classical Chinese texts into English. The comparison between Gu Zhizhong\u0026rsquo;s translation of CG and W.J.F. Jenner\u0026rsquo;s translation of JTW with the three LLMs demonstrates qualitatively distinct epistemic architectures that persist across both literary corpora.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFor \u003cem\u003eCreation of the Gods\u003c/em\u003e, the human translator Gu Zhizhong (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, left) demonstrates a distinctive epistemic architecture centred on Semantic Fidelity. Strong connections radiate from Semantic Fidelity to multiple translation codes, particularly Lexical Choice, Syntactic Structure, and Intertextual Reference. This pattern indicates that Gu consistently anchors decision-making in preserving the semantic integrity of the source text while simultaneously coordinating multiple other translation dimensions, creating what might be termed a \u0026ldquo;semantic-first\u0026rdquo; translation architecture.\u003c/p\u003e \u003cp\u003eClaude Opus 4.5 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, left) positions Narrative Flow as the primary hub from which other connections emanate, diverging from Gu\u0026rsquo;s semantic-first approach. With 96 coded units across 15 connections and a mean strength of 0.1796, Opus\u0026rsquo;s strongest connection links Narrative Flow and Lexical Choice (weight: 0.2847), with Narrative Flow and Intertextual Reference receiving equal attention (weight: 0.2753). While Opus displays engagement with Cultural Adaptation through connections to both Lexical Choice (weight: 0.2478) and Intertextual Reference (weight: 0.2478), it lacks the semantic anchoring that characterizes Gu\u0026rsquo;s network, prioritizing story coherence over precise meaning preservation.\u003c/p\u003e \u003cp\u003eChatGPT 5.2 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, left) exhibits a more distributed network with 64 units across 21 connections and a mean strength of 0.1452, contrasting sharply with Gu\u0026rsquo;s concentrated semantic-centred structure. ChatGPT\u0026rsquo;s strongest connection links Narrative Flow and Lexical Choice (weight: 0.3331), following a hierarchical pattern similar to Opus but with broader, less intense code co-activations throughout. This distributed architecture reveals an exploratory strategy that considers multiple translation dimensions without Gu\u0026rsquo;s clear hierarchical prioritization around meaning preservation, suggesting a different organizational principle in translation.\u003c/p\u003e \u003cp\u003eGemini 3.0 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, left) generated 53 coded units across 10 connections, creating the sparsest yet focused network. Gemini\u0026rsquo;s strongest connection links Narrative Flow and Lexical Choice (weight: 0.3966), with Intertextual Reference receiving secondary attention (weight: 0.3589). This configuration reveals systematic prioritization of narrative continuity, diverging from Gu\u0026rsquo;s semantic-anchored network structure. Where Gu integrates multiple codes through semantic fidelity as the organizing hub, Gemini employs a selective approach that concentrates on narrative coherence and literary allusion management, treating semantic correspondence as secondary.\u003c/p\u003e \u003cp\u003eFor \u003cem\u003eJourney to the West\u003c/em\u003e, W.J.F. Jenner\u0026rsquo;s translation (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, right) maintains the characteristic Semantic Fidelity centrality observed in Gu\u0026rsquo;s CG translation. Robust connections from Semantic Fidelity to Lexical Choice, Syntactic Structure, and Intertextual Reference form a coherent cluster, demonstrating that human translators coordinate these codes simultaneously while anchoring decisions in meaning preservation regardless of source text characteristics. This consistency suggests a stable human epistemic approach to classical Chinese translation.\u003c/p\u003e \u003cp\u003eClaude Opus 4.5 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, right) demonstrates dramatic scaling in JTW, expanding to 167 units with 21 connections while maintaining its narrative-centred orientation distinct from Jenner\u0026rsquo;s approach. The network reveals strong coordination between Narrative Flow and Intertextual Reference (weight: 0.2391) and between Lexical Choice and Intertextual Reference (weight: 0.2133). Cultural Adaptation assumes greater prominence (weight: 0.2922) compared to CG, possibly reflecting JTW\u0026rsquo;s fantastical elements. Despite this adaptive scaling, Opus continues organizing translation patterns around narrative coherence rather than adopting Jenner\u0026rsquo;s semantic-first architecture, revealing different priorities between human and artificial intelligence in managing literary translation.\u003c/p\u003e \u003cp\u003eChatGPT 5.2 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, right) expands to 129 units with 21 connections but shows decreased mean strength of 0.1196, contrasting with Jenner\u0026rsquo;s more concentrated code coordination. ChatGPT\u0026rsquo;s strongest connection shifts to Cultural Adaptation-Lexical Choice (weight: 0.2803), diverging from both its own CG pattern and Jenner\u0026rsquo;s semantic emphasis. This shift suggests some adaptive capacity based on text type, yet the model\u0026rsquo;s prioritization of cultural accessibility over meaning preservation represents a qualitative departure from Jenner\u0026rsquo;s translation philosophy, where semantic fidelity remains the gravitational center regardless of the text\u0026rsquo;s fantastical or cultural content.\u003c/p\u003e \u003cp\u003eGemini 3.0 Pro (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, right) maintains remarkable consistency with 47 units and 14 connections, with strongest connections continuing to link Narrative Flow with Lexical Choice (weight: 0.4144) and Intertextual Reference (weight: 0.4144). This structural stability across both texts contrasts sharply with Jenner\u0026rsquo;s flexible yet consistently semantic-centred approach. Where Jenner adapts specific translation choices while maintaining semantic fidelity as the organizing principle, Gemini applies a systematic algorithm that prioritizes narrative coherence regardless of source material characteristics, suggesting inflexibility in its translation philosophy.\u003c/p\u003e \u003cp\u003eThe comparative analysis across both corpora illuminates an architectural divergence: human translators construct epistemic networks where semantic preservation serves as the gravitational centre coordinating all other translation decisions, while LLMs organize their translation around narrative coherence, treating semantic fidelity as peripheral. This pattern persists across both classical texts and all three LLMs, suggesting these network differences reflect distinctions in human versus artificial intelligence approaches to cross-linguistic meaning-making.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Epistemic Networks between AI and Reference Corpus\u003c/h2\u003e \u003cp\u003eThe epistemic network analysis reveals structural differences between human and AI translation processes across both CG and JTW. Understanding these differences requires first examining the reference corpus network architecture, which provides a structural point of reference for contextualizing the AI network architectures examined below, not as a directly equivalent baseline, but as an illustration of the kind of integrated epistemic organization characteristic of proficient original writing in the target language.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe reference corpus demonstrates a densely interconnected epistemic network (see Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e) with mean connection strength of 0.1147 across 11,955 coded segments. Authors of these original writings establish robust co-occurrences among all code pairs, with the five strongest connections being Syntactic Structure with Cultural Adaptation (0.2751), Syntactic Structure with Intertextual Reference (0.2677), Lexical Choice with Cultural Adaptation (0.2436), Cultural Adaptation with Intertextual Reference (0.2068), and Syntactic Structure with Semantic Fidelity (0.1989). This connectivity pattern indicates that human writers engage in integrated decision-making where choices in one dimension systematically influence and coordinate with others, creating a holistic writing process. The network positioning shows balanced distribution across both structural-cultural dimensions (SVD1) and narrative-semantic dimensions (SVD2), enabling flexible responses to varied textual demands. Critically, the network reveals no isolated nodes or peripheral codes; instead, all writing considerations maintain substantial mutual engagement, suggesting that expert writers simultaneously manage multiple constraints rather than processing them sequentially or independently.\u003c/p\u003e \u003cp\u003eWhen examined alongside the reference corpus, Opus 4.5 exhibits markedly different network characteristics for CG translation. With mean connection strength of 0.1451 and maximum strength of 0.3474, the model shows higher mean density than some AI comparators in this study. Its strongest connections centre on Cultural Adaptation, linking it robustly to Lexical Choice (0.2922) and Intertextual Reference (0.2922). However, structural gaps emerge: Syntactic Structure connections remain weaker, with the Syntactic Structure to Intertextual Reference link reaching 0.2079, compared with 0.2677 in the reference corpus. This indicates reduced capacity for coordinating structural manipulation with allusion handling, essential for rendering the mythologically-embedded and allusion-dense passages characteristic of \u003cem\u003eCreation of the Gods\u003c/em\u003e. For JTW translation, Opus 4.5 shifts emphasis toward Narrative Flow connections, achieving mean strength of 0.1811 with maximum of 0.3880. The strongest connection becomes Narrative Flow to Intertextual Reference (0.3272), suggesting recognition of JTW\u0026rsquo;s episodic-allusive structure. Yet the Narrative Flow to Cultural Adaptation link appears weaker than in the reference corpus, potentially limiting coordinated cultural transfer within narrative progression.\u003c/p\u003e \u003cp\u003eGemini 3 Pro exhibits a more fragmented network architecture across both texts. For CG, the model achieves mean connection strength of 0.1596, but maximum strength of 0.4743 concentrated narrowly in Narrative Flow to Intertextual Reference. While establishing moderately strong links between Narrative Flow and Lexical Choice (0.4144), the model fails to connect Cultural Adaptation with Syntactic Structure in meaningful ways. This fragmentation creates more isolated processing modules rather than the integrated architecture visible in the reference corpus, with the network centroid positioned peripherally in epistemic space. For JTW, Gemini improves substantially with mean strength of 0.1765 and maximum of 0.4431, establishing strong connections from Narrative Flow to Intertextual Reference (0.3589) and moderate links to Cultural Adaptation (0.1474). However, structural weaknesses persist: Syntactic Structure remains poorly integrated, and high connection standard deviation (0.1600) indicates inconsistent strategy application across different passages, contrasting with the more evenly distributed connectivity of the reference corpus.\u003c/p\u003e \u003cp\u003eChatGPT 5.2 Pro occupies an intermediate structural position between Opus and Gemini. For CG translation, the model achieves mean strength of 0.1196 with maximum of 0.3785, prioritizing Cultural Adaptation connections at 0.2803 while maintaining moderate links to Intertextual Reference (0.2252) and Lexical Choice (0.2100). Unlike the balanced topology found in the reference corpus, ChatGPT 5.2 Pro exhibits Cultural Adaptation as an over-weighted hub while peripheralizing structural considerations, explaining its tendency toward surface-level cultural substitution without corresponding structural adaptation. The Syntactic Structure to Intertextual Reference connection falls significantly below reference corpus figures, indicating fragmented rather than integrated processing. For JTW, ChatGPT 5.2 Pro shows better balance than on CG through Narrative Flow emphasis (0.3331) and dual Cultural Adaptation connections to both Intertextual Reference and Lexical Choice (0.2100 each), with mean strength of 0.1452 and maximum of 0.3578. Nevertheless, connections remain structurally divergent from reference corpus figures by 15\u0026ndash;40% across critical code pairs.\u003c/p\u003e \u003cp\u003eAcross all comparisons, a consistent structural pattern emerges: LLMs demonstrate lower overall network density, unbalanced topologies with over-reliance on specific code pairs, and peripheral rather than central positioning in epistemic space. Where the reference corpus shows mean connection strengths above 0.11 with standard deviations around 0.10, LLMs exhibit either comparable means with higher variance (indicating inconsistency) or lower means with concentrated maximums (indicating fragmentation). These structural differences suggest that AI translation organizes meaning construction in more atomistic, hub-dependent ways, producing renderings that may lack the dimensional coordination visible in the reference corpus network. The balanced connectivity across all dimensions in the reference corpus contrasts with the AI tendency toward hub-based architectures, where certain codes dominate while others remain underutilized, a pattern that may bear on translation quality in ways quantitative metrics alone cannot capture, though the reference corpus\u0026rsquo;s different functional context means this comparison illuminates structural tendencies rather than establishes a performance standard.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Discussion","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Between Anthropogenic and Algorogenic Thumbprints\u003c/h2\u003e \u003cp\u003eThe epistemic network analysis reveals a complex relationship between AI and human translator style that resists simple categorization. Through the distinction between anthropogenic and algorogenic thumbprints (see Section 2.1), both human translators and LLMs can be understood as exhibiting identifiable recurrent patterns in Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) sense. At this level, the findings indicate that translator-like stylistic regularity is not exclusive to human agents.\u003c/p\u003e \u003cp\u003eBoth human and AI translators in this study satisfy Baker\u0026rsquo;s criterion of translator style as a \u0026ldquo;characteristic use of language\u0026rdquo; that persists across texts. Gu Zhizhong and W.J.F. Jenner display stable tendencies across \u003cem\u003eCreation of the Gods\u003c/em\u003e and \u003cem\u003eJourney to the West\u003c/em\u003e, while Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro likewise maintain relatively distinctive network configurations across both novels. This suggests that LLMs can produce recognizable and reproducible stylistic signatures, rather than merely generating isolated local solutions. In this respect, the findings are consistent with Saldanha\u0026rsquo;s (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) observation that translator style emerges through recurring patterns of choice.\u003c/p\u003e \u003cp\u003eThe more revealing difference lies in how these recurrent patterns are organized. Human translators in this study consistently structure their epistemic networks around Semantic Fidelity as the central coordinating node, with strong connections extending to Lexical Choice, Syntactic Structure, and Intertextual Reference (see Section 4). This semantic-first organization suggests that meaning preservation functions as the principle through which other translation decisions are negotiated. By contrast, all three LLMs show stronger centrality around Narrative Flow, while Semantic Fidelity occupies a less dominant position. Because this pattern recurs across both source texts and across all three models, it appears to indicate a broader tendency in LLM literary translation rather than a text-specific effect. The contrast therefore concerns not simply different stylistic preferences, but different ways of organizing translational attention.\u003c/p\u003e \u003cp\u003eComparison with the Western fantasy reference corpus sharpens this interpretation. The reference corpus exhibits denser and more balanced epistemic networks, suggesting a relatively integrated coordination of form, meaning, cultural framing, and stylistic convention. The LLM networks, by contrast, show lower density and greater dependence on a narrower range of code pairings. Particularly revealing is the weak coordination between Syntactic Structure and Cultural Adaptation in the AI networks, a relation that is much stronger in the reference corpus. This suggests that although LLM outputs may be fluent and narratively coherent, they are less effective at integrating multiple translational dimensions into a unified stylistic structure. The comparison should not be read as a direct quality ranking, since the reference corpus consists of original writing rather than translation, but it does help contextualize the structural tendencies observed in the LLM outputs.\u003c/p\u003e \u003cp\u003eTaken together, these findings suggest that human and AI translators may display comparable stylistic regularities while organizing translational choices differently. The distinction between anthropogenic and algorogenic thumbprints is useful here because it captures this contrast at the level of pattern formation rather than evaluative judgment. The findings therefore support a reading of translator style (thumbprint) in which recurrent textual signatures can emerge from more than one kind of underlying process, even if the resulting forms of stylistic organization are not equivalent.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Reconceptualizing the \u0026ldquo;Thumbprints\u0026rdquo; in AI translation\u003c/h2\u003e \u003cp\u003eThe distinction between anthropogenic and algorogenic thumbprints has important implications for Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) \u0026ldquo;thumbprints\u0026rdquo; framework. If comparable stylistic regularities can emerge from different generative processes, then translator style cannot be understood only in relation to human agency, intentionality, and socio-cultural situatedness. Baker\u0026rsquo;s framework remains valuable in identifying translation as a site of recurrent stylistic patterning, but its explanatory assumptions appear less adequate when extended to LLM translation.\u003c/p\u003e \u003cp\u003eIn this context, the distinction between anthropogenic and algorogenic thumbprints provides a more flexible theoretical vocabulary. It allows translator style to be treated as an observable pattern of translational organization without presupposing a single source for that pattern. In human translation, such regularities may plausibly be related to biography, training, interpretation, and socio-cultural positioning. In LLM translation, comparable regularities are more plausibly associated with model architecture, training distributions, and optimization dynamics. The point is not to deny the existence of regularity in either case, but to avoid treating formally similar outcomes as evidence of the same underlying process.\u003c/p\u003e \u003cp\u003eFrom this perspective, Baker\u0026rsquo;s framework needs not be rejected, but reformulated. Its central insight, that translation leaves recurrent stylistic traces, remains productive. What requires revision is the assumption that such traces necessarily derive from human subjectivity. The concept of algorogenic thumbprints may therefore serve as a provisional analytical category for describing translator-like regularities produced by non-human systems, without attributing to LLMs the forms of consciousness or agency associated with human translators. In this sense, the term is intended as a heuristic, rather than a definitive theoretical claim, and it remains open to revision as future research examines other model families, language pairs, literary genres, and prompting conditions.\u003c/p\u003e \u003cp\u003eThis reconceptualization also has methodological implications for translator style research. Traditional approaches have often treated stylistic patterning as evidence of human interpretation and intentionality, making agency both the object and, implicitly, the explanation of analysis. Such assumptions are less applicable in the case of LLM translation, where intentionality is not in itself an adequate explanatory framework. Variables such as training data distributions, architectural constraints, decoding settings, and prompting conditions may instead be more relevant. In this regard, epistemic network analysis is valuable because it models how different translator types organize translational choices relationally without presupposing identical cognitive mechanisms. It thus provides a useful framework for comparing human and algorithmic translators while remaining sensitive to the distinction between anthropogenic and algorogenic forms of stylistic patterning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Architecturally Aware Human\u0026ndash;AI Collaboration in Literary Translation\u003c/h2\u003e \u003cp\u003eThe epistemic network analysis reveals preliminary implications for how literary translation practice might be organized in an era of increasingly capable AI systems. The structural contrasts identified here, particularly the complementary strengths and limitations of semantic-anchored human networks and narrative-prioritized AI architectures, suggest a theoretical basis for exploring collaborative models. The workflow proposals that follow should, however, be understood as theoretically motivated hypotheses requiring empirical validation rather than evidence-based prescriptions ready for immediate adoption.\u003c/p\u003e \u003cp\u003eThe network comparisons suggest that LLMs demonstrate comparative strengths in dimensions where systematic pattern application proves advantageous. In this study, Claude Opus 4.5\u0026rsquo;s consistent narrative-centred architecture and strong Narrative Flow connections (weights: 0.2847\u0026ndash;0.3272) indicate reliable capacity for maintaining story coherence across extended passages. Similarly, all three models demonstrate robust Lexical Choice coordination, suggesting potential effectiveness in terminology consistency management. These observations align with Kenny\u0026rsquo;s (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) identification of AI\u0026rsquo;s advantage in applying linguistic patterns systematically at scale, a capacity potentially valuable for managing the elaborate mythological systems and supernatural terminology that classical Chinese fantasy translation demands, though direct empirical confirmation within this specific literary domain awaits further investigation.\u003c/p\u003e \u003cp\u003eThe networks simultaneously expose limitations that resonate with findings in existing translation technology research. The systematic absence of Semantic Fidelity as a central organizing principle across all three LLMs represents a weakness in literary translation, where meaning preservation constitutes what Landers (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2001\u003c/span\u003e) terms the translator\u0026rsquo;s primary obligation. Human translators\u0026rsquo; semantic-anchored networks demonstrate integrated decision-making that the models examined here do not replicate. This architectural gap finds corroboration in Li\u0026rsquo;s (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) empirical study of AI literary translation of classical Chinese poetry, which found that AI systems face significant limitations in cultural conveyance and translator subjectivity, with human intervention remaining indispensable for preserving the authenticity and depth of literary renderings. More broadly, post-editing research demonstrates that AI-generated literary output consistently requires substantial human revision, particularly where cultural density and syntactic complexity coincide, precisely the conditions characteristic of classical Chinese narrative texts (Jiang et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese patterns suggest a possible epistemic division of labour worth testing in future studies. In principle, the narrative-centred processing of LLMs could inform initial draft generation, maintaining story coherence and lexical consistency across extended passages, while human translators apply semantic-anchored revision, strengthening meaning preservation, cultural adaptation, and the coordinated decision-making that AI networks systematically underutilize. However, it is worth stressing that this specific task allocation has not been empirically tested in classical Chinese literary translation. Whether the quality improvements documented in adjacent post-editing domains (Jiang et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) extend to the demands of classical literary texts requires controlled experimental investigation before any practical implementation can be responsibly recommended.\u003c/p\u003e \u003cp\u003eRealizing such collaboration would also require addressing challenges the analysis brings into relief. The architectural divergence between semantic-anchored human networks and narrative-prioritized AI networks may create integration difficulties: AI-generated passages organized around narrative coherence could resist revision aimed at reorienting them around semantic fidelity without near-complete retranslation. As Lee (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) cautions from a posthumanist perspective, over-reliance on AI augmentation risks eroding translators\u0026rsquo; epistemic engagement with source texts, potentially weakening the very semantic-anchored architecture that distinguishes expert human translation. These challenges reinforce that the collaborative framework sketched here is a research agenda, not a deployable workflow model.\u003c/p\u003e \u003cp\u003eThe future of literary translation may therefore lie in developing what might be conceptualized as \u0026ldquo;architecturally aware collaboration\u0026rdquo;, an approach that recognizes human and AI systems as constructing meaning through different epistemic structures, each offering distinct capabilities for mediating classical Chinese literature across cultural boundaries. This study provides a conceptual framework and preliminary evidence for such an approach; the empirical validation of the collaborative model itself constitutes the necessary next step for translating these theoretical insights into practice.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study set out to examine whether LLMs develop consistent stylistic signatures when translating classical Chinese fantasy, and whether such signatures are structurally comparable to those of human translators. The ENA findings confirm that they are not, but the more consequential contribution lies in specifying \u003cem\u003ehow\u003c/em\u003e they differ and what that difference demands theoretically.\u003c/p\u003e \u003cp\u003eThe three research questions yield a coherent interpretive picture: LLMs exhibit distinct, cross-textually stable epistemic architectures organized around narrative coherence; human translators systematically anchor meaning-making in semantic fidelity; and both diverge structurally from the balanced, densely interconnected networks characteristic of original English fantasy writing. Taken together, these findings do not simply indicate that AI translates differently, they reveal that the epistemic network structures observed in this study differ systematically between the LLM and human translations. This distinction carries a consequence Baker\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2000\u003c/span\u003e) original framework was not designed to accommodate: that stylistic consistency can be generated without consciousness, intentionality, or socio-cultural situatedness. The anthropogenic/algorogenic distinction proposed here is therefore not a terminological refinement but a conceptual necessity, one that repositions translator style theory from a human-centred account toward a broader framework capable of modelling meaning-making across hybrid human-AI translation ecologies increasingly characteristic of contemporary literary practice.\u003c/p\u003e \u003cp\u003eFuture work should empirically test whether algorogenic architectures are prompt-sensitive, examine whether the semantic/narrative divergence identified here persists across non-English-centric language pairs, and investigate whether architecturally aware human-AI collaboration can produce networks that neither humans nor models achieve independently.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAuthor contributions: Conceptualisation: Kan Wu, Siqi Jiang \u0026amp; Defeng Li.Supervision: Defeng Li. Data collection: Kan Wu \u0026amp; Siqi Jiang. Methodology: Kan Wu. Data analysis and interpretation: Kan Wu. Writing\u0026mdash;original draft: Kan Wu \u0026amp; Siqi Jiang. Writing\u0026mdash;revising and proofreading: Kan Wu \u0026amp; Siqi Jiang.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThis study is supported by the Zhejiang Provincial Philosophy and Social Science Planning Annual Project (26NDJC277YB).\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eDataset related to this study is available at: 10.6084/m9.figshare.31915422\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBaker M (2000) Towards a methodology for investigating the style of a literary translator. Target 12(2):241\u0026ndash;266\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaker M (2018) In other words: a coursebook on translation. Routledge, London\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoase-Beier J (2006) Stylistic approaches to translation. Routledge, London\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBosseaux C (2004) Point of view in translation: a corpus-based study of French translations of Virginia Woolf\u0026rsquo;s To the Lighthouse. Lang Cult 5(1):107\u0026ndash;122\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBosseaux C (2007) How does it feel? Point of view in translation: the case of Virginia Woolf into French. Rodopi, Amsterdam\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCacioppo S, Cacioppo JT (2012) Decoding the invisible forces of social connections. Front Integr Neurosci 6:51\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChesterman A (2016) Memes of translation: the spread of ideas in translation theory, revised edn. John Benjamins, Amsterdam\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCsanadi A, Eagan B, Kollar I, Shaffer DW, Fischer F (2018) When coding-and-counting is not enough: using epistemic network analysis (ENA) to analyze verbal data in CSCL research. Int J Comput-Support Collab Learn 13(4):419\u0026ndash;438\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElkins K (2024) In search of a translator: using AI to evaluate what\u0026rsquo;s lost in translation. Front Comput Sci 6:1444021\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElmoazen R, Saqr M, Tedre M, Hirsto L (2022) A systematic literature review of empirical research on epistemic network analysis in education. IEEE Access 10:17330\u0026ndash;17348\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGao R, Lin Y, Zhao N, Cai ZG (2024) Machine translation of Chinese classical poetry: a comparison among ChatGPT, Google Translate, and DeepL Translator. Humanit Soc Sci Commun 11(1):1\u0026ndash;10\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan C, Lu X (2025) Beyond BLEU: repurposing neural-based metrics to assess interlingual interpreting in tertiary-level language learning settings. Res Methods Appl Linguist 4(1):100184\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHouse J (2015) Translation quality assessment: past and present. Routledge, London\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHsia CT (2016) The classic Chinese novel: a critical introduction. Chinese University of Hong Kong, Hong Kong\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang L (2015) Style in translation: a corpus-based perspective. Springer, Berlin\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJakobsen AL, Jensen KTH (2008) Eye movement behaviour across four different types of reading task. Cph Stud Lang 36:103\u0026ndash;124\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang L, Wei B, Al-Shaibani GKS (2025) Effective neural machine translation with human post-editing of Chinese intangible cultural heritage corpus into English. SAGE Open 15(4):21582440251386954\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKarpinska M, Iyyer M (2023) Large language models effectively leverage document-level context for literary translation, but critical errors persist. arXiv:2304.03245\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKenny D (2022) Machine translation for everyone: empowering users in the age of artificial intelligence. Language Science, Berlin\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLanders CE (2001) Literary translation: a practical guide. Multilingual Matters, Clevedon\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL\u0026auml;ubli S, Sennrich R, Volk M (2018) Has machine translation achieved human parity? a case for document-level evaluation. arXiv:1808.07048\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee TK (2024) Artificial intelligence and posthumanist translation: ChatGPT versus the translator. Appl Linguist Rev 15(6):2351\u0026ndash;2372\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi D (2016) Translator style: a corpus-assisted approach. In: Ji M, Oakes M, Li D, Hareide L (eds) Corpus methodologies explained: an empirical approach to translation studies. Routledge, London, pp 103\u0026ndash;136\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Q (2024) Bridging languages: the potential and limitations of AI in literary translation\u0026mdash;a case study of the English translation of \u0026lsquo;A Pair of Peacocks Southeast Fly\u0026rsquo;. Adv Humanit Res 8(1):1\u0026ndash;7\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMalmkj\u0026aelig;r K (2003) What happened to God and the angels: an exercise in translational stylistics. Target 15(1):37\u0026ndash;58\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMastropierro L (2018) Key clusters as indicators of translator style. Target 30(2):240\u0026ndash;259\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMikros G (2025) Beyond the surface: stylometric analysis of GPT-4o\u0026rsquo;s capacity for literary style imitation. Digit Scholarsh Humanit 40(2):587\u0026ndash;600\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOlohan M (2003) How frequent are the contractions? a study of contracted forms in the translational English corpus. Target 15(1):59\u0026ndash;89\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eO\u0026rsquo;Sullivan J (2025) Stylometric comparisons of human versus AI-generated creative writing. Humanit Soc Sci Commun 12:1708\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePing Y, Wang B (2024) Retranslated Chinese classical canon \u003cem\u003eJourney to the West\u003c/em\u003e: a stylometric comparison between Julia Lovell\u0026rsquo;s retranslation and Arthur Waley\u0026rsquo;s translation. Digit Scholarsh Humanit 39(1):308\u0026ndash;320\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePym A (2015) Translating as risk management. J Pragmat 85:67\u0026ndash;80\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuis AR, Rosser AA, Quandt-Walle C, Nathwani JN, Shaffer DW, Pugh CM (2018) The hands and head of a surgeon: modeling operative competency with multimodal epistemic network analysis. Am J Surg 216(5):835\u0026ndash;840\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaldanha G (2011) Translator style: methodological considerations. Translator 17(1):25\u0026ndash;50\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShaffer D (2018) Transforming big data into meaningful insights: introducing quantitative ethnography. Scientia, Madison\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShaffer DW, Collier W, Ruis AR (2016) A tutorial on epistemic network analysis: analyzing the structure of connections in epistemic, social, and interaction data. J Learn Anal 3(3):9\u0026ndash;45\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShaffer DW, Hatfield D, Svarovsky GN, Nash P, Nulty A, Bagley E et al (2009) Epistemic network analysis: a prototype for 21st-century assessment of learning. Int J Learn Media 1(2):33\u0026ndash;53\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiebert-Evenstone AL, Irgens GA, Collier W, Swiecki Z, Ruis AR, Shaffer DW (2017) In search of conversational grain size: modeling semantic structure using moving stanza windows. J Learn Anal 4(3):123\u0026ndash;139\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVandepitte S, Hartsuiker RJ (2011) Metonymic language use as a student translation problem: towards a controlled psycholinguistic investigation. In: Alvstad C, Hild A, Tiselius E (eds) Methods and strategies of process research. John Benjamins, Amsterdam, pp 67\u0026ndash;92\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Q, Li D (2011) Looking for translator\u0026rsquo;s fingerprints: a corpus-based study on Chinese translations of \u003cem\u003eUlysses\u003c/em\u003e. Lit Linguist Comput 27(1):81\u0026ndash;93\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWinters M (2004) German translations of F. Scott Fitzgerald\u0026rsquo;s The Beautiful and Damned: a corpus-based study of modal particles as features of translators\u0026rsquo; style. In: Kemble I (ed) Using corpora and databases in translation. University of Portsmouth, Portsmouth, pp 71\u0026ndash;88\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu K, Li D (2024) Unraveling Eileen Chang\u0026rsquo;s stylistic multiverse: insights from multivariate analysis with multifactorial design. Digit Scholarsh Humanit 39(3):1001\u0026ndash;1018\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYao X, Kang Y-B, McCosker A (2025) Missing the human touch? a computational stylometric analysis of GPT-4 translations of online Chinese literature. Transl Spaces 14(2):303\u0026ndash;330\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang R, Zhao W, Eger S (2025) How good are LLMs for literary translation, really? literary translation evaluation with humans and LLMs. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies, vol 1. p 10961\u0026ndash;10988\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"humanities-and-social-sciences-communications","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"palcomms","sideBox":"Learn more about [Humanities \u0026 Social Sciences Communications](http://www.nature.com/palcomms/)","snPcode":"41599","submissionUrl":"https://submission.springernature.com/new-submission/41599/3","title":"Humanities and Social Sciences Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"algorogenic thumbprints, epistemic network analysis, large language models, translator style, classical Chinese fantasy","lastPublishedDoi":"10.21203/rs.3.rs-9323835/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9323835/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study investigates algorogenic thumbprints, defined as consistent and reproducible stylistic patterns generated by large language models (LLMs) in literary translation. Using epistemic network analysis (ENA) across seven decision-making domains, lexical choice, syntactic structure, cultural adaptation, stylistic register, semantic fidelity, narrative flow, and intertextual reference, we examine translations of two classical Chinese fantasy novels produced by Claude Opus 4.5, ChatGPT 5.2 Pro, and Gemini 3 Pro, alongside published human translations and a Western fantasy reference corpus. The findings show that each LLM exhibits a stable, model-specific epistemic architecture oriented toward narrative coherence, whereas human translators structure decisions more strongly around semantic fidelity. This architectural divergence persists across texts and models. The study therefore suggests that Baker\u0026rsquo;s framework of translator thumbprints can be meaningfully extended to algorithmic translation and demonstrates the usefulness of ENA for modeling the relational structures underlying translation decision-making.\u003c/p\u003e","manuscriptTitle":"Algorogenic Thumbprints in Literary Translation: An Epistemic Network Analysis of LLM Translations of Classical Chinese Fantasy","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-29 11:55:10","doi":"10.21203/rs.3.rs-9323835/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-05-11T06:08:07+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-10T16:26:07+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-04T18:50:53+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-27T14:27:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"329076672906919478216273970862331985659","date":"2026-04-27T13:58:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"1382847197406765086094668667788244169","date":"2026-04-26T14:15:36+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"323556983084106865499680293377441933286","date":"2026-04-22T11:42:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"55247735662025951949361922977115968856","date":"2026-04-21T04:26:05+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-21T03:54:46+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-21T03:52:56+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-17T06:09:07+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-17T04:09:43+00:00","index":"","fulltext":""},{"type":"submitted","content":"Humanities and Social Sciences Communications","date":"2026-04-17T04:04:16+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"humanities-and-social-sciences-communications","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"palcomms","sideBox":"Learn more about [Humanities \u0026 Social Sciences Communications](http://www.nature.com/palcomms/)","snPcode":"41599","submissionUrl":"https://submission.springernature.com/new-submission/41599/3","title":"Humanities and Social Sciences Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Nature AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"efc59fd3-3c5b-4a59-8f89-733ec06865cc","owner":[],"postedDate":"April 29th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-11T06:08:07+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-10T16:26:07+00:00","index":42,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-04T18:50:53+00:00","index":40,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":67099036,"name":"Humanities/Cultural and media studies"},{"id":67099037,"name":"Social science/Cultural and media studies"},{"id":67099038,"name":"Humanities/Language and linguistics"},{"id":67099039,"name":"Social science/Language and linguistics"},{"id":67099040,"name":"Humanities/Literature"}],"tags":[],"updatedAt":"2026-05-11T06:12:16+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-29 11:55:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9323835","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9323835","identity":"rs-9323835","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}