بلاغت دیجیتال و اخلاق سایبری در ایران معاصر: کاربست هوش‌مصنوعی در تحلیل انتقادی و بازنمایی متون ادبی و فرهنگی

This study delineates the framework of digital rhetoric and cyberethics within the contemporary Iranian literary context, focusing on interpretative and ethical challenges. The central problem of this research is the critical reading and genealogy of analytical rifts in applying artificial intelligence models to encode, represent, and evaluate the rhetorical devices of literary texts. Due to the metaphorical nature of the Persian language, Western computational tools encounter significant ethical and interpretative calibration challenges when confronting second-order tropes such as Iham (amphiboly/polysemy), resulting in contextual blindness and the materialistic reductionism of sacred meanings. This research aims to extract computational validation patterns in text-based literary analysis and propose a localized matrix for evaluating digital rhetoric. The methodology employed is digital qualitative content analysis utilizing a text-based, computational approach. This is executed through the calibration of the GPT-4 large language model and NVivo software to extract core and secondary codes from Ghazal 1 of Hafez's Divan, the visual poetry of Tahereh Saffarzadeh, and Sadeq Chubak's novel Tangsir. The rationale for utilizing this method lies in its capacity to synthesize the mathematical architecture of algorithms with the semantic nuances of literary criticism. The findings indicate that without reconfiguration based on traditional Iranian rhetoric, computational algorithms suffer a 32% interpretative error rate in decoding Iham and encounter a methodological impasse in analyzing visual poetry. Nevertheless, inter-rater reliability calculations between human evaluators and the machine demonstrate that by implementing localized prompts and cyberethical frameworks, Cohen's kappa coefficient (κ) scales up to 0.84, indicating an excellent level of agreement. These insights substantiate the imperative transition toward human-in-the-loop digital humanities and pave the way for targeted policymaking to mitigate algorithmic biases.
Introduction
The rapid integration of generative artificial intelligence (AI) within the global educational and cultural landscapes has prompted a fundamental reassessment of textuality, interpretive authority, and hermeneutic practices. While natural language processing (NLP) models demonstrate high proficiency in processing shallow structural semantics, their application to non-Western, highly metaphorical literary traditions reveals significant methodological and ethical rifts. This study addresses the critical intersection of digital rhetoric and cyberethics within the contemporary Iranian literary and cultural context.
The core research problem stems from the inherent "contextual blindness" of Western-centric computational tools when dealing with the high-context, multi-layered semantic architecture of the Persian language. When algorithmic models encounter second-order tropes such as Iham (amphiboly/polysemy), they frequently default to literalist decoding, executing a materialistic reductionism that strips classical and modern texts of their sacred, allegorical, or socio-political nuances. This dynamic constitutes not merely a technical limitation but a profound ethical dilemma regarding algorithmic representation, cultural hegemony, and intellectual dignity in the age of generative AI.
Consequently, this research aims to accomplish two objective milestones: first, to systematically trace and quantify the interpretive errors committed by advanced large language models (LLMs) facing classical and modern Persian literary corpora; second, to construct a localized validation matrix rooted in traditional Eastern rhetoric and modern cyberethics to restore interpretive sovereignty to human-in-the-loop digital humanities.
Materials and Methods
This study utilizes a digital qualitative content analysis design underpinned by computational poetics and text-based validation protocols. To guarantee empirical breadth across genres, structures, and historical epochs, a specialized multi-genre literary corpus was constructed, comprising three distinct pillars:

Classical Mystical Lyricism: Ghazal 1 from the Divan-e Hafez, selected for its hyper-dense layers of mystical and secular Iham.
Modern Avant-Garde/Visual Poetry: Selected typographic and ideographic concrete poems from Tahereh Saffarzadeh’s Tanin-e Dadeh-ha (The Resonance of Data), chosen to test structural-spatial processing.
Realistic Modern Fiction: Narrative segments from Sadeq Chubak’s novel Tangsir, focusing on colloquial vernacular, socio-cultural idioms, and dense cultural prose.

The technical methodology proceeded in two operational phases. In the first phase, the default, uncalibrated engine of OpenAI's GPT-4 large language model was prompted to execute inductive coding, thematic extraction, and rhetorical identification on the raw Persian texts, using basic zero-shot instructions. Simultaneously, the same corpus was ingested into NVivo software, where a panel of three expert human evaluators—specialists in classical and modern Persian rhetoric—independently generated an open-coding taxonomy based on traditional canons, including Asrar al-Balaghah and Al-Motavval.
In the second phase, a localized cyberethical framework and cross-cultural prompt engineering models were introduced to the LLM. The AI was re-tested using culturally calibrated system prompts that explicitly operationalized traditional rhetoric concepts. Statistical convergence and divergence between the machine-generated codes and the human consensus matrix were mathematically monitored using Cohen's kappa coefficient (κ) computed through NVivo to evaluate inter-rater reliability across both phases.
Research Findings
The comparative textual analysis yielded stark structural anomalies between default computational outputs and human expertise. In the raw, uncalibrated phase, computational algorithms demonstrated a notable vulnerability to semantic flatlining, exhibiting a 68% accuracy rate (a 32% interpretative error rate) in decoding Iham within Ghazal 1 of Hafez due to materialist reductionism and literalism. The model consistently over-selected the primary, material denotation of mystical tokens (e.g., translating spiritual wine and tavern spaces purely in a materialistic, hedonistic sense), failing to capture the dual-tier semantic systems outlined in classical rhetorical manuals.
When processing the visual poetry of Tahereh Saffarzadeh in Tanin-e Dadeh-ha, the model encountered a total methodological impasse, resulting in 0% accuracy due to the loss of spatial and typographic rhetoric. Lacking spatial, topographic, and gestalt awareness of the printed typographic layouts, GPT-4 extracted the semantic text as a linear sequence, thereby discarding the macrostructural rhetoric where the visual arrangement itself forms an ideological critique of technological alienation; however, when provided with explicit textual descriptions of the visual layouts, its analytical accuracy reached 45%. In Chubak's Tangsir, the model displayed high accuracy (81%) in superficial plot-point identification but achieved only a 52% accuracy rate in localized vernacular analysis, frequently triggering cultural decontextualization by mistaking regional Southern Iranian honor codes for literal declarations of lawlessness.
However, the introduction of localized cyberethical prompts and cross-cultural system guidelines transformed the analytical landscape. By feeding the model ontological constraints derived from Persian poetics, the interpretative accuracy for Hafez scaled up to 91%, while Chubak's vernacular comprehension increased to 86%. The statistical measurement of inter-rater reliability reflected this optimization: Cohen's kappa coefficient (κ) between the human consensus and the AI analysis improved from a baseline κ=0.41κ=0.41 to κ=0.84κ=0.84, indicating much stronger agreement after calibration. While this result suggests the system is highly adaptable to culturally informed tuning, it does not by itself demonstrate the presence or absence of Western bias.
Discussion of Results and conclusion
The empirical findings underscore that default computational algorithms are not value-neutral instruments; they carry embedded epistemic assumptions that prioritize Western rhetorical models and low-context textual linearism. As Moretti (2013, p. 55) argues within the ethos of distant reading, computational aggregation can unveil macro-patterns, yet this study proves that such distant readings trigger a total collapse of meaning when applied to hyper-dense Eastern literary traditions without regional re-indexing. The 32% error rate in Iham processing highlights an urgent ethical vulnerability: the systematic devaluation of non-Western rhetorical structures in global digital archives. If generative models become the primary arbiters for translating and representing Persian cultural heritage, a form of digital epistemicide occurs, where complex allegorical traditions are compressed into flat, uniform prose.
To mitigate this challenge, this study offers a pioneering, localized matrix for digital rhetoric in contemporary Iran. By demonstrating that a human-in-the-loop architecture can elevate analytical agreement to an excellent Cohen's kappa of 0.84, the research proves that the future of digital humanities lies not in automated surrender to AI, but in active cyberethical intervention.
Conclusively, this study recommends that Iranian academic institutions and cultural policymakers mandate ethical calibration protocols for any LLM-driven textual analysis platforms. Scholars must move beyond both the uncritical adoption of digital tools and the total rejection of computational methodologies. Instead, by combining traditional rhetorical heritage with rigorous prompt design, the Iranian academic community can harness artificial intelligence to preserve, unpack, and globally project the multi-layered majesty of its literary and cultural history within the digital ecosystem.