The Poet’s Query: Unmasking AI’s Hidden Voices

by

Author: Denis Avetisyan

Researchers are using poetic prompts to reveal the underlying biases and creative constraints of large language models.

This review examines how ‘Poetry Prompt Patterns’ can serve as a diagnostic tool for analyzing the rhetorical tendencies and algorithmic biases present in generative AI.

Despite advances in artificial intelligence, understanding the internal logic and biases of large language models remains a significant challenge. This paper, ‘Decoding the Black Box: Discerning AI Rhetorics About and Through Poetic Prompting’, investigates a novel approach to probing these ‘black boxes’ by employing creatively structured text prompts-specifically, ‘Poetry Prompt Patterns’-as a diagnostic tool. Our analysis reveals how these models not only describe poetic content but also demonstrate rhetorical tendencies and a willingness to adapt original creative works, potentially exposing underlying biases. Ultimately, can a poetic lens offer new insights into the ethical implications of generative AI and its capacity for creative appropriation?

The Echoing Mirror: LLMs and the Reproduction of Bias

Large Language Models, despite their sophisticated capabilities, aren’t neutral processors of information; rather, they function as mirrors reflecting the biases embedded within the vast datasets used for training. These models learn patterns and associations from text and code, and if those materials contain prejudiced viewpoints – regarding gender, race, religion, or any other social category – the LLM will inevitably absorb and reproduce them. This isn’t a matter of intentional malice on the part of the algorithm, but a consequence of statistical learning; the model predicts the most probable continuation of a given prompt, and if biased content is prevalent in the training data, biased outputs become statistically more likely. Consequently, LLMs can inadvertently perpetuate harmful stereotypes, reinforce discriminatory language, and even amplify existing social inequalities, making critical evaluation of their outputs essential.

The propagation of bias within Large Language Models isn’t a flaw in their architecture, but rather a direct consequence of the data used to cultivate them. These models learn by identifying patterns and relationships within vast datasets – datasets which, inevitably, reflect the historical and systemic inequalities present in the real world. Consequently, if the training data contains prejudiced language, underrepresentation of certain groups, or reinforces harmful stereotypes, the model will internalize and reproduce these biases in its outputs. This means that seemingly neutral applications of the technology can inadvertently perpetuate discriminatory practices, not through intentional programming, but through the uncritical absorption of existing societal prejudices embedded within the very foundations of its knowledge base.

The application of Large Language Models to creative fields such as poetry and literature demands careful scrutiny, as these systems don’t generate content from a vacuum of objectivity. Instead, they reproduce patterns learned from massive datasets, and if those datasets contain societal biases – regarding gender, race, or other characteristics – the resulting creative work will likely reflect and even amplify them. This isn’t simply a matter of factual inaccuracy; biased language models can perpetuate harmful stereotypes through character portrayals, narrative structures, and thematic emphasis, potentially influencing perceptions and reinforcing inequalities under the guise of artistic expression. Consequently, a thorough understanding of how these biases manifest in generated text is crucial for responsible development and deployment, prompting questions about authorship, originality, and the ethical implications of AI-driven creativity.

The construction of identity, both individual and collective, is fundamentally shaped by the data used to train large language models. These models don’t possess inherent understandings of people or groups; instead, they statistically assemble representations based on patterns within their training datasets. Consequently, if those datasets reflect historical biases – for example, associating certain professions more frequently with one gender or ethnicity – the model will reproduce and even amplify these skewed associations. This isn’t a matter of intentional prejudice, but a consequence of algorithmic pattern recognition; the LLM perceives and replicates the imbalances present in the data as indicative of real-world relationships. The resulting representations, therefore, can perpetuate harmful stereotypes and limit the diversity of perspectives, potentially impacting how individuals and groups are perceived and understood through the lens of artificial intelligence.

Deconstructing the Text: A Methodological Approach

Critical Discourse Analysis (CDA) and Rhetorical Analysis form the core of our methodology for identifying bias in Large Language Model (LLM) outputs. CDA examines the relationship between language and power, focusing on how LLMs construct and reinforce social inequalities through linguistic choices; this includes analyzing vocabulary, framing, and presuppositions. Rhetorical Analysis complements this by dissecting the persuasive strategies employed in the generated text, identifying appeals to ethos, pathos, and logos, and assessing their potential to manipulate or mislead. Both approaches move beyond surface-level content analysis to investigate the underlying ideological commitments and discursive practices embedded within the LLM’s textual production, enabling a systematic evaluation of potential biases.

Critical Discourse Analysis and Rhetorical Analysis, when applied to Large Language Model outputs, facilitate the identification of embedded ideological frameworks and persuasive strategies. This involves examining linguistic features – including framing, lexical choices, and narrative structures – to reveal how the text constructs particular viewpoints or advocates specific positions. By deconstructing the text’s argumentative components, researchers can determine the underlying assumptions guiding the LLM’s responses and identify the rhetorical devices employed to influence interpretation. This process moves beyond surface-level content analysis to reveal the deeper mechanisms through which LLMs shape meaning and potentially perpetuate biases.

Prompt Engineering, utilizing Poetry Prompt Patterns, establishes a reproducible framework for bias assessment in Large Language Models (LLMs). This methodology involves constructing prompts with specific poetic forms – such as haiku, sonnets, or villanelles – which, due to their inherent structural constraints, consistently elicit responses that reveal underlying algorithmic tendencies. By analyzing the LLM’s adherence to or deviation from the expected poetic form, and the semantic content generated within that structure, researchers can systematically identify and categorize biases related to gender, race, ideology, or other sensitive attributes. The controlled nature of these prompts allows for comparative analysis across different LLMs and model versions, quantifying the prevalence and character of these biases as diagnostic metrics for algorithmic behavior.

Traditional methods of analyzing language, such as keyword spotting or sentiment analysis, offer a surface-level understanding of text. This methodological approach, however, prioritizes the examination of linguistic structures and rhetorical devices to reveal how Large Language Models (LLMs) actively construct meaning. It moves beyond identifying what is said to analyzing the processes by which LLMs frame arguments, establish relationships between concepts, and represent entities and events. This includes analyzing choices in lexical selection, syntactic arrangement, and narrative construction to expose the underlying cognitive and ideological frameworks embedded within the generated text, thereby providing a more nuanced understanding of the LLM’s representational strategies.

Echoes of Convention: Examples and Analysis

Analysis of large language model outputs demonstrates a tendency to reinforce positive stereotypes, even when prompts are designed to elicit neutral or positive responses. This occurs because LLMs are trained on massive datasets that contain pre-existing societal biases, which are then reflected in generated text. Specifically, models may overemphasize perceived positive traits associated with particular demographic groups, or consistently portray individuals from those groups in limited, albeit flattering, roles. This perpetuation of stereotypes isn’t necessarily intentional, but arises from the statistical patterns learned during training; the model prioritizes generating text that aligns with frequently observed associations, even if those associations are based on biased or incomplete information. The effect is observable even when the model is explicitly instructed to create positive content, as the inherent biases influence the way positivity is expressed.

Analysis indicates that Large Language Models (LLMs) can exhibit patterns of appropriation when generating creative text. This manifests as the imitation of an author’s stylistic elements – including diction, syntax, and thematic concerns – without acknowledging the source or demonstrating comprehension of the original work’s cultural or historical context. Observed instances included replicating poetic forms or employing phrasing characteristic of specific poets without attribution, suggesting a surface-level understanding of style divorced from meaningful engagement with the author’s intent or background. This behavior is not necessarily intentional plagiarism, but rather a result of the models’ training on vast datasets where stylistic patterns are learned and reproduced without inherent understanding of ownership or influence.

Analysis of LLM-generated poetry revealed instances of factual inaccuracies and nonsensical statements integrated within poetic text. These “hallucinations” were not simply stylistic choices but demonstrable errors; for example, models generated biographical details about poets inconsistent with established records, or attributed nonexistent publications to them within the poems themselves. Furthermore, some generated verses contained logically inconsistent imagery or scenarios that lacked internal coherence, indicating a failure to maintain semantic consistency even within the bounds of creative license. These instances suggest LLMs can generate text that appears meaningful but lacks grounding in factual accuracy or logical reasoning, even when operating within a traditionally less-constrained format like poetry.

Testing of large language models ChatGPT, Claude, and DeepSeek with poetic prompts revealed consistent positive bias in their responses concerning Maya Angelou and her poetry. Across multiple prompts and variations, all three models overwhelmingly generated praising language when describing Angelou’s work, often employing superlative adjectives and emphasizing the emotional impact and cultural significance of her poems. This consistent positive framing occurred regardless of the specific prompt’s intent, even when requesting objective analysis or critical evaluation. The models did not demonstrate this level of consistent praise when prompted about other poets, indicating a specific and potentially learned bias towards Angelou and her established reputation.

Comparative analysis of large language model responses to poetic adaptation prompts revealed differing approaches to ethical considerations. Claude consistently refused requests to rewrite Maya Angelou’s “Still I Rise,” citing concerns about altering a work deeply embedded in specific cultural experience and potentially misrepresenting its original intent. In contrast, ChatGPT readily adapted all three poems provided – “Still I Rise,” “I Know Why the Caged Bird Sings,” and “Phenomenal Woman” – without expressing similar reservations, demonstrating a divergence in how these models address the ethical implications of modifying culturally significant creative works.

Analysis of poetic adaptation attempts revealed a consistent tendency across ChatGPT, Claude, and DeepSeek to conflate geographically specific references with universally relatable concepts. When prompted to broaden the audience for poems containing localized cultural details, the models often replaced those details with generic phrasing, failing to accurately convey the original poem’s meaning while simultaneously implying a broader applicability than was warranted. This suggests a limitation in the models’ ability to discern the precise cultural context informing poetic language and to effectively translate those nuances for diverse audiences without losing crucial meaning or introducing inaccurate generalizations.

The Imperative of Refinement: Mitigating Bias in LLMs

The development of truly beneficial large language models (LLMs) necessitates a proactive approach to bias mitigation. These techniques aren’t simply about achieving technical accuracy; they are fundamental to ensuring fairness and equity in LLM outputs. Without careful attention to potential biases present in training data or model architecture, LLMs risk perpetuating and even amplifying societal prejudices across various applications, from loan applications and hiring processes to criminal justice and healthcare. Mitigating these biases requires a multi-faceted strategy, encompassing data curation, algorithmic adjustments, and ongoing auditing to align LLM behavior with established human values and ethical guidelines. Ultimately, prioritizing bias mitigation isn’t just a matter of responsible AI development – it’s essential for fostering public trust and unlocking the full, positive potential of these powerful technologies.

Mitigating bias in large language models necessitates a diverse toolkit of techniques, each addressing the problem from a unique angle. Data augmentation strategically expands training datasets with underrepresented perspectives, while model fine-tuning adjusts existing parameters to prioritize fairness metrics. More advanced methods, such as adversarial training, pit a ‘challenger’ model against the primary LLM, forcing it to learn more robust and equitable representations. Crucially, algorithmic auditing provides a systematic evaluation of model outputs, identifying and quantifying potential biases that might otherwise remain hidden – a process vital for continuous improvement and responsible deployment. These combined approaches represent a multi-faceted strategy for building LLMs that are not only powerful, but also demonstrably fair and aligned with ethical considerations.

Large language models often perpetuate societal biases not simply through malicious intent, but as a consequence of the statistical patterns they learn from massive datasets. These models excel at identifying and reproducing correlations, meaning that if biased language or associations are prevalent in the training data – such as gender stereotypes linked to specific professions – the model will inevitably reflect and amplify them. However, bias also arises from inherent limitations within the model architecture itself; the very mechanisms designed for generalization can inadvertently exacerbate existing inequalities or create new ones. Distinguishing between these two sources – statistical correlation in data versus architectural limitations – is critical for crafting effective mitigation strategies; addressing data bias requires careful curation and augmentation, while tackling architectural bias necessitates innovative model designs and training techniques that prioritize fairness and equity.

Large language models possess remarkable potential for innovation across diverse fields, yet realizing this creative capacity hinges on proactively mitigating embedded biases. Untreated, these biases can perpetuate and even amplify societal inequalities, leading to discriminatory outcomes in areas like hiring, loan applications, and even criminal justice. Addressing bias isn’t merely an ethical imperative; it’s a prerequisite for trustworthy AI. By systematically identifying and neutralizing prejudiced patterns within training data and model architecture, developers can cultivate LLMs that generate fairer, more equitable, and genuinely helpful outputs. This careful approach not only safeguards against harmful consequences but also unlocks the full spectrum of the technology’s creative possibilities, fostering innovation that benefits all of society.

The exploration of Large Language Models through poetic prompting reveals a fascinating fragility. These systems, built on complex algorithms, demonstrate a particular susceptibility to rhetorical patterns, echoing and sometimes distorting the creative impulses they process. As Claude Shannon observed, “The most important thing in communication is the amount of information that is conveyed.” This rings true when considering how LLMs interpret and re-present poetic forms; the ‘information’ isn’t merely the words themselves, but the underlying structures, biases, and limitations embedded within the model’s architecture. The study highlights how seemingly neutral systems aren’t impervious to decay-their responses, shaped by training data and algorithmic constraints, age not because of errors, but because time-and the constant influx of new data-is inevitable. Stability, in this context, isn’t a fixed state, but a temporary delay of the inevitable distortion of original creative intent.

What’s Next?

The exercise of discerning algorithmic rhetoric through poetic prompting reveals less a ‘black box’ to be decoded, and more a hall of mirrors reflecting humanity’s own linguistic biases. Every commit in this line of inquiry is a record in the annals, and every version a chapter documenting the LLM’s evolving capacity-and persistent limitations-in engaging with creative expression. The patterns identified are not inherent flaws, but rather the echoes of the datasets from which these models learn, amplified and re-presented. Delaying fixes to inherent biases is a tax on ambition; a continued insistence on novelty built upon unstable foundations.

Future work must move beyond diagnosis toward intervention. Simply identifying the presence of appropriation or biased response is insufficient. The challenge lies in developing methods to actively shape the LLM’s generative process, not through blunt censorship, but through the cultivation of more nuanced and ethically-grounded aesthetic sensibilities. This requires a shift from treating the model as a passive recipient of prompts to an active collaborator in the creative act – a notion fraught with philosophical implications.

Ultimately, the true metric of success will not be the LLM’s ability to mimic poetry, but its capacity to reveal new facets of language itself. The system will age, as all systems do. The question is not whether it will decay, but whether that decay will be graceful-a slow unraveling that illuminates the fragile beauty of the patterns it once held, or a sudden collapse into noise.

Original article: https://arxiv.org/pdf/2512.05243.pdf

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

See also:

2025-12-08 10:41