From Notepad AI to Social Media: How Can Text Style Transformation Mitigate Social Harm?

Abstract: The rapid proliferation of harmful and emotionally damaging content on social media platforms has intensified concerns regarding societal harm. While content moderation efforts primarily focus on detecting and removing harmful posts, less attention has been given to mitigating harm through stylistic text transformation while preserving semantic meaning. In this paper, we propose a writing-assistance framework that can reduce societal harm by transforming aggressive, toxic, or emotionally harmful comments into softer, more neutral stylistic forms inspired by Notepad AI, a simple AI writing assistant. Rather than censoring or suppressing speech, we apply controlled stylistic modifications to preserve core informational content while reducing emotional intensity and identity-based attacks. We introduce an Emotion Drift Index (EDI) metric to systematically quantify emotional change and evaluate the effectiveness of stylistic rewriting, thereby reducing harmful interactions in online environments.

• The paper introduces a controlled stylistic rewriting framework to neutralize emotionally harmful content while preserving core meaning.
• Using prompt engineering with LLMs, it quantitatively measures affective shifts with the innovative Emotion Drift Index (EDI).
• Empirical results demonstrate significant emotion drift, particularly with inspirational and humorous styles, supporting safer social media interactions.

Stylistic Text Transformation as Harm Mitigation in Social Media

Problem Context and Motivation

The proliferation of emotionally harmful content on digital platforms presents persistent challenges for moderation and user well-being. Traditional approaches focus on detection and removal but provide limited nuance in preserving user intent and information. The paper "From Notepad AI to Social Media: How Can Text Style Transformation Mitigate Social Harm?" (2604.27365) proposes an alternative path: leveraging stylistic text transformation, inspired by Notepad AI writing assistant functionality, to neutralize emotionally charged or toxic language and systematically quantify the resulting affective impact through an Emotion Drift Index (EDI).

Core Contributions

The primary technical contributions are the formulation of a writing-assistance framework capable of controlled stylistic rewriting, designed to minimize social harm while maintaining semantic continuity. The system utilizes prompt engineering with LLMs, notably Microsoft Phi-3, to create rewrites in four distinct text styles: formal, casual, inspirational, and humor. Crucially, the framework introduces EDI—a quantitative metric grounded in the Valence-Arousal-Dominance (VAD) circumplex—to measure the affective displacement induced by stylistic transformation. This index provides empirical grounding for assessing harm mitigation efficacy, focusing on emotion drift within the rewritten content at both sample and dataset-level aggregation.

Technical Approach

Data and Tools

Two benchmark datasets are employed: Toxic Comment Classification Challenge and HateXplain, both offering granular annotations for multiple harmful speech categories. For emotion detection, the SamLowe/roberta-base-go_emotions model is used, mapping 27+1 emotion categories to six core emotions, facilitating VAD-based quantitative analysis. Phi-3 LLM applies zero-shot prompt engineering for stylistic rewriting, generating variants across the four specified styles.

Methodology

The process systematically consists of:

• Filtering: Identification of harmful textual content via conventional moderation filter.
• Emotion Detection: Assigning ground-truth emotion labels via RoBERTa model, mapped into VAD space.
• Style Transformation: Phi-3 rewrites input with controlled style prompts, targeting specified stylistic configurations.
• Emotion Drift Quantification: VAD embeddings for both original and rewritten text are compared via Euclidean norm to compute ED, aggregated into EDI.
• Evaluation: Rate of emotional change and EDI analyzed across dataset, measuring impact per transformation style.

This pipeline is designed for modularity and reproducibility, intentionally relying on non-finetuned models to establish baseline efficacy.

Empirical Results

The experiments reveal robust emotion shifts attributable to stylistic rewriting. Key findings include:

• High Change Rates: Inspirational and humorous styles induce substantial emotional transitions, with inspirational style yielding up to 76% change in the toxic comment dataset.
• Directional Emotion Drift: Negative emotions (anger, sadness, disgust) are frequently re-mapped toward positive or neutral emotions (happiness, surprise), as visualized in multi-style heatmaps.
• Quantitative Correlation: EDI consistently tracks with emotion change rates; higher EDI values predict larger affective displacements, e.g., inspirational transformation on toxic comments produces EDI of 1.82.
• Semantic Preservation: Despite affective drift, deliberate prompt construction and LLM instruction ensure core informational content remains intact.

These results support the hypothesis that stylistic rewriting, properly controlled and quantified, can reduce risk of psychological distress and social harm at scale.

Limitations and Ethical Considerations

The study confines itself to six basic emotion categories for standardization, potentially omitting relevant subtleties in affective state representation. The use of separate LLMs for emotion detection and rewriting avoids bias but does not leverage joint optimization, which could improve results. The framework intentionally excludes offensive content from dissemination and addresses privacy by referencing prior work on device-level privacy-preserving stylistic transfer (Soumik et al., 4 Jun 2025).

Practical and Theoretical Implications

Practically, this paradigm extends moderation from binary removal to adaptive rewriting, creating less disruptive user experiences and potentially increasing tolerance for nuanced discourse. The EDI metric provides moderation systems with operational levers for dynamic content mitigation, establishing a standardized affective risk assessment. Theoretically, the paper demonstrates that linguistic surface features directly modulate emotional reception, providing evidence for future fine-grained social harm intervention studies.

Further development may integrate more granular emotion taxonomies, joint LLM architectures for end-to-end rewriting and detection, and real-time deployment scenarios, as well as incorporating user feedback loops to assess subjective harm reduction.

Conclusion

Stylistic text transformation, enabled by LLM prompt engineering and systematically quantified via Emotion Drift Index, offers viable mitigation of emotional harm in social media. The empirical evidence establishes correlation between rewriting styles and affective displacement, particularly the redirection of negative emotions toward neutral/positive affective states. This approach augments traditional moderation paradigms, providing a reproducible, adaptive mechanism for safer digital communication environments with strong practical signals of efficacy and a foundation for further research.