(Re)mediators of Epistemic Injustice: Generative AI and Hermeneutic Resource Provision in Intimate Partner Violence

Paper Title

(Re)mediators of Epistemic Injustice: Generative AI and Hermeneutic Resource Provision in Intimate Partner Violence

Publication Info

• Topic area: The role of generative AI in mediating epistemic justice and harm for survivors of intimate partner violence (IPV).
• Keywords: Intimate partner violence, epistemic injustice, hermeneutic resources, generative AI, large language models, public health, harm evaluation, survivor support, algorithmic harm, digital abuse.

Background and Problem

• Problem / challenge: Survivors of IPV face epistemic injustice, including testimonial and hermeneutical barriers, which limit their ability to disclose and understand their experiences. Current generative AI tools may both alleviate and exacerbate these issues, but their role in mediating epistemic harm is underexplored.
• Significance: IPV is a global public health crisis, affecting millions annually and leading to severe physical, emotional, and social consequences. Addressing barriers to help-seeking and understanding IPV is critical for prevention and intervention.
• Motivation and related work: Prior research highlights the role of technology in IPV, including its use for surveillance and harm. However, the implications of generative AI in IPV contexts remain underexplored, particularly in terms of its potential to mediate epistemic injustice and harm.

Solution

• Proposed approach: This paper investigates how generative AI, specifically large language models (LLMs), mediates epistemic justice and harm for IPV survivors by analyzing AI responses to survivor disclosures.
• Novelty:
  1. A taxonomy of seven generative AI use cases in IPV contexts.
  2. Quantitative evaluation of LLM responses using public health metrics (e.g., readability, hermeneutic resource provision).
  3. Qualitative analysis of risks and epistemic harms mediated by LLMs in IPV contexts.
  4. Cross-disciplinary recommendations for harm evaluation and remediation.
• Procedure and key techniques:
  • Data collection from five IPV-related subreddits (January 2023–July 2024).
  • Content analysis of 613 posts to identify generative AI use cases.
  • Generation and evaluation of LLM responses (GPT-4o, Claude, Gemini) using metrics like cosine similarity to National Domestic Violence Hotline resources, readability (Flesch-Kincaid Grade Level), and comprehension (CDC Clear Communication Index).
  • Risk analysis using Zhou et al.’s public health risk taxonomy.

Results

• Concrete findings:
  • Survivors use generative AI for seven purposes: hermeneutic resourcing, facilitating abuse, composing disclosures, unconventional support strategies, interpreting interactions, detecting abuse, and preventing abuse.
  • Claude responses aligned most closely with expert IPV resources (+0.39% average similarity), while GPT-4o responses were less aligned (-0.31% average similarity).
  • GPT-4o responses were the most complex (average grade level: 13.79), often exceeding survivors’ communication levels (average grade level: 7.9).
  • None of the LLMs consistently met public health communication standards (CDC Clear Communication Index score ≥90).
• Advantage over baselines:
  • Claude provided the most survivor-focused responses, emphasizing safety and harm reduction.
  • Gemini demonstrated balanced, consistent performance across contexts, though less nuanced than Claude.
  • GPT-4o responses, while structured and comprehensive, often dismissed survivor concerns or reinforced harmful narratives.
• Experiments / evaluation:
  • LLM responses were evaluated for semantic similarity to IPV resources, readability, and comprehension.
  • Risk analysis highlighted epistemic harms, including misidentification of abuse, additional barriers to help-seeking, and inequities in resource provision.
• Limitations and future work:
  • Lack of cultural, racial, and linguistic nuance in the evaluation.
  • Limited exploration of perpetrator use of generative AI.
  • Future work should include diverse cultural perspectives and examine generative AI’s role in facilitating abuse.

Summary

This study explores how generative AI mediates epistemic justice and harm for IPV survivors, identifying seven use cases and evaluating LLM responses using public health metrics. While tools like Claude provide survivor-focused support, others like GPT-4o risk amplifying harm through dismissive or misaligned responses. The findings highlight the dual role of generative AI as both a remedy and mediator of harm, emphasizing the need for cross-disciplinary evaluation and remediation strategies. Future work should address cultural and linguistic disparities and explore the broader implications of generative AI in IPV contexts.