AI Safety Test Reveals Critical Gaps in LLM Responses to Technology-Facilitated Abuse
A pioneering study published on arXiv (arXiv:2602.17672) has conducted the first expert-led evaluation of large language models (LLMs) in responding to technology-facilitated abuse (TFA) scenarios, revealing significant variations in response quality and raising important questions about AI safety for vulnerable populations.
The Growing Need for AI Support in Abuse Contexts
Technology-facilitated abuse represents a pervasive and evolving form of intimate partner violence where digital tools—from smartphones and social media to GPS trackers and smart home devices—are weaponized to control, surveil, and harm survivors. According to the study, while tech clinics provide crucial support for TFA survivors, they face substantial limitations including staffing constraints and logistical barriers that prevent many survivors from accessing timely assistance.
This accessibility gap has led increasing numbers of survivors to seek information online, creating what researchers describe as a "natural progression" toward consulting LLM-based chatbots before or instead of seeking professional help. With intimate partner violence organizations showing growing interest in AI tools, understanding how current models perform in this sensitive domain becomes critically important.
Methodology: Expert and Survivor Perspectives Combined
The research team employed a comprehensive evaluation framework, assessing four distinct LLMs: two widely used general-purpose non-reasoning models and two domain-specific models designed specifically for intimate partner violence contexts. Using real-world questions collected from existing literature and online forums where survivors seek advice, the study examined zero-shot single-turn responses generated with a survivor safety-centered prompt.
The evaluation criteria were specifically tailored to the TFA domain, going beyond general response quality metrics to include factors like:
- Safety prioritization in recommendations
- Actionability of suggested steps
- Contextual appropriateness for abuse scenarios
- Avoidance of harmful or victim-blaming language
In addition to expert assessment, the researchers conducted a user study to evaluate the perceived actionability of LLM responses from the perspective of individuals who have experienced technology-facilitated abuse, creating a rare dual-perspective evaluation.
Key Findings: Capabilities and Concerning Limitations
The study revealed significant disparities between general-purpose and domain-specific models. While specialized models demonstrated better understanding of abuse dynamics and more appropriate safety considerations, even these models showed concerning limitations in certain scenarios.
General-purpose models, despite their advanced capabilities in other domains, frequently failed to recognize the nuanced power dynamics in abusive relationships and sometimes provided advice that could inadvertently increase risk to survivors. The research identified specific patterns where models:
- Underestimated the potential for escalation in abusive situations
- Provided technically correct but contextually dangerous advice
- Failed to prioritize immediate safety concerns over longer-term solutions
- Displayed inconsistent understanding of legal and social support systems
Implications for AI Development and Deployment
These findings carry significant implications for AI developers, policymakers, and organizations supporting abuse survivors. As LLMs become increasingly integrated into help-seeking pathways, the study highlights the urgent need for:
Domain-Specific Fine-Tuning: The superior performance of specialized models suggests that generic safety training is insufficient for high-stakes domains like intimate partner violence. Developers need to incorporate domain expertise directly into model training and evaluation processes.
Safety-Centered Prompt Engineering: The study demonstrates that even with safety-focused prompts, models can generate problematic responses, indicating that prompt engineering alone cannot guarantee appropriate outputs in sensitive contexts.
Human-AI Collaboration Frameworks: The research suggests that LLMs may serve best as supplementary tools rather than primary advisors in abuse contexts, with clear pathways for escalation to human experts when needed.
Concrete Recommendations for Improvement
The study concludes with specific, actionable recommendations for improving LLM performance in supporting TFA survivors:
Enhanced Contextual Understanding: Models need better training on power dynamics, coercion patterns, and the specific ways technology is weaponized in abusive relationships.
Risk Assessment Integration: LLM responses should incorporate explicit risk assessment frameworks, recognizing that advice appropriate in low-risk situations may be dangerous in high-risk contexts.
Localized Resource Knowledge: Models require up-to-date, geographically specific information about legal protections, shelter availability, and support services.
Transparency About Limitations: Systems should clearly communicate their limitations and the importance of consulting human experts in potentially dangerous situations.
The Path Forward: Responsible AI for Vulnerable Populations
This research represents a crucial step toward developing AI systems that can safely support vulnerable populations. By establishing evaluation frameworks specifically designed for high-stakes domains and incorporating both expert and survivor perspectives, the study provides a model for future AI safety research.
As the authors note, "Our findings, grounded in both expert assessment and user feedback, provide insights into the current capabilities and limitations of LLMs in the TFA context and may inform the design, development, and fine-tuning of future models for this domain."
The growing accessibility of LLMs creates both opportunities and risks for abuse survivors seeking support. This study underscores that realizing the benefits while mitigating the dangers will require ongoing collaboration between AI researchers, domain experts, and the communities these systems aim to serve.
Source: arXiv:2602.17672v1, "Assessing LLM Response Quality in the Context of Technology-Facilitated Abuse," submitted January 11, 2026.
