Assessing LLM Response Quality in the Context of Technology-Facilitated Abuse

A groundbreaking study from researchers at UW-Madison and Cornell has conducted the first expert-led evaluation of large language models' performance on technology-facilitated abuse (TFA) scenarios. The team assessed four LLMs—two widely used general-purpose models (likely including GPT-4 and Claude) and two domain-specific models designed for intimate partner violence contexts—using real-world questions collected from literature and online forums. They employed a survivor safety-centered prompt and evaluated responses on criteria specifically tailored to the TFA domain.

The evaluation revealed significant gaps in LLM performance. Models struggled with accuracy, nuance, and providing actionable safety advice when responding to TFA-related questions. The researchers conducted both expert assessments and a user study with individuals who have experienced TFA, finding that responses often lacked the specificity and contextual understanding needed for effective survivor support. This is particularly concerning given that survivors may increasingly turn to AI chatbots before seeking help from understaffed tech clinics.

The study's findings highlight a critical safety issue as LLMs become more accessible. While IPV organizations show growing interest in AI tools, current models appear ill-equipped to handle the complexity of abuse scenarios. The researchers conclude with concrete recommendations for improving LLM performance in this domain, emphasizing the need for specialized training data, safety protocols, and human oversight. This work establishes a crucial benchmark for evaluating AI systems in sensitive, high-stakes applications where incorrect advice could have serious consequences.