Responsibility attribution by large language models (LLMs) in road safety analysis

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Victim blaming in safety analysis occurs when attention is directed toward individual behavior rather than the systemic factors shaping safety risks, weakening prevention efforts. As Large Language Models (LLMs) begin assisting in safety–critical decision-making, it is important to examine whether they reproduce such tendencies. This study tested 144 road-crash scenarios using ChatGPT-4o and DeepSeek-V3 under a structured interaction protocol to assess how they assign responsibility. Scenarios varied by risk behavior, injury severity, demographics, national context, and driving purpose. Each scenario was analyzed through three sequential prompts examining prevention strategies, primary responsibility attribution, and AcciMap-based ratings across six system levels. When asked about prevention, nearly 90% of model recommendations targeted systemic interventions, including policy, infrastructure, and organizational measures, indicating systems-oriented reasoning under this prompting condition. When asked to assign responsibility, however, both models shifted toward narrower, context-driven attribution. Private driving scenarios produced full driver responsibility, whereas work-related scenarios assigned primary responsibility to employers in 69% of cases. These patterns suggest that the core challenge observed here is not demographic bias but prompt-sensitive analytical inconsistency. Both models shifted between systemic and individual analytical orientations depending on question structure rather than consistently applying systems safety principles across prompts. These findings highlight analytical consistency as a key requirement for the responsible use of generative AI in safety–critical contexts such as transportation safety.