In a significant step toward improving road safety, Johns Hopkins University researchers have developed an artificial intelligence-based tool that can identify the risk factors contributing to car crashes across the United States and accurately predict sites of future incidents.
The tool, called SafeTraffic Copilot, aims to provide experts with both crash analyses and crash predictions to help reduce the rising number of fatalities and injuries that happen on U.S. roads each year.
The work,”SafeTraffic Copilot: adapting large language models for trustworthy traffic safety assessments and decision interventions” led by Johns Hopkins University researchers, is published in Nature Communications.
“Motor vehicle fatalities in the U.S. continue to increase, despite decades of countermeasures, and these are complex events affected by numerous variables, like weather, traffic patterns, roadway design, and driver behavior,” said senior author Hao “Frank” Yang, an assistant professor of civil and systems engineering at Johns Hopkins. “With SafeTraffic Copilot, our goal is to simplify this complexity and provide infrastructure designers and policymakers with data-based insights to mitigate crashes.”
The team uses a type of AI known as large language models, or LLMs, which are designed to process, understand, and learn from vast amounts of data. SafeTraffic Copilot was trained using text such as descriptions of road conditions, numerical values such as blood alcohol levels, satellite images, and on-site photography. The team’s model also has the ability to evaluate both individual and combined risk factors, offering a more detailed understanding of how these elements interact to influence crashes.
By design, SafeTraffic Copilot incorporates a continuous learning loop so that prediction performance improves as more crash-related data is entered into the model, making it even more accurate over time. By using LLMs, researchers can quantify the trustworthiness of the predictions—in other words, they can say a given prediction will be 70% accurate in a real-world scenario.
“By reframing crash prediction as a reasoning task and using LLMs to integrate written and visual data, the stakeholders can move from coarse, aggregate statistics, to a fine-tuned understanding of what causes specific crashes,” Yang said.
The model gives policymakers and transportation designers a trustworthy and interpretable tool to identify combinations of factors that elevate crash risk. The data can then be used to execute evidence-based interventions and more effective infrastructure planning to save lives and reduce injuries.
The researchers see the model as a copilot for human decision-making.
“Rather than replacing humans, LLMs should serve as copilots—processing information, identifying patterns, and quantifying risks—while humans remain the final decision-makers,” Yang said.
SafeTraffic Copilot has the potential to be a blueprint for responsibly integrating AI-based models into high-stakes fields, like public health and human safety. Because LLMs operate as large black-box models, users do not know how predictions are generated, deterring their use in high-risk decision-making scenarios. The team plans to continue their research to better understand how AI models can be used responsibly in those settings.
“The central focus of our ongoing research is to find the best way to combine the strengths of humans and LLMs so that decisions in high-stakes domains are not only data-driven, but also transparent, accountable, and aligned with societal values,” Yang said.
Study authors include Johns Hopkins doctoral candidates Yang Zhao, Pu Wang, master’s student Yibo Zhao, and Hongru Du, an assistant professor at the University of Virginia with a PhD from the Johns Hopkins Department of Civil and Systems Engineering.