Evaluating LLMs for Roadway Design Compliance: An MUTCD-Based Evaluation of GPT-4 and Claude 3.5

Traffic design has long relied on rule-based frameworks, empirical data, and professional judgment to ensure safe and efficient roadway operations. Practitioners frequently consult design manuals such as the Manual on Uniform Traffic Control Devices (MUTCD) and state supplements to determine appropriate traffic control strategies. While this process provides regulatory consistency, it remains time-consuming and dependent on the engineer’s ability to interpret and cross-reference complex technical provisions. Recent advancements in artificial intelligence (AI) have introduced new possibilities for automating design-related reasoning and documentation tasks. This study explores how different LLMs perform when tasked with interpreting MUTCD-based roadway design prompts. Specifically, two cases were examined: (1) the choice between installing a new signal or a single-lane roundabout at a suburban four-leg intersection, and (2) the development of a temporary traffic control plan for a work-zone lane closure on a multi-lane urban arterial. By comparing the citation accuracy and contextual interpretation of ChatGPT and Claude, this paper evaluates the reliability of LLMs in applying MUTCD provisions and discusses their potential as rapid-assessment tools to support engineering decision-making.