LLM-assisted systematic review of inaccurate mental models for driving automation systems: Constructs, KPIs and measurements

Fang, Y., Bazilinskyy, P., Martens, M. H.

Submitted for publication

ABSTRACT Research on the interaction between users and driving automation systems (DAS) is growing rapidly. To support safe interaction, researchers have examined how drivers comprehend these systems, yielding the concept of mental models. The literature shows that users naturally form incomplete representations, termed inaccurate mental models (IMM). However, research on measuring IMM remains fragmented: studies typically isolate single dimensions, relying on declarative knowledge questionnaires, behavioural indicators or related constructs such as miscalibrated trust, and no coherent framework links these metrics to the cognitive mechanisms underlying safety-critical risks. In this systematic review, we investigate how IMM is interpreted and assessed in in-vehicle human-DAS interaction at SAE Levels 1-4, with the goal of developing a unified assessment framework that supports systematic diagnosis. Methodologically, we used a large language model (GPT-5.2) as an independent second reviewer working in parallel with a human expert, introducing a collaborative human-LLM screening strategy for human-machine interface review research. Of 1,178 unique records, 53 were included (38 empirical and 15 theoretical). We organise the evidence into an initial mental model (explicit and implicit conceptions) and a situational mental model (situational understanding, situational attitude, and behaviour). Across these five dimensions, we identify 7 constructs and 30 key performance indicator (KPI) categories and map them to measurement methods, automation levels, and driving contexts. The framework supports developers and researchers in choosing complementary measures and converting abstract cognitive risks into quantifiable data, with the aim of diminishing inaccurate mental models for safe human-machine interaction (HMI). interviewpreprintautomated-drivingartificial-intelligence