As artificial intelligence becomes embedded in everyday systems, people increasingly face situations where AI assists their decisions, raising questions about when to trust those recommendations and when to act independently. A central challenge is whether people can appropriately judge when to trust an AI and when to take control, and whether factors such as poor visibility or individual attitudes toward technology affect that judgment. 

This study examined how visual uncertainty and technology readiness influenced the decision to manually override an AI system in a time-pressured driving task. Thirty participants supervised a semi-autonomous vehicle that executed lane-switching decisions by default, behaving reliably in most trials but producing errors in a subset. Three levels of weather visibility were manipulated: clear, low fog, and heavy fog. Before the experiment, participants completed the Technology Readiness Index 2.0. After the task, they completed a trust in automation scale and scenario-specific trust items, followed by a semi-structured interview. 

The results revealed a clear dissociation between two measures of override behavior. Override frequency remained stable across visibility conditions and was not predicted by technology readiness traits. However, the time taken to reach an override decision increased substantially as visibility deteriorated: under heavy fog, participants used on average 91% of the available response window, compared to 60% under clear conditions. Post-experiment trust ratings were associated with faster intervention under clear conditions and greater behavioral adaptation across conditions. Qualitative analysis identified five themes in how participants reasoned about trust and reliance. A behaviorally distinctive subgroup of eight participants who explicitly recognized their own susceptibility to automation complacency demonstrated significantly better calibration to the AI's actual reliability under the most uncertain conditions, despite showing similar questionnaire scores to the rest of the sample, suggesting that metacognitive awareness may function as a behavioral moderator that standard psychometric instruments do not capture. 

These findings suggest that measuring both the frequency and the timing of human overrides provides a more complete picture of trust calibration than frequency alone. Interface designs that make AI behavior feel interruptible and operator-led may better support appropriate human oversight in safety-critical contexts.