Conversational versus Dashboard Explainable AI for UAV Intrusion Detection: An Empirical Study of Operator Trust and Reliance

Conversational AI interfaces powered by large language models were perceived as more useful than traditional dashboard visualizations for understanding complex cybersecurity decisions in UAV networks, potentially because they help users access and synthesize relevant information more easily.

There is a critical trade-off between usability and appropriate skepticism: while the conversational interface improved perceived usefulness, it also led to lower levels of appropriate self-reliance, meaning operators were more likely to accept AI recommendations without verifying underlying evidence when the system made errors.

Natural language responses may create a false sense of confidence: the study suggests that receiving explanations in conversational, natural language form made AI advice "easier to accept," which reduced participants' tendency to question or verify the system's reasoning—a concerning finding for high-stakes security applications.

Current explainability approaches struggle with complex, multimodal data: traditional static dashboards have difficulty presenting the intricate relationships among multiple types of cyber-physical data in UAV networks, highlighting a genuine challenge in making AI systems interpretable for operators.

Responsible AI deployment requires balancing transparency with appropriate user skepticism: these findings suggest that making AI systems more accessible and user-friendly should not come at the cost of encouraging operators to critically evaluate AI recommendations, particularly in security-critical applications where errors have serious consequences.