Sixth-generation (6G) networks are expected to evolve toward AI-native and cross-layer mobile intelligence. However, many current AI-enabled wireless designs remain prediction-driven, producing estimates or decisions without exposing the reliability of those outputs to the rest of the system. This is especially problematic in applications such as teleoperation, where physical-layer AI functions, such as channel estimation or link-quality prediction, may become unreliable under fast-varying conditions. If higher layers, including MAC-layer functions such as scheduling and link adaptation and network-layer functions such as routing and mobility management, act on such outputs without accounting for their confidence or validity, errors may propagate across the protocol stack and degrade end-to-end performance. In this vision paper, we highlight that future 6G systems should move toward trust-aware cross-layer intelligence. We envision a trust-aware mechanism that complements existing 3GPP control and signaling procedures by attaching reliability-related information to AI outputs. We outline a conceptual system model spanning the physical (L1), MAC (L2), and network/service (L3) layers, with a teleoperation use case, and highlight the key open challenges for building more explainable and trustworthy AI-native mobile systems.