It has been shown that working with databases from heterogeneous sources of varying fidelity can be leveraged in multi-fidelity surrogate models to enhance the high-fidelity prediction accuracy or, equivalently, to reduce the amount of high-fidelity data and thus computational effort required while maintaining accuracy [1]. In contrast, this contribution leverages low-fidelity data queried on a larger feature space to realize data-driven multi-fidelity surrogate models with a fallback option in regimes where high-fidelity data is unavailable. The fallback is constructed from low-fidelity data. Accordingly, methodologies are introduced to fulfill this task and effectively resolve the contradictions that inherently arise in multi-fidelity databases. In particular, the databases considered in this contribution feature two levels of fidelity with a defined hierarchy, where data from a high-fidelity source is, when available, prioritized over low-fidelity data. The proposed surrogate model architectures (i) filter out contradicting low-fidelity data during pre-processing, (ii) utilize transfer learning from low-fidelity to high-fidelity data and (iii) use an upstream classifier to resolve contradictions. They are examined in the context of an engineering use case in autonomous driving introduced in [2], where the human-seat interaction is evaluated using a data-driven surrogate model, that is trained through an active learning approach. It is shown that, compared to the non-hierarchical approach, two proposed architectures can achieve an improvement in accuracy on the high-fidelity data while simultaneously performing well on the low-fidelity data where high-fidelity data is unavailable.