Virtual prototypes are simulators used in the con- sumer electronics industry. They enable the development of embedded software before the real, physical hardware is avail- able, hence providing important gains in speed of development and time-to-market. Transaction-level Modeling (TLM) is a widely used technique for designing such virtual prototypes. Its main insight is that many micro-architectural details (i.e. caches, fifos and pipelines) can be omitted from the model as they should not impact the behavior perceived from a software programmer's point-of- view. In this paper, we shall see that this assumption is not always true, specially for low-level software (e.g. drivers). As a result, there may be bugs in the software which are not observable on a TLM virtual prototype, designed according to the current modeling practices. We call this a faithfulness issue. Our experience shows that many engineers are not aware of this issue. Therefore, we provide an in-depth and intuitive explanation of the sort of bugs that may be missed. We claim that, to a certain extent, modified TLM models can be faithful without losing the benefits in terms of time-to-market and ease of modeling. However, further investigation is required to understand how this could be done in a more general framework.