Earlier quoted context omitted.
USB does use error detection (CRC), but if you transfer enough data over anything, you'll eventually see undetected errors.
If this is true, then the MTBF design parameter is unacceptably low.
What it probably does not have is a detectable error counter that marks a link as broken (e.g. in case of a bad/damaged cable, plug or controller) before the chance of an undetected error becomes unacceptably large, although it seems like a host would be able to implement that as an optional feature?