The document discusses issues with bidirectionality in model transformations. It notes that while bidirectionality is relevant, it rarely produces anticipated benefits due to several factors: transformations are often non-deterministic, existing bidirectional languages introduce opaque semantics by enforcing consistency through unknown update policies, and developers lack control over the transformation behavior. The document proposes that bidirectional transformations should generate all possible results to manage uncertainty, define update policies at design time to give developers control, and allow manual selection when automatic policies cannot be determined.