I think any analogy with physical theft is too different from data to apply to this comparatively subtle case. Especially when we get into the details of just using the output of the model to train on.