Massimo Arnoldi and Massimo Milan (from Lifeware) gave a talk on building tests from real objects. The point of the presentation seemed to be that if you have a lot of data, and you need to ensure that changes to programs don't affect the calculation of the data, then if you can translate the objects into a program that's capable of programmatically re-creating the data afterwards, you can ensure that the generated data is the same as the original data.
I'm sure I missed something in this presentation. After all, there's no real difference in doing this for loading/saving an object to a database, other than the fact that instead of running a program to instantiate the object, you're effectively unmarshalling it. I'm also not sure what this is supposed to test ... unless what you're testing is your ability to unmarshal the data again. Maybe there's better coverage of this presentation elsewhere, or you'll be able to see the video and presentation for yourself. The video is now available.