I do not have any data regarding opt-in, but my impression was that as a rule, no one opts-in, and no one opts-out of basically anything (excepting the notorious cases, e.g., Real Player). If it worked, and somehow gave an at least somewhat unbiased sampling of the data, opt-in would obviously be best.

I am completely unsatisfied with the submit-and-run model. Feature engineering seems to rely on knowing your data really intimately, and that does not seem possible in a submit-and-run model.