The idea of applying clean-room design to model training is interesting... having a "dirty model" and a "clean model", dirty model touches restricted content and clean model works only with the output of the dirty model.

However, besides how this sidesteps the fact that current copyright law violates the constitutional rights of US citizens, I imagine there is a very real threat of the clean model losing the fidelity of insight that the dirty model develops by having access to the base training data.

IMO they just don't have any idea what data are actually copyrighted and are too lazy to invest in the problem.

Even if output is blocked, if it can be demonstrated that the copyrighted material is still in the model then you become liable for distribution and/or duplication without a license.

Training on synthetic data is interesting, but how do you generate the synthetic data? Is it turtles all the way down?

That would reduce the training quality immensely. Besides, any generalist model really needs to remember facts and texts verbatim to stay useful, not just generalize. There's no easy way around that.

I'm assuming that the goal of the bloom filter is to prevent the model from producing output that infringes copyright rather than hide that the text is in the training data.

In that case the model would lose the ability to provide relatively brief quotes from copyrighted sources in its answers, which is a really helpful feature when doing research. A brief quote from a copyrighted text, particularly for a transformative purpose like commentary is perfectly fine under copyright law.

But that would only hide the problem, doesn’t resolve the fact that models, in fact, violate copyright