3 ms·

I see a parallel between the news linking case and AI training on copyrighted content (without regurgitating it back later). If AI can't use ideas from copyrigh

I see a parallel between the news linking case and AI training on copyrighted content (without regurgitating it back later). If AI can't use ideas from copyrighted works, then nobody can. Because anyone could be secretly using AI we have to apply the same strict standards of attribution to all human works. And that would lead to a chilling effect, any creative act takes the risk of incidental infringement. Should copyright owners control the external discourse about their work, or own the ideas in their works? Should we upgrade AI from protected expression to protected ideas?

cout 3y ago
> If AI can't use ideas from copyrighted works, then nobody can. Because anyone could be secretly using AI we have to apply the same strict standards of attribution to all human works. Yes, we should be applying strict standards of attribution to all works, human or otherwise. IMO one of the biggest defects of language models currently is lack of attribution; if sources are not properly cited, there is no way to verify the veracity of what is written. Unfortunately, apart from the technical difficulty of implementing attribution, the exact combination of sources is part of a model's "secret sauce", so there is an economic incentive to not cite.