I'm also curious why they took action on ethical basis, yet don't take action on legal bases. There are models overfitted on Copyrighted data, there are datasets with personally-identifying information, etc. Why not start with just following the law first?
@alexjc
-
AI Training on Copyrighted Data: Double Standards in Copyright Enforcement
By
–
They said (paraphrase) "we train the audio models on CC data because training on copyrighted data would be problematic." Thus, they imply that training on copyrighted images is also problematic — but they don't care because there's no strong lobby like RIAA there.
-
Audio Model Admissions and Copyright Law Discovery Issues
By
–
Admitted it in interviews about the audio model. Many statements made are problematic if you look at them in the context of discovery. (About copyright law, yes. What about it?) https://
x.com/RedSagitter/st
atus/1614657872532766723
… -
Raw latent access improves reconstruction quality
By
–
Yeah, it's not well explained and disparaging — but there's some truth to it. You could obtain better reconstructions if you had raw access to the latents.
-
Fair Use as Defense Against Copyright Infringement Claims
By
–
They say "unauthorized copies" is that what you mean? If so, then it's an appropriate statement. Fair Use, which many in the community claim is at play here, is a defence you use to clear the liability of infringement — but the infringement is admitted.
-
Safeguards and Fair Compensation for AI-Generated Content
By
–
Yes, I can understand that perspective. I think more safeguards and fair compensation need to be in place for that to work. e.g. adding a kNN search to all generated outputs to notify of infringement on training data, and dedicating 10% income to charities supporting artists.
-
Opinion Polarization in AI Art Debate and Copyright Concerns
By
–
I think the worst part is that this opinion seems to be the root cause of the polarization in the AI Art debate. If he came out and said the same thing as for their music models ("they are infringing so we're being careful") then the debate would become reasonable in comparison.
-
Future of coding as source code representation
By
–
If "coding" means producing a source code representation, then the quote makes some sense as one possibility in the future.
-
Annotation Quality Comparison and ML Impact on Professional Fields
By
–
Does the paper also compare the quality of the annotations? Re: Tsunami. There was a recent thread about how the market for translation was destroyed by ML. It will happen one field at a time… so the impact is incremental.
-
AI Artists Must Understand Fair Use Implications Carefully
By
–
This is another reason why AI Artists and companies in this space should not be running around willy-nilly claiming Fair Use without really understanding the consequences.