Useful generative AI is only possible by the same collection of copyrighted material previously deemed illegal. I imagine a model trained solely on material explicitly marked for AI use would be significantly worse.

On the flip side - almost (say, 99.99%) of Human Engineers, Artists, Technicians, Scientists, etc... have mental models trained on copyrighted material.

Nobody (to my knowledge) has ever said that you can't train on copyrighted material - what you aren't allowed to do is copy or directly plagiarize. Something that all the generative systems are going to great pains to remove from their system where possible.

Are they doing a perfect job - nope. But they'll get better, and this is good - copyrights are supposed to prevent replication, not use of their material.

Lots of people have said that it's illegal to train on copyrighted material without a license to it.

Also, performance licenses for movies, plays, recorded music, and copyrighted scores are all required. The lack of copying is not relevant there, the performance alone can be infringing.

I'm intrigued - can you point me to any credible commentator or article that makes the argument "it's illegal to train on copyrighted material without a license to it" - that seems entirely contrary to the spirt (as I understand it) to copyright law, which is it grants you a right to copy.

I like that you bring up music - almost every musician you have ever listened to (some exceptions of course) developed their talent by learning from others - chords, bridges, etc... And I'm just as certain that close to 0% of them had a "license to learn" from their material.

I'm talking about ML training. I think human training is expressly covered by fair use (i.e. copying for educational purposes). Sorry for the confusion, I misread your comment.

>Are they doing a perfect job - nope. But they'll get better

I lost all optimism with tech hoping "they'll get better" quite a while ago. No, it's time to regulate them before they burn the bridge this time.

It uses a lot of copyrighted content without permission.

What permission would be needed? It's READING it.

Are there problems when it reproduces it verbatim? Absolutely, but that's not what the copyright maximalists are talking about. They're saying it's a violation of their right of reproduction when someone simply reads their work.

I remember when people were upset that someone would (gasp!) link to their site without permission. Especially if it was a "deep link".

Expanding copyright in this novel way obviously is going to lead to a whole slew of problems, least among which is going to be ensuring regulatory capture by the largest of the largest simply because only they will have the money to license reading.

It's a shakedown by industries that no longer have a viable business model.

>What permission would be needed? It's READING it.

and then storing it in a database, as shown by the ability to nearly replicate images with enough prompting before they band-aided a fix over it (which does not remove it from their database). It's clearly not just "reading". You can argue the same for a human mind, but it's a lot easier to peer into a mind of code for now (and honestly, by the time we can accurately read brainwaves LLM's won't even be in the top 10 of ethical concerns anyway).

All that aside, web scraping has been legally contentious for over a decade. This mass scraping for commercial LLM usage is honestly making a horrible argument for that already dubious factor.

>Expanding copyright in this novel way obviously is going to lead to a whole slew of problems, least among which is going to be ensuring regulatory capture by the largest of the largest simply because only they will have the money to license reading.

It's probably for the best, since at that point at least the owners of the data are getting paid (though there's other grey areas to iron out. Especially with User-generated content being sold as if the site "owns it", while being legally exempt from being sued for hosting it). the opposite effect just means the corporations win indirectly instead of directly, with less money flowing around. A company that can outspend the competition can also out spend on hardware to process faster, scrape more, and polish he final effects. There's no endgame here where the corporation loses and the indies win, short of some absolutely radical policy changes.

Generative AI is computational plagiarism of diffuse - but still copyrighted, sources.

> computational plagiarism

Plagiarism is copying without attribution. Transformative works do not qualify as plagiarism.

when does it credit/attribute who/what/when/where the data it copied from?

You don't need to credit for a transformative work. LLMs don't regurgitate like you think they do.

> LLMs don't regurgitate like you think they do.

"Copilot has been found to regurgitate long sections of licensed code without providing credit — prompting this lawsuit that accuses the companies of violating copyright law on a massive scale."

https://www.theverge.com/2022/11/8/23446821/microsoft-openai...

Like I said, they don't regurgitate the way you think they do. Doesn't mean it can't regurgitate data if it's overfit to the training data.

>they don't regurgitate the way you think they do.

Copyright is about the outcome, not the destination. Short of edge cases where two contemporary inventors in different parts of the world funnel upon the same novel idea at the same-ish time, the method doesn't change how we interpret infringement.

So the fact that it's capable of doing it is enough to bring about legal suspicion.