1y ago

Japan determines copyright doesn't apply to LLM/ML training data

infosec.town Taggart :donor: (@mttaggart)

Japan determines copyright doesn't apply to LLM/ML training data. > On a global scale, Japan’s move adds a twist to the regulation debate. Current discussions have focused on a “rogue nation” scenario where a less developed country might disregard a global framework to gain an advantage. But with J...

cross-posted from: https://programming.dev/post/8121669

Taggart (@mttaggart) writes:
Japan determines copyright doesn't apply to LLM/ML training data.
On a global scale, Japan’s move adds a twist to the regulation debate. Current discussions have focused on a “rogue nation” scenario where a less developed country might disregard a global framework to gain an advantage. But with Japan, we see a different dynamic. The world’s third-largest economy is saying it won’t hinder AI research and development. Plus, it’s prepared to leverage this new technology to compete directly with the West.
I am going to live in the sea.
www.biia.com/japan-goes-all-in-copyright-doesnt-apply-to-ai-training/

Programming @programming.dev

ericjmorey @programming.dev

1y ago

Japan determines copyright doesn't apply to LLM/ML training data

infosec.town /notes/9o2c3aijben6rgxe

78 22

Machine Learning @programming.dev

ericjmorey @programming.dev

1y ago

Japan determines copyright doesn't apply to LLM/ML training data

infosec.town /notes/9o2c3aijben6rgxe

26 3

66 comments

Nice, time to train one with all the Nintendo leaks and generate some Zelda art and a new Mario title!
- train one with all the Nintendo leaks
  This is fine
  generate some Zelda art and a new Mario title
  This is copyright infringement.
  The ruling in japan (and as I predict also in other countries) is that the act of training a model (which is just a statistical estimator) is not copyrightable, so cannot be copyright infringement. This is already standard practice for everything else: You cannot copyright a mathematical function, regardless of how much data you use to fit to it (that is sensible: CERN has fit physics models to petabytes worth of data, that doesn't mean they hold a copyright on laws of nature, they just hold the copyright on the data itself). However, if you generate something that is copyrighted, that item is still copyrighted: It doesn't matter whether you used an AI image generator, photoshop, or a tattoo gun.
  
  But you could ask it to generate something in the style of Zelda, without being Zelda.

I think this is a difficult concept to tackle, but the main argument I see about using existing works as 'training data' is the idea that 'everything is a remix'.
I, as a human, can paint an exact copy of a Picasso work or any other artist. This is not illegal and I have no need of a license to do this. I definitely don't need a license to paint something 'in the style of Picasso', and I can definitely sell it with my own name on it.
But the question is, what about when a computer does the same thing? What is the difference? Speed? Scale? Anyone can view a picture of the Mona Lisa at any time and make their own painting of it. You can't use the image of the Mona Lisa without accreditation and licensing, but what about a recreation of the Mona Lisa?
I'm not really arguing pro-AI here, although it may sound like it. I've just heard the 'licensing' argument many times and I'd really like to hear what the difference between a human copying and a computer copying are, if someone knows more about the law.
- Um - your examples are so old the copyright expired centuries ago. Of course you can copy them. And you can absolutely use an image of the Mona Lisa without accreditation or licensing.
  Painting and selling an exact copy of a recent work, such as Banksy, is a crime.
  … however making an exact copy of Banksy for personal use, or to learn, or to teach other people, or copying the style… that’s all perfectly legal.
  I don’t think think this is a black and white issue. Using AI to copy something might be a crime. You absolutely can use it to infringe on copyright. The real question is who’s at fault? I would argue the person who asked the AI to create the copy is at fault - not the company running the servers.
  
  Painting and selling an exact copy of a recent work, such as Banksy, is a crime.
  … however making an exact copy of Banksy for personal use, or to learn, or to teach other people, or copying the style… that’s all perfectly legal.
  And that was the bait and switch of OpenAI! They sold themselves as being a non-profit simply doing research, for which it would be perfectly legal to consume and reproduce large quantities of data... And then, once they had the data, they started selling access to it.
  I would say that that alone, along with the fact that they function as gatekeepers to the technology (One does not simply purchase the model from OpenAI, after all) they are hardly free of culpability... But it definitely depends on the person trying to use their black box too.
  
  Thanks for your response. I realize I muddied the waters on my question by mentioning exact copies.
  My real question is based on the 'everything is a remix' idea. I can create a work 'in the style of Banksy' and sell it. The US copyright and trademark laws state that a work only has to be 10% differentiated from the original in order to be legal to use, so creating a piece of work that 'looks like it could have been created by Banksy, but was not created by Banksy' is legal.
  So since most AI does not create exact copies, this is where I find the licensing argument possibly weak. I really haven't seen AI like MidJourney creating exact replicas of works - but admittedly, I am not following every single piece of art created on Midjourney, or Stable Diffusion, or DALL-E, or any of the other platforms, and I'm not an expert in the trademarking laws to the extent I can answer these questions.
  
  To be at fault the user would have to know the AI creation they distributed commits copyright infringement. How can you tell? Is everyone doing months of research to be vaguely sure it's not like someone else's work?
  Even if you had an AI trained on only public domain assets you could still end up putting in the words that generate something copyrighted.
  Companies created a random copyright infringement tool for users to randomly infringe copyright.
  
  Your example is a dude who paints unsolicited on other people's property. What kind of copyright does a ghost have?
- Here's the thing... Generative AI had a plagiarism/remix phase. It raised some serious questions about copyright
  It lasted for a matter of weeks.
  We're all still stuck up on it, but go to civit.ai
  Play with it. Look at what people are creating.
  If you're not convinced, put up a bounty for something extremely specific
  Art has changed. There's no putting it back in the bottle, this is the tiniest leading edge of the singularity
  
  Just a small warning, I just played around with civit. Tried to make some Images, also wanted to try to make some nsfw images. Anyway be really careful what you prompt, I accidentally generated some images with very young people I never intended.

Japan tech economy go brrrrr
- Or it leads the way in producing the most useless, misleading bullshit more efficiently. We'll see.
  
  I didn't say it'd produce good things other than economy points.
  
  Is there a c/WTFJapan? Asking to keep track of future developments
- Maybe that would finally get them to stop using fax machines.
  
  Not sure this is the flex you think it is. The US health industry utilizes fax to send client health information millions of times a day, and it is considered a secure communication.

What's stopping somebody from making an LLM that can reproduce media that was used in its training with close to 100% accuracy? If that happens, then we'll have a copyright laundering service.
- Reproducing copywrited works would be a problem. Consuming them is not.
  In your example, a copyright case would be able to move forward and be tested in court. I would think it stands as good of a shot at prevailing in that example. It would be the same as a case against someone who wrote a script for a website to reproduce copyrighted work on command. The difference is this isn't that. And if and when it does that, the ai can be tuned to prevent it from continuing to do it.
  
  Hi chatgpt7, I like legend of Zelda tears of the second kingdom, please code a similar game but change the colour of the grass from light green to medium light green.
- If you make it reproduce copyrighted media, it is a problem.
  As long as the stuff it generates doesn't resemble any copyrighted works, even if it was trained on copyrighted works, I don't see why that should be problem.
  
  I don't even think there's a problem recreating it, you just can't distribute it.
  For personal use it's fine.
  Its not like Disney is suing everyone drawing micky mouse in their personal art workbook
- What media is an LLM going to be able to reproduce that I can't already reproduce with a copy paste?
  
  That's not the point. If you rip a dvd, you babe the movie, but you can't sell DVDs with the movie, because it is copyrighted. After the "AI" has recreated it, the copyright is gone, so you can sell that version with impunity.
- Copyright infringement is about the act of reproduction, not the tools used to reproduce it. The court effectively said the LLM itself is not illegal just like a photocopier or CD/DVD burner is not illegal. It's illegal if someone used an LLM, or photocopier, to make an unauthorized copy of a protected work though.
- It will go to a judge and the judge will say that changing three pixels doesn't make it derivative. Regardless of the method of transformation, the same fair use and parody laws apply.

The japanese artists will be so happy /s

Correct ruling.
- Why can't I pirate stuff without risking jail like AI companies?
  
  If you read a book you can talk about it, quote it, draw characters from that book, write your own ending, etc.
  Isn't that kind of the same? Let's say some day we have an AI with near human intelligence, why can't the AI be trained on copyright works, just like humans, all our school books are copyrighted works?
  
  Why do you think you can't? What do you think was illegal?

From the "source" (the Japanese one, not the broken link at the bottom):
AIによる解析・学習についての現段階の日本の著作権法上の見解を確認、権利侵害の懸念も政府に訴えた一方、生成・出力段階での出力自体の著作権の扱いや元データの著作権の扱いが未確認ですので今後質疑等で確認する予定です。
生成系AIの活用、著作権者を守るための新たな規制が必要だ
It's an analysis of the current copyright law in Japan. It does not mean that they won't update the law eventually. What a terrible article.
- Yeah i wish I had run the source article through Google Translate first.
  https://programming.dev/comment/6159535

66 comments