Copyright and why it's broken. - Tom Scott
So recently there has been a lot of debate on AI-generated art and its copyright. I've read a lot of comments recently that made me think of this video and I want to highly encourage everyone to watch it, maybe even watch it again if you already viewed it. Watch it specifically with the question "If an AI did it, would it change anything?"
Right now, AI-generated works aren't copyrightable. https://www.artnews.com/art-news/news/ai-generator-art-text-us-copyright-policy-1234661683/ This means you can not copyright the works produced by AI.
I work in games so this is more seemingly relevant to me than maybe it is to you. https://techcrunch.com/2023/07/03/valve-responds-to-claims-it-has-banned-ai-generated-games-from-steam/ Steam has outright said, earlier this month, that it will not publish games on its platform without understanding if the training data has been of images that aren't public domain.
So right now, common AI is producing works that are potentially copyright-infringing works and are unable to be copyrighted themselves.
So with this information, should copyright exist, and if not, how do you encourage artists and scientists to produce works if they no longer can make a living off of it?
A lot of people misunderstand the purpose and implementation of copyright.
To make things even more confusing, most people only understand the US version of copyright. The US has worked really hard to get the rest if the world to align their copyright law with the US, but they’re all subtly, and sometimes not so subtly, different, even before you get to the different legal systems used to interpret and implement copyright law.
The first thing is: copyright is about the right of a human creator to control how a copy of their work is made available (publicly in some countries, any in others) for a limited time. In some places that right can be temporarily or permanently assigned or sold to someone else.
So when it comes to AI, one group argues that unless the rights holder licenses the work to be part of a training set, it’s not allowed to be part of a training set.
Another group says that it doesn’t matter; if the human creating the training set got the work legitimately, they can use it to train a model.
The models themselves do not contain any copyrighted material and so are outside the discussion.
Copyright is flexible; for music, we have different copyright rules for recording, reproducing, and performing music for personal or public consumption, with different copyright for music and lyrics and associated video.
This could also be applied to AI, with training models being a new “recipient” of copyrighted content. We could enshrine in law how this can be done legally. But so far, we haven’t, so it falls back to “is this fair use?” and “are different copyright permissions needed when a human is not consuming the work?”
The end bit, that AI created works aren’t copyrightable, is already settled. However, any work a human does to tweak or select AI generated content, if it is itself creative, is copyrightable.
So yes, copyright should exist. People need to re-learn what it actually is though, and additions may be required to the laws in order to enable AI generated art to promote valuable skills and knowledge.
100%.
It gets tricky, though. For example, I'm using a website called "Sudowrite" to help me write a novel. I've been kicking this idea around since 2007. I have a general idea for what it should look like, but I always struggle with Act 2.
Literally over a decade's worth of notes. And not a good Act 2.
But I was able to use ChatGPT and Sudowrite (especially its "Story Engine" tool) to finally understand what Act 2 was missing. And now I'm able to rewrite what I've already done, making it better. AI is a tool just like a word processor is a tool.
Lest anyone think I'm writing an ad here - I'm not. Per their FAQ, Sudowrite says that I own the copyright on anything that I generate with their stuff.
But if I don't modify it, that's clearly not true (as you mention). Furthermore, I can actually have it suggest things that might run counter to that idea.
I've had it suggest lines from Kafka - good lines, too. I've read Kafka, so I recognized them... but what if I didn't? I don't own the copyright on those lines, as Tom Scott points out in OP's video. Kafka's original German is public domain... most translations are not.
You can highlight some text in the tool and say "Write this in the style of Douglas Adams." It knows who Douglas Adams is. It knows what his work sounds like. And the only way it knows is because its model was trained on his work. When I did this, one of the suggestions included Zaphod Beeblebrox, which was certainly not mentioned in my text. It also suggests spaceships and aliens and futuristic gadgets, all written in the kind of prose that you'd expect from Hitchhiker's Guide to the Galaxy.
How would it know that, if it hadn't read Hitchhiker's Guide to the Galaxy?
It's why Sarah Silverman is suing OpenAI. While the model may be a bunch of statistics, it also must know what her text is like - to some degree. We can argue over how, but going back to the AI suggesting Zaphod Beeblebrox... if I didn't know HGTTG maybe I'd think that's a cool name for a character? How can Sudowrite say I own the copyright when it's clear that they don't own it, either?
Which sort of brings me back to the beginning. AI has the potential to be a wonderful tool - again, like going from a typewriter to a computer. I have had this idea for literally 16 years now, and Sudowrite was literally a game changer. I knew all of act 1, act 2 was... ehhhh, and then act 3 was never satisfying without a good act 2. I knew where I wanted to go, but not how to get there. AI really helped, because it understands story structures - and how to make good stories (with some prodding - it's not perfect). And now, whenever I'm stumped, I can type some stuff into the prompt and it'll generate ideas for me.
But that only works if we really figure out where the line is for copyright. I'm trusting what Sudowrite is telling me... but I'm taking a risk, because what if they're wrong?
Keep in mind that you cannot copyright ideas. You copyright the specific implementation of those ideas. So simply mentioning Zaphod Beeblebrox doesn't trigger anything to do with copyright, but if you had a word for word introduction of Zaphod Beeblebrox copied straight from Hitchhiker's Guide that would be a problem.
Then after that you can normally have "de minimus" use, where you can use a small portion (think quotes, short phrases etc). As long as your work can't be used as a replacement for the original then you are normally in the clear (legally speaking, as far as copyright goes).
But none of this goes into trademark law, but that's a whole other thing and are generally easier to avoid too (e.g. don't use the word "coke" to sell your drink).
TL;DR: As long as the AI doesnt spit out a verbatim copy of the original, and then if you copied it verbatim from the AI, then you'd be in trouble.
I am not a lawyer, just a dude interested in copyright law.
I don’t know about US law, but in Europe you certainly can, and it’s an issue over and over again ending up in courts.
Yes it does. Fanfiction e.g. is considered infringement of the creator’s right, and that doesn’t extend to the exact verbatim text but to general plots, names, etc. It’s even infringement if you write a story about „Härrie Pötter“, since it’s immediately obvious that it’s based on Harry Potter.
Some years ago a German discounter sold a costume that was an obvious reference to the TV depiction „Pipi Langstrumpf“, a famous character by Swedish author Astrid Lindgren. Mind you, in essence it was just a really cheap wig and a dress somewhat resembling a tv character. The company owning the rights on the character sued and won.
Edit: oh! I missed the part where there were two courts that decided it was infringement but in the end the highest court overruled that and said it’s not: https://www.lto.de/recht/hintergruende/h/bgh-urteil-pippi-langstrumpf-romanfigur-urheberrecht/
Which only proves that all of it is completely arbitrary and just a matter of opinion. /Edit
In another case, someone took a photo of a soldier, cut out the soldier, turned it into an outline, and printed and sold t-shirts of that. If you took the shirt and put the photo next to it, it was immediately obvious it was based on the photo. Here the court had no issues, because in their opinion it was too far away from the original work to be compared with it. 🤷♀️
So, it’s quite impossible to draw a line between an idea and an implementation, and that’s why thousands and thousands of infringement cases are ending in front of courts, and in the end the only relevant factor is the opinion of the court.
It's a bit more complicated than that, as evidenced by the fact that J. K. Rowling was not the first person to write a story about a magic user named Harry Potter.
AI created works are copyrightable and guidance from the U.S. Copyright Office isn't law, so it's also not settled. Guidance reflects only the office’s interpretation based on its experience, it isn't binding in the courts or other parties. Guidance from the office is not a substitute for legal advice, and it does not create any rights or obligations for anyone.
I read all that in Tom Scott's voice.
That’s good; I wrote it in his voice too!