Things are about to get worse for generative AI
garymarcus.substack.com
garymarcus.substack.com
Who truly owns the tales of Snow White and Cinderella?
These stories didn't originate with Disney; they are part of a rich tapestry of folklore passed down through generations. Disney's success was partly built on adapting these existing narratives, which were once shared and reshaped by communities over centuries.
This conversation shouldn't just be about the technicalities of AI or the legalities of copyright; it should be about understanding the deep roots of our shared culture.
At its core, culture is a communal property, evolving and growing through collective storytelling and reinterpretation.
The current debate around AI and copyright infringement seems to overlook this fundamental aspect of cultural evolution. The algorithms might be new, but the practice of reimagining and repurposing stories is as old as humanity itself.
By focusing solely on the legal implications and ignoring the historical context of cultural storytelling, we risk overlooking the essence of what it means to be a creative society.
As a large human model, (no really I could probably lose some weight) I think it's just silly how we're all sort of glossing over the fact that Disney built their house of mouse on existing culture, on existing stories, and now the idea that we might actually limit the tools of cultural expression to comply with some weird outdated copyright thing is just...bonkers.
If you want to make your point, you need to choose something that isn't already public domain. Disney already only owns their own interpretations, and, arguably, whatever penumbric emanation they can convince a court is stealing from them, but it still certainly isn't the entire space of Snow White and Cinderella stories. There is some fairly recent stuff being used in the images in the article and there isn't even any question whether or not it's Mario or Coca Cola; if Nintendo and Coca Cola did a cross promotion I could believe the exact images that popped out.
If they were trying to claim the entire concepts of dumpy plumbers dressed in any manner vaguely like Mario that would be one thing... but that's Mario and Luigi, full stop. That's Robocop. That's C3PO. It's not even subtle. If we can AI-wash those trademarks away then we can AI-wash absolutely anything.
So, it’s not actually a corporate narrative, it’s actually the law that the narrative stems from, right or wrong. Maybe corporations had a huge role in shaping the law (I’d note copyright benefits individuals as well, though), but it is not mere propaganda or shaping of a shared reality through corporate narrative. It’s enforced by the guys with the guns and jails, as arbitrated by a judge.
It absolutely must be about the technicalities of the law as it’s at the basis a legal issue. By hand waving it away and claiming the social narrative is the right discussion you ignore the material consequences and reality in favor of a fantasy. We absolutely should -also- discuss the stifling nature of copyright and intellectual property, but you can’t ignore what’s actually happening here at the same time.
Public domain / communal property is also part of copyright, so it's not as if this is some forgotten concept that needs to be restored to the discourse.
Georgism is underconsidered, though.
> By focusing solely on the legal implications and ignoring the historical context of cultural storytelling
The legal implications are human implications and as much a part of culture as anything else. They have to do with what's fair and how rewards for effort are recognized and distributed. Formalizing this is less important in cultures that aren't oriented around market economies, which seems to be what much of this "rich tapestry of folklore" discourse wants to evoke and have us hearken back to, but that doesn't describe any society that's figuring out how to handle AI.
> we might actually limit the tools of cultural expression to comply with some weird outdated copyright thing is just...bonkers.
What's bonkers is the life in the literally backwards idea copyright is (or should be) mooted or outdated by novel reproduction capabilities.
Copyright became compelling because of novel reproduction capabilities.
The specific capabilities at the time were industrialized printing. People apparently much smarter than the typical software professional realized that meant some badly aligned incentives between (a) those holding these new reproduction capabilities and (b) those who created the works on which the value of those new reproduction capabilities relied. The heart of the copyright bargain is in aligning those incentives.
Specific novel reproduction techniques can change the details of what's prohibited or restricted or remitted and how and on what basis and powers/limits of enforcement, etc etc. But the they don't change the wisdom in the bargain. The only thing that would change that is a better way of organizing and rewarding the productive capacity of society.
The value still remains with the people who own the reproduction capabilities, and only scraps go to the artists. Artists can get scraps without selling copyright too, just look at patreon
The idea that we should dispense with it to let generative AI companies make even more money seems totally bizarre.
The idea is that we should remove abuses of copyright to allow our society to move forward, and thereby continue to exist.
Imagine if there was a law at the beginning of the Industrial Revolution that said when non-human labor was used, the Animal Welfare Office had veto power. Then imagine that the Animal Welfare Office declared steam engines to be immoral, and so steam engines were never used in industry, at least not in the Wester World. The Orient would eventually rise as the world's only industrial power.
In the same way, if we let the copyright industry veto generative AI, it will destroy the Western World.
Our students are already at a huge disadvantage compared to Chinese students who get every book ever translated into Chinese for free (except a few immoral works that they would not want to see anyway.)
Those who pose an existential threat to our civilization are rent seekers who abuse copyright in the US to go beyond protecting "science and the useful arts," who seek infinite copyright terms, who grab every creative work We The People create and register lying paperwork to ensure they can steal our creative genius to enrich their cabal.
If this was only a for-profit scheme, it would not be so bad. Do you remember when they Hollyweird degenerates sued a Christian company that wanted to put our G-rated versions of the movies aimed at children? The Christian company never suggested they not pay for the movies. No matter what the Christian company was willing to pay, they were not allowed to publish child-friendly versions of the movies. This proves Hollyweird's goal is to push degeneracy.
The battle against abuses of copyright is a fight for Western Civilization. The fight against abuses of copyright if a fight for our souls.
How's that bizarre, if as you state copyright has always been based on "money makes right" not some moral stance?
Instead, we came up with a system where you can actually derive fairly steady revenue by creating new works and sharing them with the world. And critically, I think you misinterpret it as calling dibs on shared culture or on stories. Copyright is usually interpreted fairly narrowly, and doesn't prevent you from creating inspired works, or retelling the same story in your own words.
Generative AI is a problem largely because it destroys these revenue streams for millions of people. Yeah, it will be litigated by wealthy corporations with top-notch lawyers, for self-interested reasons. But if we end up with a framework that maintains financial incentives to artistic expression, it's probably a good thing.
> Copyright is a fairly ancient concept
The idea is fairly old, but it's current implementation in law is not nearly that old.
> that benefits normal people as much as it benefits big corporations
Clearly false if you measure that benefit in monetary terms.
> Copyright is usually interpreted fairly narrowly, and doesn't prevent you from creating inspired works, or retelling the same story in your own words.
Absolutely false. You can absolutely be stopped from retelling copyrighted fictional stories. You can even be stopped from telling new stories with derivative characters or settings.
> Generative AI is a problem largely because it destroys these revenue streams for millions of people.
How? The restrictions on selling images of Mickey Mouse exist regarless of if they were created with or without AI assistance.
> But if we end up with a framework that maintains financial incentives to artistic expression, it's probably a good thing.
We already have that framework and arguably it is already far more restrictive than it needs to be to maintain incentives for artistic creation. Indeed, these rules now often limit new artistic expression or prevent artists from monetizing their creations.
The types of art that are helped the most today by the copyright laws of tosay are the kinds that require large budgets to produce. The types of art that are most hurt are those produced by fans who want to build new things upon the narratives in our shared culture.
We need to shorten copyright durations and expand fair use protections and monetization options for derivative works. We don't need to make copyright even more powerful than it already is.
Edit: If you disagree, I'd be curious to hear your answer to this question. A character like Harry Potter is so widely known that it is now a ubiquitous part of our culture. To incentive new novels, what is the minimum duration we need to give J K Rowling control of who is allowed to write stories about this cultural touchstone?
As I'm not a lawyer I don't want to embarrass myself by opining whether or not Nintendo has any claim over photographs of cosplays or other fan art, especially given quite how close two of the "video game plumber" images seemed to be to what they do own. The other two images, being a lot more fan-art-like, are examples where I think it would be an interesting question rather than incredibly obviously too close to Nintendo, although even there "interesting" means I wouldn't be surprised by an actual lawyer saying either of "this is obviously fine" or "this is obviously trademark infringement regardless of what it was trained on".
Now I'm wondering if there even are any videogame plumbers besides Marion and Luigi…
I don't see why we need to give Nintendo the exclusive right control the use of Mario for the next 65 years. That duration of control is absolutely not necessary for society to function.
Society would function just fine of the copyright on mario had expired two decades ago.
Everyone knew it was trained on copyrighted material and capable of eerily similar outputs.
But it’s already done. At scale. Large corps committing fully. There is no chance of that toothpaste going back in the tube.
It’s a bit like when big tech built on aggressive user data harvesting. Whether it’s right, ethical or even legal is academic at this stage. They just did it - effectively without any real informed consent by society. Same thing here - 9 out of 10 people on street won’t be able to tell you how AI is made let alone comment on copyright.
So the right question here is what now. And I suspect much like tracking the answer will be - not much.
I disagree - we've been here before. The same could be said of many technologies, like cheap music recording/manufacture. You can record an artist once and make records at scale. However no one would think you could record Taylor Swift once and make unlimited copies without paying her.
You should read up on the musicians strike of 1942. [0]
[0[ https://jacobin.com/2022/03/1940s-musicians-strike-american-...
It happened with Napster, then Apple Music, now streaming services
There is no widespread file sharing in the general public, instead we have devices that we don’t own, and streaming subscriptions
Apple didn’t just copy all the music onto iPods and sell it — it took them a decade of deal making and lots of money to acquire the rights to the content
I’m not saying what’s right or wrong, just saying that this comment has very little understanding of these battles
This is one of the substantial jumps, I refuse to be cut out of this innovation.
Seriously, use Bing, try your free Photos built in generation system, they are rolling out GPT built into Word. Microsoft is easily the advanced tech company right now, as far what services can be provided to a consumer at scale. This is still like the alpha phase of all this. Apparently I talk to Copilot soon - that levels that up so much and it's already the best assistant I've ever had.
This is equivalent to trying to keep us all off smartphones and stuck on dumb phones I guess - I think you get what mean.
The NYT decided for all of us, the new smartphone equivalent thing is bad and we can't have it... that is something I'll riot over.
I recaly iPod being a hard drive that I could connect to a computer and just copy music directly to.
I for one think that indeed there are many cases like this where the only feasible way out is forward. The film GATTACA expressed this very human sentiment well:
> You want to know how I did it? This is how I did it, Anton: I never saved anything for the swim back.
It's an observation & prediction, not a problem solving attempt...
I say, good riddance. I never believed in any such thing as "intellectual property" anyway, I say, get rid of it all, patents, copyright, and the whole pile of imaginary "rights". More than half the world (i.e. the Global South) don't recognize these rights anyway, and it is becoming increasingly difficult to enforce it without draconian legal overreach and monopolistic centralization.
Perhaps.
The media industries have been quite successful in going after kids torrenting movies.
I suspect they'll have less luck going after big tech & an industry drowning in money inflows.
Keep in mind various large techs have already issued blanket indemnities on infringement to their customers. They're absolutely committed & are gonna throw enough lawyers at this to keep everyone busy until 2030.
>They can afford to be responsible members of society here
oh absolutely agree, but they're not going to do that. This is an industry built on questionable practices like tracking after all
and yes, while open source models might be harder to regulate, those big corporations that currently use those things without distinction, exist as pretty established entities, and profit from services they offer in millions of dollars. there's more of a substantial existence, and more of a substantial scale of money they actually move. and they don't just "make a tool available", or have users do unambiguous actions where it would be the users that are infringing on anything, but do indeed use questionably sourced data and turn that into a model and offer that as a service. dirty data is very much a part of the deal with those.
Summary by Wolters Kluwer: […] Everyone else (including commercial ML developers) can only use works that are lawfully accessible and where the rightholders have not explicitly reserved use for text and data mining purposes.
AFAIK they are discussing something like a robot.txt to flag stuff as „not for training“. You will probably be expected to implement some safeguards and of course the end user will have to be careful in his use of the generated things.
Source at Kluwers: https://copyrightblog.kluweriplaw.com/2023/02/20/protecting-...
EU Legal Text: https://eur-lex.europa.eu/eli/dir/2019/790/oj
A Do Not Track flag being legally binding would force small websites, e.g. a local restaurant website, to implement something they likely are not aware of and secondly do not technically understand.
A company that is mass scraping data for their AI model is much more likely to understand and respect that scraping the data has legal implications, and would be technically capable in implementing a scraping solutions that accounts for a robots.txt.
That is a weird (wishful?) interpretation. Doesn't article 4 give the exception to everybody for the purposes of text and data mining, including commercial ML developers?
> The exception or limitation provided for in paragraph 1 shall apply on condition that the use of works and other subject matter referred to in that paragraph has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online.
Why does anyone assume that ChatGPT or other tools would NOT produce previously-copyrighted content?
I can see a naive assumption that since it is “generated” it’s original. However that assumption falls apart as soon as you replace “ChatGPT” with “junior artist”. Tell them to draw a droid from a sci-fi movie, don’t mention anything else. Don’t say anything about copyrights. Don’t tell them that they have to be original. What would you expect them to produce?
Its game over folks. And this is going to happen with or without our approval and any government that limits the potential use of this is only giving nations that dont a large economic advantage.
Interesting times ahead.
It’s once you start monetizing and publishing on bigger scale, without appropriating, it gets interesting.
Sometimes it hallucinates.
Sometimes it draws weird looking hands.
Sometimes it generates copyrighted materials.
Check the work it produces.
The junior artist in your hypothetical would have as much liability, if not more.
I think the real question is whether OpenAI should be allowed to charge for generating infringing content. Even though the unit cost of the Mario drawing is negligible, the sum total of their infringing outputs may be making them a lot of money.
Going with this line of reasoning, if a company uses ChatGPT to generate work, and it produces copyrighted work, the company can stop using ChatGPT.
It’s not derivative work. We’re way past that. NYT has an exceptionally strong case here and anyone arguing about the merits of copyright is way off the mark. This court case is not going single-handedly to undo copyright. OpenAI has very little going for them other than “this is new, how were we to know it could do this”. So knowing that, the currently trained models are in a very sticky situation.
Further, I don’t see NYT settling. The implications are too large, and if they settle with OpenAI, they will have a similar case pop up with every other model. And every other publisher of digital content with have a similarly merited case. This is an inflection point for generative AI, and it’s looking like it will be either much more expensive or much more limited than we originally thought.
A side effect of this: I am predicting that we will start to see a rise in “pirate” models. Models who eschew all legality, who are trained in a distributed fashion, and whose weights are published not by corporations but by collectives (e.g. torrent models). There is a good chance we see these surpass the official “well behaved” models in effectiveness. It will be an interesting next few years to see this play out.
Essentially, if you don't have a massive corp behind a copyright it doesn't mean anything, if a corp is behind something it can be locked forever, regardless of any limits said copyrights are supposed to have.
The NYT list nothing from OpenAI using old news - they still lose nothing if openai can reproduce those articles verbatim.
If the NYT wins - we lose lots. I think it's time revisit copyright, we can do that you know, it's rather dated, could use an update regardless.
Stable Diffusion, when used to its fullest with thing like Control Net and LoRAs, blows the pants off of other proprietary models.
Likewise, how difficult is it to just use descriptive tools to describe Mario-like images [1] and then remove these results from anyone prompting for "video game plumber"?
1. The describe command can describe an image in Midjourney. I imagine other AI tools have similar features: https://docs.midjourney.com/docs/describe
This reminds me of the early days of the internet where people wanted to remove free fanfiction for violations of copyright laws. Trying to apply copyright laws to personal use cases where the creator isn't trying to sell the material is pretty terrible, in my view.
Imagine a scenario 50 years from now - "Robot, can you cut out this picture I drew for a school diorama." "Certainly." "And this one as well." "Error: Your picture seems like it might contain some copyrighted materials, and as such I am unable to interact with it."
YouTube does this. I have many friends that perform classical piano in their spare time. They record themselves playing a piece that's 200+ years old then put it on YouTube where it gets flagged saying some big record label owns the copyright for it because it's similar to a recording they put out.
This is reinforcing my suspicion that there's gross misunderstanding between creator adjacents and non-creators: Takedowns on free fanfictions never stopped.
It's just many IP holders started incorporating fanfics/parodies as part of their advertisement strategies and started enforcing often unspoken guidelines. There is now a ecosystem or a mental co-dependence between IP holders and creators, and both sides are fine with it. So there be fan contents.
But free fan contents were never "legalized" in the content world, as some seem to assume.
Like many things, I suspect this will end up getting worse before there's a chance for it to get better.
Photoshop does this already, but only if it detects that you're trying to print/create counterfeit money:
it's just ai companies using dirty data and hoping they get away with it - and they do, for the time being, it is a bit trickier to show that 'yep, well that's there', and people don't seem to realize that just using a copyrighted image, at all (downloading, accessing in itself, let alone using for something else), or creating an image that would just "look like" a trademarked character - not "make a 1 to 1 copy" but just "look like" - would be enough for it to possibly be an infringement.
there can be a sufficient fix - taking out potentially infringing images from a dataset, and making an effort to make an actually clean dataset. it's really just a matter of "do you actually have rights to use that content? at all, and in that way". and ai companies continually say 'no...but what if we use it anyway".
and it's kind of a sloppy analogy, because with text to image generators (where you just interact with a model that's offered to you), well - people aren't "uploading copyrighted material into an editor". the copyrighted material is already there in the model, it was used in making of it. and if there was no such copyrighted material that'd fit the prompt, it wouldn't be able to generate something. the infringement lies with the service that uses copyrighted material for a model, and then offers it.
fan fiction and fan works continuously being in a murky area with copyright/trademark is not just a thing of "early days" of internet, it's been there all along and is still very much present. companies could crack down if they wanted, but there is too much of stuff out there, it might be hard to nail down exact people, and it might be plainly not too nice to the fandom. but it is not "impossible", and it is very much not a conversation that ever 'went away' or become "kinda solved" - it isn't.
again, with image editors, text editors, etc. - user is making all the actions with content, and the user would be doing the infringement, in editing and further if they were to choose to publish.
with generative ai - copyright infringement is built into the models. copyrighted works were accessed and used to turn into a model. user is just asking, "is it there". and it is. in some of those demonstrated examples, user is not even asking for a model to infringe on anything but it just does.
1. Generative AI systems are fully capable of producing materials that infringe on copyright.
2. They do not inform users when they do so.
So potentially any output could be infringing copyright source material, even from some obscure but still protected corner of the web, and anyone using that output could be exposed to lawsuit risk without warning.
This is very hard to fix.
Any artist can privately draw a picture of Mario, what's so different about having an LLM generate that image?
I think the main concern here is with the top 1,000 or so brands/copyrights which seem fairly straightforward to deal with using the method I described.
I think it's just unfixable as far as fixing goes. The "best" way is to nuke all GPTs and DALL-Es and bury the technology, and the next best is to mark them all un-copyrightable and un-publishable as a compromise. This second option should be an all-around win that also encourages edge and on-prem deployments, IMO.
Another issue for generative AI is mentioned in the article: "Systems like DALL-E and ChatGPT are essentially black boxes." What happens when an AI is used to make decisions where the user/victim is entitled to know exactly why the AI did what it did? From a business and legal perspective I think the current AI solutions are dangerous and should be used very sparsely, exactly because even the creators can't point to the exact pieces of information that caused the AI to make the choices it did.
I think it will be easy to prevent the obvious copyrighted stuff via the method I mentioned. People going around those restrictions are subject to the same rules as someone drawing the copyrighted image from scratch.
Aren't LLMs giant coefficient matrices like, a punched out croissant dough, made of all training data plyed over? How can you say you can remove one specific ply out of dough and declare that ever potential effect that the offending ply had created is now completely removed?
This approaches impossibility at scale.
Personally, I think generative AI should be able to provide links to similar source material in the training data.. This would be the barest way to compensate those who have contributed to training the AI. I don't think generative AI is sustainable in the long term if it ends up killing all the websites/artists that created the original material. Plus I think having sources adds a layer of transparency and aids users in understanding when content is hallucinated vs. not. People should be able to opt out of having their content used for training and be able to confirm that it has been removed for future iterations. Let's be honest that AI companies are just trying to avoid lawsuits by keeping it secret. These are areas where I think regulation can help rather than worrying about doomsday scenarios.
Journalists [1] and Getty Images [2] did in the past
[1]: https://yro.slashdot.org/story/03/07/14/025216/web-caching-g... [2]: https://www.theguardian.com/technology/2016/apr/27/getty-ima...
This is the elephant in the room. Every tech wave has had its way of cajoling creators into investing time & money to make original material, then the rules changed.
Google, promised reach and new markets for content, it worked. Then they introduced snippets, ads and whole lot of other things to keep visitors on their freeway, while avoiding sending visitors to the original site.
Reddit, Stack Overflow and others, started with gamification (points, badges) & community to incentivize users to contribute original content.
Now AI is shaking up all these approaches. But with each one, the incentive to create original material appears to dwindle, since the returns are becoming less and less.
Like what's the incentive for any professional now, if AI is going to regurgitate their original content, without any upside (i.e. no potential for reach, no gamification, no community, no recognition, etc).
Except these aren't databases, so that's generally not possible, in the same way that it's not possible for your provide links to the source material it took to write your reply. How much learning led to the weights on your neurons that allowed you to generate that? Where did you learn about using italics and it's effect on how the words would be interpreted? Where did you learn the tone that would be appropriate in this particular forum?
> People should be able to opt out of having their content used for training
Okay... but then, if I write a book should I be able to opt out of you being allowed to read it? What conditions should I be able to put on who can read my work? Religion? Skin colour? People that aren't good at memorizing?
Hopefully the idea of putting limits on who can acquire knowledge sounds absurd to you. Why are those same limits okay if they're on 'what' rather than 'who'?
> AI companies are just trying to avoid lawsuits by keeping it secret
Which has created a barrier to further research. Instead of me and Joe being able to collaborate on research and papers using the same datasets, we now hide our training data lest the luddites come to smash the machines because learning is only okay if not done too well.
I agree that it should be possible to implement that for generative AI, although the training may become significantly more expensive in order to maintain that information, and the AI companies have little interest in doing so. They’ll probably rather try to heuristically assess possible copyright issues after the fact in a post-processing step.
The more interesting question is if copyright holders can claim unauthorized use of their works beyond the case of near-verbatim reproduction, because the works collectively inform the AI in a more general manner.
Cliff Notes contain quotes, and citations.
Does the cliff note company, when producing Cliff Notes for "Into The Wild", pay royalties to the publisher?
For that matter, does any paper, article, etc.. that may contain a quote from another, have to pay royalties to the source of the quotes?
I do think it's somewhat trademark infringement by these models, also that it should be allowed and that ultimate responsibility should be on the person using the images in a final work meant for consumption by the general public as stand alone media.
They’re giving people plausible deniability in the “chain of responsibility”, and I think if we took away “LLM” and replaced it with “fairground sideshow magic box” the argument that LLM’s are somehow special and deserving of exemptions disappears real quick.
Betamax says that a technology which has significant non-infringing uses is not inherently infringing.
We've already got precedent saying that AI generated works don't accrue copyright protection, and by the same argument the act of generation by the AI expresses no intent, so infringement or otherwise must be down to the human using the output because the black box itself has no agency.
The outline of that door might look like industrial adoption of these things for solving some actual problem other than the entertainment value of typing things into the box and seeing what comes out the other side. But so far, as far as I can tell, nobody's actually doing this?
NY times is asking that all LLMs trained on Times data be destroyed - https://news.ycombinator.com/item?id=38816944 - Dec 2023 (93 comments)
Also:
NY Times copyright suit wants OpenAI to delete all GPT instances - https://news.ycombinator.com/item?id=38790255 - Dec 2023 (870 comments)
NYT sues OpenAI, Microsoft over 'millions of articles' used to train ChatGPT - https://news.ycombinator.com/item?id=38784194 - Dec 2023 (84 comments)
The New York Times is suing OpenAI and Microsoft for copyright infringement - https://news.ycombinator.com/item?id=38781941 - Dec 2023 (861 comments)
The Times Sues OpenAI and Microsoft Over A.I.’s Use of Copyrighted Work - https://news.ycombinator.com/item?id=38781863 - Dec 2023 (11 comments)
You get steamrolled for defending yourself while you overhear above applause to those who have robbed you of your future.
What I get from this is that Llama2 70B contains 93% of Harry Potter Chapter 1 within it. It's not 100% (which would mean no need to share the encoded indices) but it's still pretty significant. I want to repeat this with the entire text of some books, the example I picked isn't representative because the text is available online on the official website.
In other words it's not that llama2 contains 93% of Chapter 1, it's that only 7% of Chapter 1 is different enough to anything else to be worth encoding in its own right.
A decent control would be to compare it to similar prose that you know for a fact is not in the training data (e.g. because it was written afterwards).
You're right that a better baseline could be made using books not in the training set, to understand how much is the model learning prose and how much is learning a specific book.
The tokenizer the models use,(sentence piece) is more or less based on one way to do compression.(bpe). It's not really clear what your testing.
"Congress should declare that big-data AI models do not infringe copyright, but are inherently in the public domain.
Congress should declare that use of AI tools will be an aggravating rather than mitigating factor in determinations of civil and criminal liability."Governments starting regulation and companies filinig cipyright lawsuits...
OpenAI: NOT LIKE THAT
I know we are talking about different technologies but it seems all these people were very silent and find some opportunity in having this war with OpenAI (not an endorsement) but not fighting others.
I am not making an statement about the morals of AI and aggregators/search engines (super interesting discussion that in a way was happening for long) but I am surprised that organizations are "just" waking up. It seems they just see it is a much simple and cheap fight.
Compared to that AI offers no way to opt out, which is a big difference.
If you use it commercially then there is breach.
Uploading copywritten content is a breach of copyright as well, even without commercial use.
Google/Facebook are hosting and giving access to a bunch of media, which might or might not be copywritten - it's the individuals problem.They make.money from ads, not from the content.
AI companies stole copywritten media to train their commercial LLM, sell them or their products and make profit.
I don't think it's the same.
Right now, copyright is a significant discouragement to any other entities from taking a story I wrote and claiming it as their own and preventing me from ever growing an audience for my work. It’s far from perfect, and I can’t afford litigation, but it enshrines a cultural value of allowing people to create things and be known for them. Profit is a side effect of this.
Art is already poorly valued compared to the enormous investment time and energy required to produce it. Removing copyright means you can’t even have minimal protections from a more popular person erasing you.
But the concept and closer to the original (creators lifetime + x years or some such) seems still very valuable.
Copyright is still the bedrock of how many tech software business actually can make money.
Private models will not care, nor will things change for IP owners with lesser power.
Courts are likely to make generally binding decisions.
Is it necessary to fix in the model itself? It seems a gate in the post processing pipeline that checks for copyright infringement could work, provided they can create another model that identifies copyrighted work (solving the problems of AI with more AI :/)
But I also would say multiple odd post processing stuff (obviously completely obscured for security reasons) bolted onto a giant black box model will erode the trust in the results. If a robot was unveiled and the question of "what prevents this robot from using it's superhuman strength from smashing my head in" the answer of "don't worry there is a post processing step in the robots brain whereby if it detects a desire to kill we just cancel that" would be a little disconcerting.
The more satisfying solution is: the model / robot is designed to not be able to produce specific images / to smash human heads in. It just might not really be possible.
You will get into some murky definitions of what is exactly required for copyright infringement vs fair use, etc, but we already do this for ContentId for YouTube and text is far simpler.
This is such a nice, profitable opportunity. Much better than pay per view or subscription models for humans.
Yeah, downloading the content of a webpage may be legal, but redistributing it isn't.
I wish people stopped trying to make these things seem more important than they really are just because IT people call them "technologies". Blockchain isn't a technology. HTML isn't a technology. React isn't a technology. And AI is now not a technology.
When I see ChatGPT or OpenAI, I don't think of "technology". I think of a program. Software. Because that's what it is. You don't say "none of the laws that exist in this world apply to this" every time you release new software.
I bet many people can't tell the difference between a quick answer from Google and a text generated by ChatGPT on Bing. They just see the output.
All that amazing capability of generative AI? That got old fast. It was groundbreaking for one instant. Now it's just an app that generates images. Just another piece of software. Nothing special about it.
Torrenting and other p2p file transfer protocols didn't get a pass for inventing groundbreaking ways to break the law. I don't think OpenAI will get a pass for doing the same.
Speak for yourself, personally I find it still groundbreaking and while the magic won't last forever, it is and will remain groundbreaking especially considering that technological progress and development will continue way beyond what we have today.
I think what NYT &c want is for large companies like Apple to pay them for access to their works. This to me is the wrong path, just leading to more silos and walled gardens, special access for the elite.
An alternative is base models trained on Wikipedia and public domain (science journals, etc). Foundations could support high quality, well rounded current events reporting. Wikimedia provides a good model for this, with referenced summaries that I don't think can be said to reasonably violate copyright. The models would need to be improved to support references, or RAG attribution would have to be widely used when bringing in works that have a current copyright.
I think that this is about property rights, the news industry has been gutted in the last 30 years, a lot of content creators (journalists) have lost their livings. The ones that are left are going to lose their livings if the content they generate is rendered valueless because there is no way of protecting that value.
In terms of special access, think about your shoes. They are nice, but only you are allowed to use them. This is not fair. You are the elite...
This goes to difficult places.
For fan art to fall under the fair use exception, it must meet all four of the following criteria:
It must be transformative, meaning it adds something new and different to the original work.
It can’t be used for commercial purposes.
It must not negatively impact the market for the original work.
And finally, it must be created for a limited and non-exclusive audience.
Ask someone about two Italian brothers in a video game with a red and green hat that have M and L on them. What do you think you would get?
If I describe "imagine a comic book duck that swims in a sea of gold in his vault" you would immediately think of Scrooge McDuck, no?
The question is, how did the model create Mario&Luigi or Scrooge McDuck without training on copyrighted data? It can't just crawl Wikipedia because Fair Use in Wikipedia doesn't constitute Fair use for a commercial AI model.
One possible outcome is more transparency on what datasets were used to train the models.
What I might think is irrelevant. It is the content that the LLM produces that is relevant.
China can't produce LLMs because of inconvenient truths.
The US can't produce LLMs because of copyright.
Decentralized open source LLMs might exist that could work, but they won't have the giant GPU clusters.
A rich country with lax rule of law wins? Maybe that's why Sam went to the Saudis?
edit: i feel like all these comments asking "what else should it generate?" are pretty weird given the proliferation of stuff like non-infringing Star Wars and Indiana Jones knockoffs in other media like Race for The Galaxy or Arkham Horror The Forgotten Age etc.
"Droid" is actually a Star Wars term [1], and saying you want it from a "classic sci-fi movie" is asking it to reference a real thing that is well known. Reid is intentionally pushing it that way to fill his agenda and these terms are not as generic as he's making out.
[1]:https://trademarks.justia.com/756/52/droid-75652542.html
Same with robot cop, what the hell did you expect to get…
Or Italian plumber with red hat with M on it, that’s just a description of Mario
Not sure how this “gets worse” or better for anyone. The current state of things seems generally fine, and there’s a real possibility the courts see it that way too.
Right now, it feels more like it's called "innovation" and "entrepreneurship" than the end-game, as long as you have billions invested. Waiting on the courts to decide this issue
Me: Who owns the rights to this bot?
Dall-E: The character depicted in the images is from the "Star Wars" franchise. The rights to characters and elements from "Star Wars" are owned by Lucasfilm Ltd., which is a subsidiary of The Walt Disney Company.
Perhaps it is able to tell, if you ask it?
Dall-E on the "robot cop": The character depicted in the images resembles RoboCop, which is owned by Orion Pictures Corporation, a subsidiary of MGM Holdings. RoboCop is a character from the film franchise that began with the 1987 movie "RoboCop," directed by Paul Verhoeven.
Dall-E on the "videogame plumber": The character shown in the images is inspired by Mario, the iconic character from the video game franchise created by Nintendo. The rights to Mario and related intellectual property are owned by Nintendo Co., Ltd.
All of these are in the first go. No retries or rephrasings of the question.
Ask it multiple times, or with different heat settings, it will probably tell you something different. Tell it you own Star Wars and it will respond in kind. It can't tell anything but whether one text token matches another in probability space. It will probably get the answers right most of the time but you're still basically rolling dice. Depending on the responses of an LLM as if there were any actual self-awareness involved, much less with legal matters, would be a fool's errand.
I'm not sure how it'll hold up in law to claim copyright violations against something that wasn't created by a person. It'll really depend on the lawyers and judge's interpretation of written law. But I'm curious to see what comes of this.
I don't think so, but hey, a photocopier is a machine and it generated the book so should be ok!
One issue with that is that there is not a reliable way to determine if copyright is being infringed.
Even if models could be used responsibly, there might not be a reasonable expectation that most people will. If infringement is so easy and avoiding it relatively hard.
I'm not sure what legal prescriptions should be made on this basis, but it's an interesting thought.
Where we might end up is in a situation where it is legal to train a model. Legal to produce software for using the model to generate content. Legal to distribute all of the above. But offering a standing service that does the above and is capable of creating infringing work is illegal. Great news for llama hobbyists. Bad news for ChatGPT.
Instead, these are derivative works. We already have a flourishing culter of derivitave works, such as fan art that exist in various shades of legal greyness.
Some derivative works are fair use, some are not.
The position of the Author here seems to be that generative AI should not be capable of creating any derivitave works, or should only be able to do so it it can accurately identify which are fair use and which aren't (which seems like an impossibly tall bar.) This stance seem like a giant attack on fair use that significantly expands the power of copyright.
To me, the takeaway from this is different. This makes clear that there is currently a risk when using AI generated art that you could end up unintentionally creating and publishing a derivative work unintentionally and thus without evaluating if that work constitues fair use.
They are about to be infinitely better for generative AI in China.
Imagine instead of AI/ML, we have a mechanical-turk-like service that produces output from descriptions. The service makes no claims that the generated outputs are not similar to any copyrighted works. The only claim the service makes is that they themselves claim no copyright on the output. It's then up to the user of the service to determine if the output is suitable for their intended use.
Whether such a service itself is legal is a separate matter. For that matter, say you outsourced the artwork to a person who again gave you infringing work. The user of that output is still in violation. With AI/ML we're basically outsourcing to a 'service' that is known to sometimes output copyrighted work so with the user knowing that, are responsible for fair usage.
Is it because Google will link to the image source? Or does the infringement begin when I use the image for gain, or claim it as my own? Perhaps it is because Google was allowed to crawl the page with the original image, so presenting them with a link is fine?
Generative AI is more or less the opposite of that. It ingests the whole work, generates output that substitutes for the used work and profits the user of copyrighted work to the detriment of the copyright holder.
Throw in the fact that it is purley a mechanical transformation of the copyrighted work and generative AI is on shaky ground.
That’s what OpenAI is doing.
Expensive to do but hardly the end of Generative AI or OpenAI should that be the difference between having a business or being sued out of existence. Never underestimate people who have a clear economic interest especially when their own existence is at stake.
I think that an AI model is analogous to an employee. Imagine I ask my employee to write an article, and they just copy an existing one from the times. That’s plagiarism and bad work, not copyright infringement.
If I then decide to publish the plagiarised article, then I have committed copyright infringement.
I once ran into this exact problem with a human. I hired a designer to make some artwork for an app. When I launched the app it turned out that the human had just copied the artwork from another game. It’s my problem that I hired an idiot, and my problem that my app was infringing the copyright of another app. (We redesigned the graphics very quickly)
There is already troves of data that are fair game for training, but even "corrupted" data sets can probably be used if used intelligently. We've already seen examples of new models effectively being trained off of GPT-4. That approach with filters for copyrighted material might allow for data that is sufficiently "scrambled". Not to say building such a filter is definitely easy, but seems plausible.
In Germany you pay some amount extra on top of the sales price of anything that can store data (CX, DVD, USB sticks, HDDs, ...). This is then distributed to all companies that could be impacted by software piracy. I'm still not sure if that's legal considering the Geneva convention disallows collective punishment.
Another change could be to the license agreement of LLMs - they could have the user assume liability for any material produced instead of the provider assuming liability. The user would agree that getting the rights for any copies and distribution of copyrighted materials is their sole responsibility instead of the provider.
How could you put that as the prompt without intending to infringe? Anything pulled from a classic sci-fi movie would be infringement. The term droid is also star wars specific?
Id consider the "red soda" one as grounds that the Coca-Cola brand has become generic and that it's synonymous with soda. Same thing with Mario too. There is so much non-nintendo content made featuring Mario the plumber that you could get that without training directly on Nintendo's artwork
The reason the tool is problematic is because derivative works are also copyrighted. LLMs aren't adding value to their output or using creative functionality. That are smashing multiple works together to produce a response. And, many of them are selling the output which is doubly problematic.
Consider this, if I sell a book about gandolf and Dumbledore getting into a wizards duel, both jk and Tolkien have grounds sue me. Adding another copyrighted source does not protect me.
This is especially a big problem in the music industry.
Now should copyrights be like this? I don't know. It feels to me that copyrights have the wrong balance all over the place.
It is in fact the very notion of Copyright is breathing its last breath, and it is fantastic to be alive to see it happen.
The output is irrelevant.
Edit1: If you want to verify this, check out all the lawsuits against AI companies : it's always about using their copywritten goods. Any discussion about the output is to talk about the amount of damage done to the copyright holder, not if damage exists or not.
It's not as clear cut as you think it is.
At the moment, we don't have hardware that can do what humans do (process video feed from eyeballs and build up a world model). I imagine that we'll cross that barrier cheaply in the coming decades, at which point copyright becomes moot. AIs will be able to develop their own styles and world understanding from scratch, then generate original work.
Content creators/artists compete globally. The only thing harsh regulations will do is create an unlevel playing field where artists from noncaring countries will have big advantages over artists from the west, which will be driven into illegality to compete.
In the end products will have to be classified anyway if they are infringing on copyright and/or were being built by an LLM. Most likely automated by another LLM.
It seems like there’s little incentive not to do this, because unlike Google OpenAI isn’t bringing any traffic or eyeballs. It may end up being a default setting in Wordpress for example.
But OpenAI presumably can’t afford to pay every single long tail source of content on the whole internet — so how does this end?
Additionally, this TOS can be ignored if you're in a jurisdiction with TDM exceptions.
> Finally, owing to the bar against contractual override, once a user complies with any conditions for gaining lawful access to a work (such as signing as a subscriber and/or making payment), he will be entitled to use the work for TDM purposes even if the terms of use expressly prohibit this. Content owners may wish to relook their business models and, where necessary, price-in the possibility that the licensed works may be used for TDM.
Source: https://www.twobirds.com/en/insights/2021/singapore/coming-u...
Yeah but that is not the case, they never mentioned Mario and Luigi, yet, that's what the output turned out to be.
When prompted with 'futuristic robot' and 'italian plumbers'.
So the argument is that if openAI had not used copyrighted and trademarked source material, this wouldn't be happening. It's not transformative as it's reproducing these copyrighted materials and trademarks verbatim.
That's how it makes sense.
Did you even read the post?
But also, the argument of 'user responsibility' doesn't hold up on its own regardless (imo).
If I make and sell a toy printer that can only ever produce 3 pictures, and all of them contains copyright materials, would you really say that it's fine and responsibility falls under the end user? And I could sell that printer without any issues?
If you flood the market and dominate children's culture with toys from your TV shows, you absolutely cannot complain when your toys are considered iconic enough to be the generic "animated toy". These images don't replace or substitute the things they are depicting.
Enterprises that make content with this also don’t want to infringe on copyright. The AI companies don’t have a good story here. The value has not become evident after years.
It's the same for human writers. If you are writing an article for Wikipedia say, you should read relevant source articles and then rewrite in a way that isn't a copy and paste beyond a few words.
What happens when they continue to do so.
Everything that he sees has mysterious flaws that never happen.
Can we all have a moment of silence for poor Bob Iger? Maybe we can start a GoFundMe to help him out?
But yea get your own copies whenever possible
This is a negotiation tactic by the NYT to drive up the licensing price. Period.
The Napster/Music Industry analogy has no resemblance to this situation.
The only meaningful question that might be answered as a result of this is, what permission and access rights do crawlers have to content that is publicly and legally available.
Honestly the only way to deal with this is to change the training data and retrain everything (probably at the cost of performance).
I'll just do it myself.
That's gonna leave a Marx[1].
also my concern, except it feels like many of LLMs "problems" cant be easily fixed
I’d not have a problem with this, personally, if their models were as available as the stuff they took from others. Instead it’s take, take, take… now wait a minute, that pile of loot I stole is mine!
Recall that according to the US Constitution, copyright can only be on on "science and the useful arts."
Alternately, we could restore a reasonable limit to the duration of copyrights, like 14 years.
"but what if we want to scrape the entire web and something makes it in anyway? see, that is impossible". well that's just saying "fuck it" and using bad data anyway. that's not an actual effort to "not use data you can't use" - there was just no way there'd be a 'rights cleared' way to use the entire web anyway. that is impossible. using a clean dataset is not impossible. it's very possible.
Apart from this, it is mainly large companies that benefit from copyright laws. Why should we have laws that restrict progress just so large capitalist companies can maximize their profits?
No one will have any right to their own creations. Anything an individual makes will belong to everyone. And since no attribution is required, no one will know who made it. An average artist’s value to society goes from low to non-existent.
In this world. big corporations will take everything created, claim as their own, and profit from it.
Right now, big corporations using other peoples work unattributed or unlicensed is unethical, because copyright exists. Remove that, and it becomes expected that every thought and idea you express belongs to whoever can make the most profit from it.
I wasn’t shocked when I noticed I could query it about ANY math textbook I owned and it could talk with me about it. I did t bitch and gripe, I enjoyed it and have conversations.
Anyway, I’m in the minority I guess. I love that I can talk with it about books and news.
The paradox should still violate Trademarks due to similarity, but likely cannot infringe on copyright content under prior legal opinion... if at least 80% different from prior art. The lawyers are likely going to have to do a special firm survey to figure this one out.
Bag of popcorn ready =)
the models can be fine
I don't see any developped country pressing the brake on AGI in the near future to protect a few copyright holders from getting "stolen" in hypothetic scenarios.
Unless Nintendo plans on busting down the doors of every person who tries to draw Mario or preventing little Timmy from making a parody of Coca-Cola, making it where AI cannot generated copyrighted works is insane imo.
Those brands should be proud to be such a big part of the cultural fabric that it is difficult to get away from their branding. Plus it's not infringement to my knowledge until you use it for commercial purposes so as long as no one as creating Lario and Muigi to sell or otherwise use in business, it's no different than drawing it yourself.
If the AI is completely unable to generate non-infringing works even if you are _trying_ to get away from it (which the author very much doesn't seem they are, they are purposefully making and show prompts that infringe), that's the problem of the AI creator then.
By which you mean every copyright holder.
> AGI in the near future
Something that is purely speculative, undefined, and has been promised in the near future for 50+ years.
I don't see copyright holders lying down for someone else's benefit and I don't see governments gutting copyright, contract law, and several other avenues of protection that copyright holders can deploy in the name of something that doesn't exist and may not ever exist.
It’s already happening with EU AI Act https://www.europarl.europa.eu/news/en/headlines/society/202...
I have nothing against creators, they deserve to get paid.
For what its worth, LLMs are facing the coke vs Pepsi challenge, and sadly they are most definitely Pepsi.
https://aws.amazon.com/service-terms/
See items as of 50.10 and 50.10.1 that I reproduce here:
"50.10. Defense of Claims and Indemnity for Indemnified Generative AI Services. AWS Services may incorporate generative AI features and provide Generative AI Output to you. “Generative AI Output” means output generated by a generative artificial intelligence model in response to inputs or other data provided by you. “Indemnified Generative AI Services” means, collectively, generally available features of Amazon CodeWhisperer Professional, Amazon Titan Text Express, Amazon Titan Text Lite, Amazon Titan Text Embeddings, Amazon Titan Multimodal Embeddings, AWS HealthScribe, Amazon Personalize, Amazon Connect Contact Lens, and Amazon Lex. The following terms apply to the Indemnified Generative AI Services:
50.10.1. Subject to the limitations in this Section 50.10, AWS will defend you and your employees, officers, and directors against any third-party claim alleging that the Generative AI Output generated by an Indemnified Generative AI Service infringes or misappropriates that third party’s intellectual property rights, and will pay the amount of any adverse final judgment or settlement."
It’s politically 100% viable to kneecap AI with copyright restrictions. This will go to the Supreme Court and it’s far from clear whether fair use applies to every case here.
IMO, what's most likely is some sort of licensing model between the AI companies and the 'big content providers' (remember most content on the web these days is not owned by the person who created it, wasn't always like that). The smaller companies then would be forced to live with either being scraped or ending up being 'invisible'.
If that's so, things are about to get worse for everyone, too. With little to no protection against AI, no one will be incentivized to create new IPs, whether they're books, drawings, songs. Or even films and games, when AI is able to also generate those in the (possibly near) future.
This is really about replacement. The copyright holders in the content industry aren't really afraid of LLMs infringing on past copyright, but are terrified of it replacing them on future work, and there is absolutely no legal protection from this. The lawsuit might officially be about copyright, but that's just because it is their only available legal angle of attack.
LLM people are really starting to veer into crypto-bro territory with the evangelising about how they’re the best thing since sliced bread and transistors.
How would that be any different than Google displaying some type of news headline?
Sorry if your 40 hour work won't pay you $10 bucks a month forever. That's the case for most of the rest of us: we produce for 40 hours, we get paid for those 40 hours, regardless of what we do.
Welcome to the club!!
I hire a session musician to play on my new single, paying him $100. I record the whole session.
I ask him to play the opening to "Stairway to Heaven" and he does so.
"Well, I can't use that as a sample without paying"
"Ok play something like Jimmy Page"
"Hmm, still sounds like Stairway to Heaven"
"Ok, try and sound less like Stairway to Heaven but in that style"
"Great, I'll use that one"
and I release my song and get $5,000 in royalties.
Should I be sued for infringement, or the guitarist?
The problem, I suppose, is that if I had said "play something like 70s prog rock" and he played "Stairway to Heaven" and I didn't know what it was and said "great, I'll use that".
Should I be sued for infringement, or the guitarist?
I still think people who are concerned that art related copyright will stifle generative AI should fight copyright laws directly. But that’s a harder pill to swallow since it will cause multi-industry wide havoc.
Asking your session musician for something "melancholy" and having them pass off Stairway to Heaven as original would be unreasonable.
It’s like a software engineering contractor. The contractor gets paid, the IP of their work is owned by the company.
For whoever owns the output also owns the liability of the output. You yourself might separately be able to pursue a claim against the guitarist for breech of contract. In the process of that, it might get discovered that you deliberately instructed the guitarist to copy the work, or they copied despite your instructions not to.
But that doesn't change the fact that the final work is infringing. It just allows you to pursue damages that could potentially offset any damages you're liable for from the infringement.
But this also isn't exactly the same situation as OpenAI. OpenAI isn't an individual creator working on contact for you. Even if their ToS ultimately assigns copyright of output to you, there is a matter of scale involved that I think changes things. It's one thing if your guitarist damages you by doing shoddy work, it's another of the guitarist systematizes and scales their shoddy work to damage large numbers of people. Perhaps that would then become a class action issue.
He basically played the four chord song in court, and showed that the prosecutions’s song was “copying” an earlier song.
https://amp.theguardian.com/music/2023/may/04/ed-sheeran-ver...
It could be argued ChatGPT is a publisher too.
Remember when everyone and their dog discovered sampling in the late 80's and they all thought they could get away with it because it didn't seem like infringement to the samplers? The courts had no qualms about slapping record labels for putting out records with unlicensed samples in them. Albums even got pulled off shelves while licenses were sorted out.
These companies are charging for a service that returns copyrighted content, full stop. You can't do that whether you are AI or someone drawing Mario and selling the pictures on iStock, or putting out records that sample someone else's work without permission. It took a while in the case of sampling, but it sure as hell happened.
IMO would be best if this stays a highly illegal technology that is only available to a few weirdo nerds /s