I don’t think it makes sense for both model builders and the model’s users to separately obtain licenses for the same works used in the training set.
A model trained on several copyrighted data sources cannot somehow be used in a way depending on a subset of those sources.
So all parameters of usage and compensation should be settled by contract between the model builder and copyrighted data supplier, before the copyrighted material is used.
Or to put it simply: using copyrighted material to create a model would NOT be considered fair use.
That’s it. That’s the standard. No complicated new laws required.
Model builders obtain permission to use copyrighted material from copyright holders based on any terms both agree to.
Terms might involve model usage limits, term limits, one time compensation, per use compensation, data source credits, or anything else either party wants.
The likely result will be some standard sets of terms becoming popular and well known. But nobody has to agree to anything they don’t want to.
I slightly disagree, in that I think the person using the tool should bear the burden of copyright. I.e. if the model outputs something under copywrite it merely can't be republished. In this same way, i can use Photoshop on proprietary data but I can't necessarily sell the results.
I'm so torn. On one hand, what you suggest seems to be a nearly ideal balance between advancing scientific progress and legal liability. By placing the legal burden to publish generated works on the person actually trying to publish, it allows for a more nuanced legal approach (i.e. the difference between "there are similarities to this work, but it's murky" or "you %100 stole that work").
On the other hand, is the company running the model themselves not already publishing all of that work and profiting from it? It seems unfair that their bottom line gets to be bolstered because they can produce work based on any artist, whereas the consumers of that work may need to end up walking on egg shells in order to publish them.
Like I said, I'm torn as far as how it "should be". I know how I want it to be though. I would love if AI continued training unabated. The results have been amazing, and I believe it would be a shame if the effort was slowed down by legislation.
> is the company running the model themselves not already publishing all of that work and profiting from it?
no, because the model is transformative enough that it cannot be said to be a derivative works of the training set.
The model is in essence a form of distilled information, extracted from the training set. Information cannot be copyrighted - only expressions can.
Therefore, a model producer should have the right to use any pre-existing work, in the same way a person can, to study and internally memorize and extract information.
The reproduction of any of the training set data constitutes a copyright violation, but this is not done by the owner of the model, but by an end user of the model.
My point is that if a court finds that a generated image is indeed similar enough to constitute an infringement when a subscriber of for instance MidJourney attempts to publish it, has that work not already been "published" to the subscriber? And has MidJourney not profited by gaining a subscriber based on the work of others?
I wonder if that analogy represents the same thing. Speaking purely from a non-legal perspective on the ethics in my mind:
When you use Photoshop on propriety data you're providing the original data and choosing what manipulation to make (i.e. what tool) and directly creating the output. It makes sense that if you redistribute this it may be copyright violation.
When you use Copilot or ChatGPT for programming you're typically asking a non-proprietary question or accepting suggestions it's making based on non-proprietary (or proprietary to you) code in the file. You also don't dictate the manipulation process a black box deep learning model does (i.e. I haven't asked it to do something that could be reasonably thought to be a copyright violation).
Am I then responsible for the fact that Copilot is fooling me with effectively copy-pasted copyrighted code when it's being presented to me as generated by the software and I haven't instructed the software to commit a copyright violation? I'm not sure if intent matters for copyright, I assume it doesn't but perhaps that's a missing piece to this.
Diffusion models are gray to me, if you're asking/prompting with "Mickey Mouse riding a horse" I can see the argument that the prompt itself can be interpreted as asking the model to commit copyright violation and the user is just hiding behind a layer of abstraction. If I ask the model to spit out "a picture of a smiling cartoon woman" and it generates a Betty Boop lookalike is that still the users fault?
It seems to me like passing the burden to the user could be reasonable but would need some safe harbor type of exception. It'll be really interesting to see what the courts decide.
(1) how do you know if the image that just generated is substantially similar to an existing copyright work? Maybe if some registration tool existed, but other wise the burden is too great
(2) what is stopping someone from generating millions of images and copy righting all the "unique" ones? Such that no one can create anything without accidental collisions.
> how do you know if the image that just generated is substantially similar to an existing copyright work?
This is already a problem with biological neural nets (i.e. humans). I remember as a teenager writing a simple song on the piano, and playing it for my mom; she said, "You didn't write that -- that's Gilligan's Island!" And indeed it was. If I had made a record and sold it, whoever owned the rights to the Gilligan's Island theme song could have sued me for it, and they would (rightly) have won.
There's already loads of case law about this; the same thing would apply to AI.
> what is stopping someone from generating millions of images and copy righting all the "unique" ones? Such that no one can create anything without accidental collisions.
Right now what's stopping it is that only humans can make copyrightable material; whatever is spat out from a computer is effectively public domain, not copyrighted.
1. lots of established law and case law (at least in the US), this is already a well-settled problem and folks have the tools and proper venue to bring infringement claims. Yes, federal copyright infringement litigation is prohibitively expensive for many issues. There is a now a "small claims court" for smaller issues. [1]
2. Those works cannot be copyrighted (at least in the US). [2]. And hey, someone already tried copyrighting every song melody [3]
Copyright holders are already protected from (I.e. can legally prohibit) distribution of obvious copies, or clearly derivative works.
Regardless of how they were produced by hand, copy machine, Photoshop or with a model.
The new problem is that artists styles are being “stolen” by incorporating their copyrighted work into models without their permission.
And that problem can easily be solved if using copyrighted material to create models is declared NOT fair use.
Artists could still allow models to be built from their work, but on their terms. If they wish to do that.
A famous artist, that doesn’t mind being commercial, could sell their own unique model to let fans create art in that artist’s style, while not having their style “ripped” by others.
Or just keep their style to themselves, for their own work, as artists have done for centuries.
(Of course, with greater effort, their style could still be recreated - styles are not protected unless they are trademarked - but the recreation would have to be done without using the artist’s copyrighted works.)
This is probably a somewhat unpopular opinion on HN, but it is where many of the artists I work with are generally trying to get to. Consent, compensation, and credit.
This is the best path forward I think. And it will become increasingly sensible as things continue to evolve. AI wasn't necessary to violate copyright before, and it isn't necessary today.
The determination of copyright violation should be made against the output of the model in the event that someone uses it for commercial purposes.
If the models have a risk of generating copyrighted content, it will be up to the consumers of the system to mitigate that risk through manual review or automated checks of the output.
A divergence, but I see a lot of posters asserting that "humans learn by copying other people, but we don't call that a violation of copyright when they draw"
People casually asserting that software is equivalent to humanity will be a non-negligible thing to consider, as irritating and poorly-founded as it seems.
If the reproduction isn't pixel-perfect, but merely obvious and overwhelming, how do you refute that philosophically to people who refuse a distinction between 50GB and a human life?
> People casually asserting that software is equivalent to humanity will be a non-negligible thing to consider, as irritating and poorly-founded as it seems.
> If the reproduction isn't pixel-perfect, but merely obvious and overwhelming, how do you refute that philosophically to people who refuse a distinction between 50GB and a human life?
Software equivalence to humanity is a very philosophical question that many sci-fi writers have approached. But our primary issue related to this technology does not depend on anyone making a determination there.
The challenge is that losses to livelihood from this technology are going to come from far broader impacts than copyright alone. Copyright disputes are just the first things to get everyone's attention.
Let's say we err on the side of protection of copyright, and all training data must be fully licensed, in addition to users being responsible for ensuring outputs did not accidentally reproduce something similar to a copyrighted work, even if it was part of the licensed training dataset. Great! This fixes the problem of lost value for the owners of copyrights. Companies will face a slight delay and slightly increased costs as they license content; however, in the end, model capabilities will be the same and continue to increase.
The number of jobs that actually cannot be performed without humans will continue to dwindle — livelihoods will be lost at essentially the same scale despite upholding copyrights.
The only way we can handle a technology capable of reducing most need for human labor is by focusing on planning and executing a smooth transition toward an economy with more people than jobs — aiming for minimal human suffering during this process.
A mass loss of human jobs does not need to mean a mass loss of livelihood if our society is prepared to transition to a universal basic income. After all, human life is far more than just a job. We have the opportunity for much more fulfilling lives if we plan this transition well. We must understand that this is a far larger issue than copyright - copyright disputes are just one of the first symptoms of this disruptive process.
A human is still entering the prompt to generate the possibly copyrighted image/text. I don't think copyright law should care about the implementation. It's ok to copy a style if you use paint brushes or photo shop. But not ok if you use a statistic model?
> Or to put it simply: using copyrighted material to create a model would NOT be considered fair use.
The more I think about it, the more something along these lines seems like it might be the right way to think about it.
When you play a DVD, for example, you copy the bits off the DVD, into the memory of your DVD player, and onto your screen; this is all explicitly considered "fair use" copying. But if you then copied those fair-use bits off the screen onto a thousand other screens, that violates copyright.
When you, as the human watch the DVD, bits of it get copied into your brain; but you don't then copy the bits of your brain to millions of other people -- they each have to make their own copy.
We could make the law for LLMs follow a similar logic: That having an LLM watch a video or read a text is similar to having a DVD player read a DVD or a web browser copy information from a website. It's good for that limited use case, but the resulting copy cannot be copied again without a license.
This would allow (say) researchers, or even individuals, to do their own training and so on without a license; but when anyone wanted to create something that they wanted to scale up, they'd have to get licenses for everything.
That would fundamentally keep things balanced as they are now with creators and other creators. The big problem isn't that a handful of other creators may be copying their style; that growth in competition is limiting because of the expense of duplication. It's that millions of electronic engines can copy their style.
> When you, as the human watch the DVD, bits of it get copied into your brain; but you don't then copy the bits of your brain to millions of other people -- they each have to make their own copy.
If you ripped The Little Mermaid, redrew every frame to combine it with The Fresh Prince of Bell-Air and moved things around in scenes to make it look like Ariel is Will Smith responding to sit-com dialogue, then it'd be fair use, regardless of how many people you show this new version to.
> I don’t think it makes sense for both model builders and the model’s users to separately obtain licenses for the same works used in the training set.
I'm torn on who should pay, and where and when. In the world of patents, there's often an option/split. Say a chip manufacturer wants to build H265 decoding into their hardware. The chip manufacturer could buy the license. Or the purchaser (who probably is building some sort of board or device around the chip) could pay for the license. Or they could disable that functionality in the end product, and the consumer could pay for a license (or not, if they don't care about that feature).
The most common is usually the middle option: the end-device manufacturer (or brand that eventually sells the product) will pay for the license.
But I'm not sure if this works all that well for an AI model. With hardware, the license is usually paid per unit. It's easy to see that one chip = one license. If the model builder buys a license, that model could be used one time or 100 million times. Tracking use like that probably isn't all that practical, but I think it's safe to say that a 100-million-use model should probably pay more for a license than a single-use model.
So maybe the model builder should be responsible for attaching a comprehensive "copyright history" to the model, and users should have to pay for a license based on their use? Again, not sure how to track that. But I guess general software licensing has similar problems when you can "hide" usage.
Yes, someone using a model can’t know if the generated text/image/sound is a nearly identical copy of the original material they don’t recognize. If use of the output of these systems comes at significant legal risk then then such systems become nearly useless.
> if the generated text/image/sound is a nearly identical copy of the original material they don’t recognize
how does the industry today deal with artists that "copy" off some other works? This isn't a problem with AI at all - just that AI provides a tool to generate such works faster.
Someones comes to me to ask for a drawing of Batman or to write an erotic story around Supergirl. I can do it, but I cannot claim ownership over the characters. And I think I will quickly get a letter from DC or Marvel if I try to do this at scale.
> I can do it, but I cannot claim ownership over the characters.
of course not. But you can claim ownership if you don't call those characters their original names, and make sufficient changes to the design (how sufficient is determined by a court of law - thus expenses).
> DC or Marvel if I try to do this at scale.
The show 'invincible'[1] has a character that is a basic copy of superman. And yet, you will find that they don't get a letter from DC.
I think that’s one of the issue. The transformation done by these tools are mechanical. Even if it may be extensive. The human input is too small. Omniman may have similarities with Superman, but he is not him in the larger context of the story. LLMs can not yet be that consistent for marketable output that deserves to be copyrightable.
I’m perfectly fine for LLMs to aid with spell checking and alternative phrasing (image is a grayer area). Bu the ideas of prompts and prompt output being copyrightable is something I oppose.
the prompts have become somewhat creative these days. If you have a look at the prompts on https://civitai.com for example, you can argue they are a form of creative expression. Just like hand rolling assembly code might be.
High detail, dynamic action pose, masterwork, professional, fantasy, neo classical fine art, of a beautiful, primordial and fierce, ((angel-winged-woman,:1.9)), archangel, (MiddleEastern:1.6), with very long, flowing, wavy white hair, peach colored streaks, with a sexy, slender, fit body, wearing an ethereal, light violet, light aqua, faded gold, tie-dye, linen and Chantily lace, (knee length:1.5), strapless dress with a tattered hem, a Platinum and gold Cuirass, platinum vambraces, platinum and lace Gladiator Boots, long broadsword in a Baldric, at night, in a metropolis warzone, during a thunderstorm, dimly lit, thin, vibrant streaks of crimson light, outlining her body, fantasy illustration, in the style of Osamu Tezuka, George Edward Hurrell, Albert Witzel, Hiromitsu Takeda, Clarence Bull, Gil Elvgren, Ruth Harriet Louise, Takaki, Milton Greene, Huang Guangjian, and Cecil Beaton,, High detail, dynamic action pose, masterwork, professional, fantasy, neo classical fine art, of a beautiful, primordial and fierce, ((angel-winged-woman,:1.9)), archangel, (Columbian:1.6), with very long, flowing, wavy white hair, peach colored streaks, with a sexy, slender, fit body, wearing an ethereal, light violet, light aqua, faded gold, tie-dye, linen and Chantily lace, (knee length:1.5), strapless dress with a tattered hem, a Platinum and gold Cuirass, platinum vambraces, platinum and lace Gladiator Boots, long broadsword in a Baldric, at night, in a metropolis warzone, during a thunderstorm, dimly lit, thin, vibrant streaks of crimson light, outlining her body, fantasy illustration, in the style of Osamu Tezuka, George Edward Hurrell, Albert Witzel, Hiromitsu Takeda, Clarence Bull, Gil Elvgren, Ruth Harriet Louise, Takaki, Milton Greene, Huang Guangjian, and Cecil Beaton,
That’s a perfect example, they said “during a thunderstorm” does that image look like it’s in a thunderstorm? Sure, the output of the prompt relates to what was said, but they influenced the output rather than controlled it.
Further, it’s well known that simply telling an artist what you want even including quite detailed descriptions isn’t enough to get copyright over the resulting image.
The difference is the artists assertion that it’s either original or a copy from something else. DALLE 2 can’t tell you if it’s original or not. These AI’s have no idea and the company or group that created them doesn’t review individual output so they can’t say either.
whoever pressed the button to run DALLE will make the assertion, just like whoever was running photoshop to make the image today would make the same assertion.
A photoshop user controls what data photoshop uses, a DALLE user doesn’t. Even a prompt as generic as “Cat” could be producing an obviously derivative work if you compare it to the original. This is true for all prompts.
> A photoshop user controls what data photoshop uses
the point was that the user of the program is making their declaration, whether it's photoshop or DALLE. How does the business verify that their staff artists aren't producing copyright infringing material, just from memory?
The liability falls to them to verify the copyright status of the output they're asked to make. A business paying a photoshop user to produce a picture has just as much (or as little) trust in them as the button presser for DALLE.
This gets complicated, having no reason to know that something is copyrighted is a defense.
So if your employee installed pirated 3rd party software you’re facing strict liability. However, if a third party is reproducing their collage roommates drawing from memory then it’s effectively impossible for you to verify if something is a derivative work.
Dalle is effectively Getty images, if you’re buying works from them you can only assume it’s free of copyright issues.
Problem is, how can you determine if the model contains copyrighted material? The laws governs copyright through ownership, so in order to claim copyright infringement you have to be able pinpoint a specific person and prove that their work is somehow embedded in the gradients, which is not practically possible at the point. It's just like how you can't practically enforce copyright on encrypted data unless you ban encryption altogether.
1. If you know your copyrighted material was in the training dataset is that not sufficient?
2. From a legal perspective do you actually have to prove it's embedded in the gradients? If I draw an exact copy of Mickey Mouse from memory and sell it I didn't think Disney had to prove I've ever actually seen Mickey Mouse before or point to where the image of him is embedded in my brain.
Disney has a trademark on Mickey mouse, but that does not mean that they automatically get copyright on all pictures of Mickey Mouse drawn by others (they don't)
A model trained on several copyrighted data sources cannot somehow be used in a way depending on a subset of those sources.
So all parameters of usage and compensation should be settled by contract between the model builder and copyrighted data supplier, before the copyrighted material is used.
Or to put it simply: using copyrighted material to create a model would NOT be considered fair use.
That’s it. That’s the standard. No complicated new laws required.
Model builders obtain permission to use copyrighted material from copyright holders based on any terms both agree to.
Terms might involve model usage limits, term limits, one time compensation, per use compensation, data source credits, or anything else either party wants.
The likely result will be some standard sets of terms becoming popular and well known. But nobody has to agree to anything they don’t want to.