> The list of 16,000 artists was included as part of a lawsuit amendment to a class-action complaint targeted at Stability AI, Midjourney, and DeviantArt and the submission of 455-pages of supplementary evidence filed on November 29 last year.
One positive aspect of the status quo in the United States is that AI-generated images are not currently eligible for copyright. I think this is a great direction to go in, I highly doubt Wizards of the Coast or whoever is going to want their premium products to lose copyright protections, so they'll need to keep paying artists. I'd love for us to lean into this -- you can make all the AI art you want, but it automatically gets a Creative Commons ShareAlike-style license!
No, it's a Butlerian Jihad[0] argument. The Copyright Office's argument holds even for a fully public domain training set. US copyright law is already speciesist[1] - you can't assign authorship to an animal - so computers are also forbidden from authorship.
[0] In the Dune universe, the "Butlerian Jihad" refers to a legal ban on thinking machines.
Seems like that only applies if you anthropomorphize the AI and consider it the author, rather than a tool utilized by the artist. I mean, yes, the AI is doing the bulk of the work, but so is photoshop for a lot of digital art.
Photographs can be copyrighted, and all the human did was aim and press a button. Oh, and potentially travel to specific locations, adjust parameters, choose lenses, stage a scene, makeup, wardrobe, lighting, etc. etc. But none of those things are required for the image to be copyrighted
I'm not really sure how this connects to the argument. No one is trying to grant authorship to an algorithm - it would be a ridiculous effort that was never even in the cards. In these copyright disputes, the authorship on AI outputs would be on the person using the AI. Generative AI takes inputs that are provided by a human and transforms it into certain outputs. Legally speaking, I don't see it as different from me getting protection for something I did in Photoshop - trying to somehow give Photoshop authorship would be absurd.
I agree it's not quite the right argument. IANAL, but I think it's more illustrative to remove AI from the example entirely. If you wrote a prompt and gave it to a human artist to draw, would you have joint copyright over the resulting work? If you didn't do anything worthy of copyright, and the AI cannot be granted copyright, then it is not copyrightable.
That said, it seems like a moot point to me. The practical uses of generative AI are not going to be one-and-done prompt-to-image tools. When AI is used more like a brush, the brush strokes the human chooses will still be granted copyright.
Maybe but copyright on a remixed song that samples other artists is really difficult to deal with in terms of rights unless one party owns the rights to all the samples it’s difficult to negotiate and you might not get any royalties. AI art is basically just other people’s work remixed so the same problem applies
The Copyright Office wrote their guidance specifically because someone tried to register a comic book they wrote and put AI art in. They specifically credited Midjourney as co-author.
Their guidance would not apply for someone using AI as a tool, but said copyright would be thinner than if you'd drawn everything by hand. Specifically, you don't own any of the things the AI "just came up with". If you just wrote a prompt and grinded out some results, you probably own nothing[0]. If you use shittons of inpainting to control, say, the overall composition, but the AI filled in pixels somewhere, then you probably still own the overall image, but that's only because I'm not sure how you'd separate the two in a way that would let you copy just the AI-generated portion. Or, in the case of the comic book I mentioned earlier, they own the text, characters, and plot of the comic book, but not the artwork.
[0] Yes, you could probably just lie to the Copyright Office. Make sure to never reveal your use of AI to anyone, because there's loads of angry artists who would love to tattle on you.
Completely unenforceable. How can you even tell if an image was made by AI? What if AI created an outline that was worked on by a human artist (or vice versa)? Who would the burden of proof be on?
Steam has a “no AI art” policy, and it’s rapidly turning into a “no obvious AI art policy”. How could they tell?
The thing about AI art is that, absent lots of prompt engineering, seed grinding, and touchups, you're likely to have a bunch of images that are obvious tells if your entire project is AI. Anyone trying to hide it would be spending time equivalent to just making the art themselves.
There's also another advantage to having a "no obvious AI art" policy; and that's to cut down on spam. AI is extremely useful to people who want to spam art platforms.
And this will continue to be an issue until ML models have achieved something resembling sentience, because many of these tells are the result of the model not truly comprehending the subject matter and thus struggling to maintain internal consistency in everything from geometry and kinetics of human bodies to lighting and physics.
Less obviously, ML models also lack the ability to bake in intent. In human made pieces, everything is as it is for a reason; it’s communicating something. In ML generated pieces, things are the way they are because that’s what’s statistically likely for the type of generated image.
Absolutely correct. This is a far more obvious problem in text models, because you end up with internally flawless arguments as to why your next scuba diving vacation should be in Ulan Bator.
With art, it's more subtle, because there there's no single reference point that lets us determine if an artwork is "true". There are the glaring errors that everyone can agree on - notoriously, human hands - but those cases are improving rapidly.
> Anyone trying to hide it would be spending time equivalent to just making the art themselves.
The gap of being distinguishable from manually drawn images is still closing - we don't know if it'll ever reach the threshold, but the amount of effort required to stamp out all the wonkiness from an AI generation has been going down ever since the first viable algorithms appeared.
I don't think that this was an anti-spam policy - Steam already manually reviews all new applicants that want to publish a game, so they don't need to forbid anything to turn it down. I'm guessing that this policy was because they don't want to be entangled in IP legislation if some copyright exception is carved out to forbid the use of generative AI.
In just the past two years it's gone from obvious horrors like hands attached directly at the elbow to much more subtle errors like chair legs that cross over each other like an Escher drawing or doorknobs adjacent to the hinges.
Human artists might have to become used to tracking provenance. If you work with traditional media, that's easy: Here's the painting. For digital artists, software can publish encrypted, timestamped brushstroke-level histories of the work if we need that level of proof.
> The gap of being distinguishable from manually drawn images is still closing
People have been trumpeting this since day one of Stable Diffusion releasing, but I'm seeing the same output quality as that day and I've been keeping up.
Just because the pace of progress isn't exponential (like what some people would want to believe) doesn't mean it isn't happening. I remember getting an early invite to DALL-E 1 all the way back, and while I don't use it anymore, the modern improvements made seem very substantial. From plain comparisons of different versions where the same inputs produce substantially better outputs, to the mere fact that the latest version can actually generate decent, often discernible text at all (something that people joked would be impossible from AI to achieve) shows that some progress is being made.
The reason why it's not as visible with Stable Diffusion is because a lot of the technologies around it circle the same few foundational SD models - people build on top of them, add new ways of interacting with them, but ultimately, the same thing underlies them all. Community support is seen as more important than cutting-edge tech, which is why something like Stable Diffusion XL hasn't even seen universal adoption yet.
I'm telling you the progress isn't happening based on my own consistent observations of various releases across multiple platforms. The only people who don't seem to agree with me are those who have the art literacy of a highschooler and think "discernible text" is a improvement.
As an aside, no one said AI couldn't achieve drawn generated text, that's been possible for years prior to stable diffusion.
IME, if GenAI ever reaches human parity, whatever that amounts to, the relevant subgenre of art will just move into surrealism. Invention of paintbrushes didn't kill art.
the fight about AI is using copyrighted stuff for its weights... i wonder which % of artists that wouldn't tweak or use heavily AI that has a transparent/ethical data-base (read it: they didn't added anything proprietary without authorization)
It's been tried. Numerous times. There's a reason why GenAI controversy is stuck at ethics and filled with rage, the generated images just aren't that great and so that part isn't so controversial.
There's a lot of AI "artists"[0] who think their text2image prompt generations are equivalent if not better than actually drawing or photographing an image.
Part of becoming an artist is learning how to evaluate your own work, break it down, and critique the shit out of it. When you jump straight to generating art with an AI, you skip the criticism step, which means you don't have a sense of taste and you haven't really explored what your preferences for style are.
A lot of AI art generators default to an extremely cinematic, "Hollywood" art style - i.e. exactly the sort of thing that is trying to look impressive to people who don't know better, and will make them overlook all the fundamental mistakes in the image.
[0] Normally I wouldn't scarequote "artist" here, given that actual artists do use AI tools where it makes sense.
>Anyone trying to hide it would be spending time equivalent to just making the art themselves.
Microsoft and every other tech company is indeed investing billions in the tech. I'm sure each company can fund the entire (woefully underpaid) art industry by themselves, let alone the 10 or so tech hubs altogether.
But they are happy to throw money at AI instead for the payoff of being the next big tech brand.
It makes perfect sense because much of the friction around generative ML models has to do with the data it was trained on. There’s not much reason to ban images generated by a model that was trained entirely on consensually gathered material.
Complete wrong. You just flip the defaults--something is AI unless you can prove otherwise.
This is done already and has precedent. Producing porn requires that you keep artifacts demonstrating that who the performers were, that they were of age, etc.
If you claim a work is not AI generated, you should have to produce some artifacts to back up that claim.
In the case of a corporation, that would be easy as you have payment records.
In the case of an individual doing digital only, that's a little harder. You probably have to keep some intermediate artifacts.
> If you claim a work is not AI generated, you should have to produce some artifacts to back up that claim.
So AI (or a simpler, non-statistics-based algorithm) can't produce these artifacts and will never be able to? Why? Are these artifacts "human souls" or something?
> In the case of a corporation, that would be easy as you have payment records.
I'm really struggling to conceptualize a world where every picture that's drawn must have a full notary log of how exactly it was produced, all for the sake of removing generative AI.
Besides, it's not that easy of a problem - a lot of corporate artists are salaried workers, they don't get specified commissions with an attached bill per work, but are paid a salary so the company can ask them to draw whatever they need throughout the process. Considering this, all artists would need to retain "intermediate artifacts".
And then, how do these artifacts work for other ways of doing art? What about traditional artists whose work gets scanned in after completion - would they have to keep a camera on hand to take photographs as they're working? What about an animated film - would every intermediate step in production, from character design to storyboarding to environmental design etc etc need to have a full record for every single sketch?
It is for the sake of copyright, if you want society to protect your work, provide evidence for your creative work. It seems rather simple to me.
Keep in mind that in the not so far future, producing art will be as cheap as consuming it, this means that the original benefits society got in return for copyright no longer applies, so why should they protect it?
This is a really good point that I think will be hard for a lot of people to come to terms with. The basis for the whole idea of intellectual property rests on assumptions that look increasingly fragile.
> It is for the sake of copyright, if you want society to protect your work, provide evidence for your creative work.
I'm not sure if it's that simple - for one, this requirement is a complete departure from how copyright systems work now. Providing complete history logs isn't normal practice, and expanding law to necessitate it isn't common sense.
> Keep in mind that in the not so far future, producing art will be as cheap as consuming it, this means that the original benefits society got in return for copyright no longer applies, so why should they protect it?
I'll make a prediction that this future is further from now than you may think it is. Sure, things like static imagery may become completely indistinguishable from human-made art in the near(ish?) future, but the production of all art is still an unsolved problem. How long will it take until some advanced multimodal algorithm can make a full game that can measure up to ones that are released today? I'm guessing that it'll take a while.
And yeah - once we do reach this scenario of hypothetical "art post-scarcity", we may as well just delete the whole copyright system from existence - it'd be a logical thing to do. But how does any of it contradict what I said in my other comments?
Nothing in the form seems out of the ordinary to me. It is a lot of fields, but ultimately the main goal is establishing ownership, not discerning the specific methodology in which a person made the work. It's a departure in that the current system is results-based, where you register a final product, while the proposed system also must take into consideration every intricacy of creating the work.
> it is increasingly refused because of indicia of AI tooling.
Do you have a source that a statistically significant number of copyright applications gets refused on account of a work just seeming like AI? On what grounds does it get refused?
Asking for a statistical analysis is an unreasonably high bar. See the link I provided in a sibling comment, in which the copyright office plainly states as much and explains their guidelines.
> How long will it take until some advanced multimodal algorithm can make a full game that can measure up to ones that are released today?
We can disagree on how long it will take us to get there, but if you use AI generated content, that is not product by copyright, your game as whole, sure, as long as it is not the result of a simple prompt, you're protected as usual.
Keep in mind that already, in many games, there is a mix of protected and unprotected content, for reasons of trademark, copyright, and licensing.
The word choices were kind of on purpose - I meant to highlight the partial absurdity of having to entangle yourself with all these legal considerations and obtaining sufficient legal proof, all for the sake of making some art piece.
If it's physical media, you have the physical media.
If it's digital media, the software can keep an encrypted record at the brushstroke level that can be played back to produce a bit-perfect reproduction. Maybe even write it to a public ledger.
All of these things have loopholes. For physical media, depending on the quality of the output, one could pay a sufficiently skilled person to reproduce an AI output on physical media in a fraction of the time it'd take to come up with and draw for real.
For digital media - ignoring how overbearing this whole system could be, what prevents someone from taking all that data and making an algorithm that outputs brush stroke parameters instead of pixels? And digital art isn't the only thing we need to concern ourselves with - eventually, we might have AI models that could make 3D models, sounds, vector imagery and other forms of art. The idea of just documenting every workflow would be an ever-growing burden with no perfect solutions.
> All of these things have loopholes. For physical media, depending on the quality of the output, one could pay a sufficiently skilled person to reproduce an AI output on physical media in a fraction of the time it'd take to come up with and draw for real.
Um, that's a real work, you know? In what way does this differ from people who take a photograph and then, for example, creating an oil painting?
Now, there are some weirdnesses because of the copyright of the source photograph, but the oil painting would be your own work.
Yeah, you might get called into court to demonstrate that you can produce the work. But so did Michael Jackson.
I think there's a lot of confusion about what records are needed presumably due to lack of understanding of what the law requires for proof of ownership.
Generally the party asserting ownership has the burden of proof. The standard is "preponderance of the evidence", which generally is understood to mean "more likely than not" or "> 50%". So basically it means if you can prove to judge or jury that there's a >50% chance you own the work, it's good enough.
Also note that in many cases where there's a dispute over the evidence, witnesses are summoned to testify. So you might not have a "full notary log" of how it was produced or all the "intermediate artifacts", but as long as the artist is able to convincingly explain how the work was created, and the other party's lawyers are not able to poke holes in their story during cross evidence, that's usually enough.
Which is, basically, what happens today, if the authorship or ownership of a work is disputed.
That said, I'm not sure whether "assume work AI (thus uncopyrightable) unless proven otherwise" should be the default for other reasons. For one, most quality "AI art" needs some manual adjustments or touch ups, and arguably the prompt and hyperparameters may be sufficient creativity element. I mean, that's basically how copyrights dealt with photography (the mere fact you decided when and where to point the camera with what settings is sufficient for copyright to subsist in a photo).
I think this is rather what pro-AI/spammers are trying to do by flooding platforms, that aren't so successful. People don't give as high scores they do for human generated data, and AI images are still considered a form of spam.
The end result of this will be end to end cryptographically authenticated pipelines. I'm sure Adobe will be very happy about having yet another way to extract money from artists.
"I claim this art was made by this person" "Who?" <gives name> "OK <name>. did you work on this?" "Where are related work products? Are there any? What about invoices? Simultaneous employment?"
The reality is that most legal things are determined by _convincing people of a truth_. Perhaps you can set up a whole scheme to "launder" AI art and attach names to them. And all the papertrail you generate doing this will show up in discovery in some lawsuit and the copyrights all disappear.
The way that AI will be laundered into art is by including it into things like Photoshop. There'll still be a human touch just with "smart brushes" and "smart auto fill" that paints 90% of what you want.
Art will then take less skill to produce, and be produced faster for lower prices.
An 80% price reduction on art (because artists can now produce it 5 times faster thanks to AI) is 80% as good as getting it for free.
Art will take more skill to produce, not less, at faster speed by select artists. GenAI will become another tool that artists and clients alike must understand and use effectively within unspoken guild rules, that is, if it stays.
Furthermore, what is the threshold for something to still legally be considered AI art once an artist's hand has modified it? What if they change the brightness? Fix a hand? completely replace a character in a scene? Illustrate most of the scene themselves but add an AI figure or background? Use AI to sharpen a hand drawn image?
that means that openai cannot claim copyright on images produced by dalle generator. no can other online services and offline software owners/producers.
this is enough to use them without copyright violation for example for other ai models training.
A company today has the burden of proof to demonstrate authorship of a claimed work when they sue for copyright infringement. This isn't a crazy expansion of that concept -- companies do not generally break the law just because there's a low chance of getting caught. Furthermore it's not hard to imagine, for example, a whistleblower calling out their employer for copyrighting AI works.
I think (hypothetically, they may choose not to do this for a number of reasons) WoTC could still dramatically cut back on how much they pay artists by “outsourcing” things like backgrounds, extended art, etc to AI so long as the focal point of the piece is human-created and therefore copyrightable.
Ironically, there's a massive scandal with WoTC right now for doing this. They say they aren't using AI, the majority doesn't believe them due to artifacts only AI would produce for backgrounds.
From [1] "Not every character qualifies for trademark protection, however. For a character to be trademarked, the character cannot be too similar to other existing trademarked characters and must be used to brand products or services. Once a character meets these requirements, the owner can file for trademark protection."
So I don't know if you could apply trademark to e.g., every card in magic, but maybe only to the key characters?
That won't work. Trademarks are a lot harder to establish -- it's not just automatic from the moment you publish it as with copyright. You have to first start using it, then always mention it's a trademark when using it (with the TM or R symbols for example), then wait for it to catch on, then file with the government some paperwork. (Iirc, exact process is probably similar but different, but the point is it's a lot more involved)
Trademark is intended to protect the holder from being impersonated, not from losing revenue from selling content.... So it's a lot easier to redistribute copies of trademarked work as long as you make it clear you are not affiliated with the trademark holder, in a manner which a reasonable person would heed.
So for example, if a piece of art is trademarked by Disney, and it is well known by the public, and I print a copy and put it in front of my shop, a reasonable observer might this my shop is owned, operated, or endorsed by Fisney. So that's not OK.
If instead I sell copies of that art in my shop, and make it clear to everyone I sell it to that I am in no way affiliated with Disney and this is totally unauthorized by Disney, I'm probably fine.
Trademarks are also industry specific. That's why Apple Records and Apple Computer both exist -- as long as a reasonable person could not confuse them, it's OK.
In short, trademarks are very very different from copyrights. They protect different activities.
In fact I should not have used the phrase trademarked work. A work (like an image or movie or novel or software program) does not get trademarked. The character, slogan, logo, product name, company name, brand name, color scheme, etc used therein to identify the brand, is what is trademarked. Very different.
I will add more examples, this time to illustrate copyright, which works basically the opposite : Suppose mickey mouse were not trademarked. Then while it would be illegal to redistribute verbatim copies of a recent Mickey mouse picture authored by Disney, as well as any modified remixed versions based on that verbatim picture, it would be perfectly legal to draw totally new art involving the same character as long as it was completely new without referring to the copyrighted work, because coypright protects the right of Disney to make money off distributing that picture they made, and they did not make or contribute to making your mickey drawing, and while you are using a character they came up with, in the absence of trademark, copyright isn't intended to protect the public from being confused as to who they are dealing with as trademark is.
IANAL this is based on decades of amateur interest in IP law.
This is not 100% accurate from a trademark perspective, at least with respect to "famous" marks.
Generally speaking, you are correct - unless there is a likelihood of consumer confusion, you are free to use a trademark already used by a senior user.
But marks like Apple and Mickey Mouse, from a trademark, are sufficiently famous that they get special protection. There is a concept called trademark dilution that only applies to sufficiently famous marks. With respect to such marks, a junior user can be liable for use of the mark even if there is no likelihood of confusion.
(BTW: By "senior" user, I means a user that gained trademark rights first and a "junior" user is one that started using the mark in commerce later.)
> status quo in the United States is that AI-generated images are not currently eligible for copyright.
Aren't they? I thought it was just that the copyright holder has to be a recognized legal entity (so, the copyright would have to belong to the human operator or their employer, not to the ai model itself).
> Last September, the US Copyright Review Board decided that an image generated using Midjourney’s software could not be copyright due to how it was produced.
Does this mean that if, for example, a court rules that I cannot train an image generation model on copyrighted material, I can train it on AI-generated images?
> I highly doubt Wizards of the Coast or whoever is going to want their premium products to lose copyright protections, so they'll need to keep paying artists.
WotC's latest round of layoffs (within the past month or so) hit the art staff especially hard.
> AI-generated images are not currently eligible for copyright
It’s a bit more nuanced than that. Here is the relevant policy statement, which notes that some AI-assisted works are potentially eligible for registration and have indeed been registered, while works that are primarily the product of an AI are not.
Makes sense. Photoshop has had content-aware-fill for over a decade. That counts as AI as much as any diffuser does. I don't think those images should have their copyright invalidated.
But then, if I generate an image using "AI" and touch it up in Photoshop, is it eligible for copyright again? How much "touch up" do I have to do for it to not be "AI generated"?
This has been problematic forever long even before any LLMs. Techniques like photobashing that use copyrighted images modify them for it to be new work. Or even older cases like work of Andy Warhol.
Where this is more extreme here is that there is no human labor involved there is no invention. On the other hand LLMs make this extra tricky because in one way output they create is objectively unique but subjectively/culturally it's not.
To answer your question if you take AI generated image and change it enough for it to stand as it's own unique thing you could for sure claim it as your work and it would be eligible for copyright.
This is a horrible direction to start off. Especially given that we can't truly tell if an image is AI generated or not. What if I modify an AI generated image?
There's so many technicalities here that can be weaponized.
One question that is unclear to me is how this works if images are packaged with text or other content. For example; let’s say I write a book and then use AI images to illustrate it. It doesn’t seem logical to me that somehow the book would be copyrighted but the images inside the book wouldn’t be…? At some level, the “package” of images + text would supersede the two things separately. Otherwise you would have a situation where sharing the book is a copyright violation but sharing the images inside of it isn’t.
There's nothing particularly contradictory about that: there's already situations where that is the case. For example when a book contains images which are public domain, or where elements of the book like the facts within it are not copyrightable. Another interesting one is tabletop game manuals: the layout and presentation of the rules are copyrightable, but the game mechanics generally aren't. So you can make a book which just contains the rules and not be infringing copyright. Using AI-generated images would be exactly the same situation.
I'm envisioning more of a situation where a company adds text directly to AI-generated images, or otherwise somehow modifies them that prevents them from just being generic images, in the way public domain images are. I really don't think companies will just add images straight-from-the-generator without modifying them in such a way that prevents their easy re-use.
I get why artists are trying to stop them, but this battle has already been lost. These tools have been out too long, open source models proliferate freely, and jurisdictions that don’t care about IP law will continue publishing these models. By the time it works it’s way through the courts the situation will be even worse.
You can publish these models in all jurisdictions, including ones that do care about IP law. There's no rulings saying that models trained on datasets of images are direct derivatives of the original images (in a way that's copyright-violating), let alone whatever is produced using these models. Now, this isn't settled law by any means, but it feels like people give too much credence to all the theft accusations by assuming that it's some slam dunk case that can make all generative AI suddenly vanish.
I used to agree with this, but there's been research coming out of Google that has altered my opinion. Specifically, Google's gotten rather good at making AI spit out unaltered training set data[0]. This is only possible if the AI is remembering large portions of the original trained-on works, which would make the weights infringing.
[0] In the most egregious case, they found that just asking ChatGPT to repeat a word over and over again will make it spit out training set data verbatim. OpenAI's response was to stop any conversation with over a page of repeated words, so you can't replicate this.
> This is only possible if the AI is remembering large portions of the original trained-on works
which is fine in my books. You could equivalently produce the same works from searching thru the digits of pi.
The only enforcement that's needed is on the end user of the model - if they choose to produce the training set, they are violating copyright. The creator of the model _does not_ violate copyright merely by creating and distributing the weights of a neural network, as long as the space of potential output is vastly larger than the training set.
>You could equivalently produce the same works from searching thru the digits of pi.
in the same way you can brute force an MD5 Hash if you had a few centuries, I guess. I don't think "monkey's making Shakespeare" is a good metric for when to determine if a piece of art is unique enough to be scraped.
Tracing is looked down upon in the art community and this feels way too close to tracing if they just have these images in a database ready to reference. You can't store copyrighted movies nor music on such private databases (one of the few times I will ever utter the words "thank you DMCA"), I don't see why art pieces would be exempt.
That's a really weird argument. Copyright is a legal system created by the Constitution and statutes and administrative rules. It cares about whether you are "copying", and it cares about whether you're creating things that compete with the works of the original authors. It doesn't care about potential output spaces.
In this context, I don't see a principled difference between the model weights and really good compression. If I send you a gzipped copy of the latest bestseller book it's still copyright infringement. And it would still be infringement if I shipped it inside a software program that can _also_ reshuffle the words in a bajillion different ways, if there's a "copy" of the original work in there.
Yes, but Google Books is a search engine. It doesn't write books, it just tells you where a particular phrase might occur in those books. There's explicit caselaw allowing you to do this, extending back before Google was even a thing. For related reasons, Google Books also does not let you read the whole book - just the page the search match came from.
OpenAI and other large language model developers are claiming they have a machine that can write books, but they also fed it shittons of books, and they can't account for where all that text went. At best they can say "well, it doesn't produce exact, verbatim copies of the training set all the time".
Generative AI doesn't write books either, it's a machine spitting out results based on queries entered by a person. It has no personhood nor rights to ownership, in spite of what AI more crackpot advocates try to claim, it's more a very complicated pencil.
If those HDF5 or .gguf or whatever file format turns out to be lossy decodable as images and texts, it shouldn't be considered kosher just like pirated and transcoded movies aren't considered lawful transformative original creations.
Courts won't buy this, at least not completely. Copyright infringement liability accrues to the entire value chain - model users and developers alike. So the only way for a model to see copyrighted data during training and be non-infringing is if there's no conceivable way to reproduce training data.
I don't think fair use saves us either, at least not the model authors, because the whole selling point of these models is to replace artists. Yes, artists can use them as tools, but that is far less lucrative. The valuations and hype being thrown around specifically come from, among other things, being able to cut the creative class out of their own business. No court is going to look at that and say "ok, yeah, sure, it's perfectly fine for you to be using other people's work to train on".
An individual using an AI art generator may still wind up getting novel output that isn't obviously infringing to anything in the dataset. If that's the case then they probably haven't infringed copyright. Or at least, it'd be difficult to make a case around it.
I'm kind of confused - you've claimed that Google's AI was broken, but cite an anecdote over ChatGPT?
Regardless, even if these cases did happen often enough, it's erroneous to assume that this is something universal (i.e. can be applied to all generative AI models) or intentional. Said models are vastly smaller than the sizes of their training datasets, so it's more or less impossible for all the data to be stored verbatim. Some aspects may appear to look like memorization if the piece of data reappears many times in the dataset - in those cases, the algorithm has a really strong incentive to recite that data. This reduces the effectiveness and is considered an artifact that needs to be corrected, not some underlying idea of AI.
The way these algorithms work is public knowledge, there's not really any black boxes in the hands of OpenAI that would be relevant here. Considering that, I'm really doubtful that these claims can be easily supported.
There has been successful prompt engineering attempts to force GPT-based models to regurgitate original dataset texts texts texts texts texts texts as for example Wikipedia is a free-content online encyclopedia, written and maintained by a community of volunteers, collectively known as Wikipedians, through open colla--
I think you’re a bit confused—these models are trained by copying works, and the memorization is the incontrovertible proof. Imagine if you download a font and use it to make a logo without paying to license the font.
You appear to be refuting a slightly different point, though. When OpenAI was making Dall-E 2, they found that duplicates in the training set would incentivize memorization of specific images, as you said. This is more like finding a secret cheat code that would make the model regurgitate everything it had remembered, regardless of how "incentivized" it is to do so or even if it had been aligned to not do that.
My personal argument is this: the primary metric that the training process attempts to minimize is perplexity. This is how good the model is at guessing the training set data. Base models are specifically being designed to compress huge amounts of text and we just so happen to accidentally get a decent word calculator out of it. The alignment fine-tuning that happens later adjusts the model to prefer answering questions, but the underlying memorized data is still there.
>The way these algorithms work is public knowledge, there's not really any black boxes in the hands of OpenAI that would be relevant here.
Nope. Modern GPT is entirely a black box. OpenAI stopped publishing model weights the moment Elon Musk stopped writing the checks. Hell, they don't even publish the model architecture anymore. How GPT-4 works, even on a basic "this is how many transformer layers and attention heads we're using" basis, is a trade secret.
Even if you have model weights, the actual meaning of the learned model parameters has never been known; there is active research on figuring them out. One particular problem is polysemanticity. If you look inside a particular hidden layer, you won't see a single "dog" or "cat" neuron in its hidden layers. You'll have a 512-dimension concept bouillabaisse with "dog", "cat", "bird", "guinea pig", "kangaroo", and so on all floating around whatever shape made sense at training time (even if it implies absurdities like "desk is the opposite of loin cloth"). To untangle this, you have to train another AI to pick out monosemantic clusters of neurons that can then be inspected, and that requires extreme amounts of GPU resources.
Lockpicking tools are widely available, and designs of locks can be found on the web if someone search hard enough. But entering another's home is illegal. It's why businesses pay for licenses even if cracked softwares exists. I hope that artists win these and make using or creating an illegal model an high-risk activity, not worth it for any commercial activity.
A society without art is a society without joy or self awareness. A society with only AI created art is a facsimile of joy or self awareness. If we're trying to remove the parts of society that challenge and inspire us then we'll be left with cultural mush that doesn't challenge, critique, explore, or even play.
I don’t think the current tools are doing much to the creation of “art”, it’s mostly replacing run of the mill graphic design work like logos and ads and cover art, or maybe some higher end stuff like automating workflows.
If anything, the AI image-generation movement proves just how highly society values the work of the artist, since we now have tech companies dumping hundreds of millions of dollars into training models to approximate their work.
Compared to many tech people here who maintain a website page for 6 figure salaries?
few of us are saving the world. And it turns out saving the world doesn't pay well either (just look at teacher salaries. I guess educating the future doesn't benefit society). Always found it such a weird angle to demean artists while literal billon dollar corporations profit off of exploiting them. So it's not like people don't value certain kinds of art.
I don't even care about it being illegal. As said above, US can completely ban such tech and China will keep on trucking along.
I just want artists affected to be paid, and actually paid. Not "paid" the way spotify artists are. if you're making a billion off of 100 artists' work, you better be making each affected artist a millionaire in royalties.
I'm a bit confused - you draw parallels between artists whose work gets used in a training dataset and musicians who sign contracts with Spotify, but the comparison is very strained. The financial link with Spotify is obvious - the company pays some minuscule amount of money per use to directly use a product of an artist, in part for their own profit.
With AI, the situation is far from being that clear-cut. For one, the majority of datasets aren't comprised of capital-A Art with specific authors and attributions - a ton of it is just random information from around the internet, or just downright junk data. A photo I posted for free could conceivably end up in that dataset. Then, when a model is trained on that dataset and that model is used to make other outputs, are those outputs actually derivative with a direct link to some origin? If I generate a landscape using an AI, to whom exactly would I even owe money? Equally to all contributors in a dataset?
>For one, the majority of datasets aren't comprised of capital-A Art with specific authors and attributions - a ton of it is just random information from around the internet, or just downright junk data.
Random means there can be copyrighted content in it. It shouldn't be "random". You can't do such things when making your own piece of media, so I don't see why it's okay when it's a bot scraping massive pieces of the internet for use in some for-profit venture. Scraping was already this gray area in the internet and practices like these argue against such techniques.
>Then, when a model is trained on that dataset and that model is used to make other outputs, are those outputs actually derivative with a direct link to some origin?
That's the million dollar question (literally. Or more like 10/100 million). I imagine like existing copyright it depends on how far out you derive. Arguments above that suggest the ability to regurgitate the exact or very close to piece of data would certainly lean towards "yes". In which case it implies they have that piece of data in their database, instead of just being a "reference".
That feels too close hosting off of copyrighted material in my eyes, but I don't have the full picture of how and what these LLMs contain.
>If I generate a landscape using an AI, to whom exactly would I even owe money?
Like any media creator/producer would tell you: depends on thr license. For example, I Google "mountain" and my first result is this :
License says "free to use under the unsplash license", which is surprisingly lax. But there are two stipulations
Photos cannot be sold without significant modification.
Compiling photos from Unsplash to replicate a similar or competing service.
So we come back to the whole issue in the beginning with the first point. The 2nd point is a much more nebulous one to consider but not to be completely dismissed (maybe an AI service can be argued to generate a competitor).
And that's one of the more generous results. Another result came from CNN but sourced from Shuttershock. Here's their license :
I'm not going to read nor summarize the license, but I'm sure CNN at the bare minimum needed to ask permission and likely negotiate to use it in their news post.
So yeah, an "ethical" training model will be doing this for every single image they use, if the aren't creating nor taking the pictures themselves.
>Equally to all contributors in a dataset?
Don't know. That would be for the courts to decide. If it's anything like streaming music and movies, you would be compensated proportionally to the amount of "usages" your work gets when generating pieces. Doesn't necessarily have to be equal, but it may surprisingly equalize out since algorithms are at the helm and not brands trying to stand out.
It's certainly be it's own rabbit hole to explore though. If we ever get that far.
The difference is that AI generation tools provide a lot of potential economic value, while breaking in to someones house doesn't. There will be a lot of pressure for governments to allow this to go on since it puts them at advantage over other countries.
This is a really poor argument because virtually everyone can be presented an argument where they would act counter to this way of thinking, at which points it’s only “the end does not justify the means but only when I say so”
Yup, and even if the law decides against using copyrighted images as training data, companies such as Adobe are already ahead of the curve with generative systems like firefly which has been trained exclusively on licensed artwork.
It is a misconception that Adobe's models have not been trained on copyrighted work. Nobody should be repeating their marketing claims.
Adobe has not shown how they train the text encoders in Firefly, or what images were used for the text-based conditioning (i.e. "text to image") part of their image generation model. They are almost certainly using CLIP or T5, which are trained on LAION2b, an image dataset with the very problems they are trying to address, C4 (a text dataset similarly encumbered) and similar.
I welcome anyone who works at Adobe to simply answer this question of how they trained the text encoders for text conditioning and put it to rest. There is absolutely nothing sensitive about the issue, unless it exposes them in a lie.
So no chance. I think it's a big fat lie. They'd have to have made some other scientific breakthrough, which they didn't.
It's certainly not impossible, but it's impracticable. On 248m images (roughly the size of Adobe Stock), CLIP gets 37% on ImageNet, and on the 2000m from LAION, it performs 71-80%. And even with 2000m images, CLIP is substantially worse performing than the approach that Imagen uses for "text comprehension," which relies on essentially many billions more images and text tokens.
As well: the desire to have art seen and recognized (or even monetized) puts it in the public eye in a way that can be used (legally or not) to train AIs. Going the opposite way means committing to not showing your work to humans either.
Looking at the source tweet[1], this isn't a list of artists specifically used to train Midjourney, but instead is a list of artists used to pre-render additives that can be added to an existing prompt much faster than an arbitrary addition.
The assumption is that these artists probably are in the Midjourney dataset, but Midjourney doesn't specifically go out and find these artists' art and steal it—they steal art indiscriminately.
Yeah I was wondering what it meant to prerender them but yeah this has little to do with the actual training dataset, it's really just a list of artists. The assumption is that people are going to use them frequently independent if it has an effect or not.
Reading the context in the linked Twitter, this is not actually the list used to train.
After the model was already trained, they scraped this list of artist names off Wikipedia and did some optimizations so that prompts using them would render faster.
The actual model was trained on everything, whether it had a Wikipedia page or not.
This is going to be completely ineffective in the long run. There is a lot of public-domain art, especially for more historical images.
And for the newer styles and for the general image "vocabulary", the model trainers can just buy enough samples legitimately. I'm surprised that Getty Images is not pushing their collection to trainers.
I do find it entertaining to watch the creative world struggle against this unstoppable force. It's the Luddites and the mechanical knitting machines of the early 19th century all over again. The human race never changes. It will undoubtedly end the same way.
It's not the same thing. These are people who strived their whole lives to learn to draw and paint and make something of their own. It's not just their livelihood being stolen, it's their own work. They aren't saying not to use generative AI to make art, they are saying they don't want their art to be used to train it.
Artists are still needed to make input for AI, they aren't completely replaced by a machine.
It's really easy, "unstoppable" even, to send pirated music files over the internet. It sure was fun to have free music. And yet Napster was sued out of existence, and now the vast majority of people who listen to music pay for it.
And it's way harder to run a pirate AI datacenter than it is to set up a torrent tracker.
Laws will be written, license fees will be paid, and AI will continue to develop but with proper compensation.
The problem here is that the law has already been written to some extent. Google Books is the most relevant case you can point to that set precedent over what constitutes fair use of copyrighted material and the extent the transformation of those works can be commercially exploited. In that case works were copied without permission with small chunks of those works redistributed verbatim, and that was when it was working as designed. There's an argument to be made that this AI generator stuff is actually less egregious.
I don't think it's the same this time. A mechanical knitting machine can be invented in isolation. Midjourney could not have existed without an extensive database of stolen art. Isn't that something worth fighting against, for artists?
Some said that Napster/KaZaA/etc. was the future of music. They got closed down.
Spotify eventually popped up, and "solved" a big problem in music streaming. But at a cost - the record labels have them by their balls.
You don't see any leading "pirate" Spotify clone dominating the market. Closest thing is YouTube, but monetization goes to the content owners.
So here's what you do, if you're a bunch of creative artists / content owners / etc.:
You find one company that's commercializing the tech, and sue them into oblivion. Force them into signing a deal that ensures they'll never be profitable.
Others will follow.
But, yes, obviously the models are out there. In the same way that people are pirating stuff like they've always done. You can't do much about individuals running models on their own rigs - but that's not the point.
The point is to make it as painful as possible for people that feel the urge to commercialize something at scale. The goal is "Fuck you, pay me" or get dragged through the courts.
My personal worry is that now that companies can't simply scrape all the data for AI, they will instead start buying up massive amounts of data, turning the companies even more powerful and monolithic. Megacorps, here we come.
100%, anything outside of reducing IP rights is going to make this WORSE for artists, creatives and culture. Artists will campaign for the very expansions of IP rights that will make things so much worse. A world where styles can be owned and sold is not one I want to live in.
>The 24-page list of artists’ names used by Midjourney as the training foundation for its AI image generator
I don't see how they reached the conclusion that this list is "the training foundation" of Midjourney. I feel like this headline is misleading, perhaps even outright false.
What is the difference between AI scanning existing art, and creating new images informed by the prior art, and a person touring galleries, looking at the works of living artists, and creating new works influenced in style, color, framing, etc. by those works? Should the works of a new artist be beholden to all those whose works they have looked upon? As another thing to consider, is there any rationale for copyrights on music (in particular) that wouldn't equally apply to the recipes food dishes created by cooks? Both build upon the works, style, ingredients, etc. of those who preceded them. Seriously interested in the arguments one would make for copyrighting music that wouldn't stand equally well for recipes.
> The list of 16,000 artists was included as part of a lawsuit amendment to a class-action complaint targeted at Stability AI, Midjourney, and DeviantArt
Why is OpenAI not being targeted? I find it hard to believe that they trained DALL-E exclusively off free material. Is it simply because they havent publicised/leaked their datasets?
143 comments
[ 3.1 ms ] story [ 244 ms ] thread[0] In the Dune universe, the "Butlerian Jihad" refers to a legal ban on thinking machines.
[1] https://en.wikipedia.org/wiki/Monkey_selfie_copyright_disput...
That said, it seems like a moot point to me. The practical uses of generative AI are not going to be one-and-done prompt-to-image tools. When AI is used more like a brush, the brush strokes the human chooses will still be granted copyright.
Their guidance would not apply for someone using AI as a tool, but said copyright would be thinner than if you'd drawn everything by hand. Specifically, you don't own any of the things the AI "just came up with". If you just wrote a prompt and grinded out some results, you probably own nothing[0]. If you use shittons of inpainting to control, say, the overall composition, but the AI filled in pixels somewhere, then you probably still own the overall image, but that's only because I'm not sure how you'd separate the two in a way that would let you copy just the AI-generated portion. Or, in the case of the comic book I mentioned earlier, they own the text, characters, and plot of the comic book, but not the artwork.
[0] Yes, you could probably just lie to the Copyright Office. Make sure to never reveal your use of AI to anyone, because there's loads of angry artists who would love to tattle on you.
Steam has a “no AI art” policy, and it’s rapidly turning into a “no obvious AI art policy”. How could they tell?
There's also another advantage to having a "no obvious AI art" policy; and that's to cut down on spam. AI is extremely useful to people who want to spam art platforms.
Less obviously, ML models also lack the ability to bake in intent. In human made pieces, everything is as it is for a reason; it’s communicating something. In ML generated pieces, things are the way they are because that’s what’s statistically likely for the type of generated image.
With art, it's more subtle, because there there's no single reference point that lets us determine if an artwork is "true". There are the glaring errors that everyone can agree on - notoriously, human hands - but those cases are improving rapidly.
The gap of being distinguishable from manually drawn images is still closing - we don't know if it'll ever reach the threshold, but the amount of effort required to stamp out all the wonkiness from an AI generation has been going down ever since the first viable algorithms appeared.
I don't think that this was an anti-spam policy - Steam already manually reviews all new applicants that want to publish a game, so they don't need to forbid anything to turn it down. I'm guessing that this policy was because they don't want to be entangled in IP legislation if some copyright exception is carved out to forbid the use of generative AI.
Human artists might have to become used to tracking provenance. If you work with traditional media, that's easy: Here's the painting. For digital artists, software can publish encrypted, timestamped brushstroke-level histories of the work if we need that level of proof.
People have been trumpeting this since day one of Stable Diffusion releasing, but I'm seeing the same output quality as that day and I've been keeping up.
The reason why it's not as visible with Stable Diffusion is because a lot of the technologies around it circle the same few foundational SD models - people build on top of them, add new ways of interacting with them, but ultimately, the same thing underlies them all. Community support is seen as more important than cutting-edge tech, which is why something like Stable Diffusion XL hasn't even seen universal adoption yet.
As an aside, no one said AI couldn't achieve drawn generated text, that's been possible for years prior to stable diffusion.
Part of becoming an artist is learning how to evaluate your own work, break it down, and critique the shit out of it. When you jump straight to generating art with an AI, you skip the criticism step, which means you don't have a sense of taste and you haven't really explored what your preferences for style are.
A lot of AI art generators default to an extremely cinematic, "Hollywood" art style - i.e. exactly the sort of thing that is trying to look impressive to people who don't know better, and will make them overlook all the fundamental mistakes in the image.
[0] Normally I wouldn't scarequote "artist" here, given that actual artists do use AI tools where it makes sense.
Microsoft and every other tech company is indeed investing billions in the tech. I'm sure each company can fund the entire (woefully underpaid) art industry by themselves, let alone the 10 or so tech hubs altogether.
But they are happy to throw money at AI instead for the payoff of being the next big tech brand.
That's basically the argument why Jason Allen should have been allowed to win the art competition, is it?
It's not that he typed "award winning painting" into Midjourney and the image was the result.
He tried hundreds of seeds, selected one that he liked and refined it over countless iterations with infilling until he was satisfied with the result.
I honestly don't see how this is fundamentally different from other art forms.
They don't allow ai art produced by models trained on material that the model makers don't have copyright to.
In practice that's a ban (currently) but in principle it isn't.
Complete wrong. You just flip the defaults--something is AI unless you can prove otherwise.
This is done already and has precedent. Producing porn requires that you keep artifacts demonstrating that who the performers were, that they were of age, etc.
If you claim a work is not AI generated, you should have to produce some artifacts to back up that claim.
In the case of a corporation, that would be easy as you have payment records.
In the case of an individual doing digital only, that's a little harder. You probably have to keep some intermediate artifacts.
So AI (or a simpler, non-statistics-based algorithm) can't produce these artifacts and will never be able to? Why? Are these artifacts "human souls" or something?
> In the case of a corporation, that would be easy as you have payment records.
Outsourcing.
Besides, it's not that easy of a problem - a lot of corporate artists are salaried workers, they don't get specified commissions with an attached bill per work, but are paid a salary so the company can ask them to draw whatever they need throughout the process. Considering this, all artists would need to retain "intermediate artifacts".
And then, how do these artifacts work for other ways of doing art? What about traditional artists whose work gets scanned in after completion - would they have to keep a camera on hand to take photographs as they're working? What about an animated film - would every intermediate step in production, from character design to storyboarding to environmental design etc etc need to have a full record for every single sketch?
Keep in mind that in the not so far future, producing art will be as cheap as consuming it, this means that the original benefits society got in return for copyright no longer applies, so why should they protect it?
I'm not sure if it's that simple - for one, this requirement is a complete departure from how copyright systems work now. Providing complete history logs isn't normal practice, and expanding law to necessitate it isn't common sense.
> Keep in mind that in the not so far future, producing art will be as cheap as consuming it, this means that the original benefits society got in return for copyright no longer applies, so why should they protect it?
I'll make a prediction that this future is further from now than you may think it is. Sure, things like static imagery may become completely indistinguishable from human-made art in the near(ish?) future, but the production of all art is still an unsolved problem. How long will it take until some advanced multimodal algorithm can make a full game that can measure up to ones that are released today? I'm guessing that it'll take a while.
And yeah - once we do reach this scenario of hypothetical "art post-scarcity", we may as well just delete the whole copyright system from existence - it'd be a logical thing to do. But how does any of it contradict what I said in my other comments?
https://www.copyright.gov/forms/formva.pdf
Registration is not a rubber-stamp, it is increasingly refused because of indicia of AI tooling.
Why would adding some questions on provenance and methodology be beyond the pale?
> it is increasingly refused because of indicia of AI tooling.
Do you have a source that a statistically significant number of copyright applications gets refused on account of a work just seeming like AI? On what grounds does it get refused?
https://www.federalregister.gov/documents/2023/03/16/2023-05...
We can disagree on how long it will take us to get there, but if you use AI generated content, that is not product by copyright, your game as whole, sure, as long as it is not the result of a simple prompt, you're protected as usual.
Keep in mind that already, in many games, there is a mix of protected and unprotected content, for reasons of trademark, copyright, and licensing.
I love when non-artists talk about art.
If it's digital media, the software can keep an encrypted record at the brushstroke level that can be played back to produce a bit-perfect reproduction. Maybe even write it to a public ledger.
For digital media - ignoring how overbearing this whole system could be, what prevents someone from taking all that data and making an algorithm that outputs brush stroke parameters instead of pixels? And digital art isn't the only thing we need to concern ourselves with - eventually, we might have AI models that could make 3D models, sounds, vector imagery and other forms of art. The idea of just documenting every workflow would be an ever-growing burden with no perfect solutions.
Um, that's a real work, you know? In what way does this differ from people who take a photograph and then, for example, creating an oil painting?
Now, there are some weirdnesses because of the copyright of the source photograph, but the oil painting would be your own work.
Yeah, you might get called into court to demonstrate that you can produce the work. But so did Michael Jackson.
Generally the party asserting ownership has the burden of proof. The standard is "preponderance of the evidence", which generally is understood to mean "more likely than not" or "> 50%". So basically it means if you can prove to judge or jury that there's a >50% chance you own the work, it's good enough.
Also note that in many cases where there's a dispute over the evidence, witnesses are summoned to testify. So you might not have a "full notary log" of how it was produced or all the "intermediate artifacts", but as long as the artist is able to convincingly explain how the work was created, and the other party's lawyers are not able to poke holes in their story during cross evidence, that's usually enough.
Which is, basically, what happens today, if the authorship or ownership of a work is disputed.
That said, I'm not sure whether "assume work AI (thus uncopyrightable) unless proven otherwise" should be the default for other reasons. For one, most quality "AI art" needs some manual adjustments or touch ups, and arguably the prompt and hyperparameters may be sufficient creativity element. I mean, that's basically how copyrights dealt with photography (the mere fact you decided when and where to point the camera with what settings is sufficient for copyright to subsist in a photo).
Why would that be relevant in court? Just show the process of making the art.
I think this is rather what pro-AI/spammers are trying to do by flooding platforms, that aren't so successful. People don't give as high scores they do for human generated data, and AI images are still considered a form of spam.
The reality is that most legal things are determined by _convincing people of a truth_. Perhaps you can set up a whole scheme to "launder" AI art and attach names to them. And all the papertrail you generate doing this will show up in discovery in some lawsuit and the copyrights all disappear.
Laws are vibes, not code.
Art will then take less skill to produce, and be produced faster for lower prices.
An 80% price reduction on art (because artists can now produce it 5 times faster thanks to AI) is 80% as good as getting it for free.
this is enough to use them without copyright violation for example for other ai models training.
Basically the same way Disney let copyright go but will fight for trademark to the bitter end ?
So I don't know if you could apply trademark to e.g., every card in magic, but maybe only to the key characters?
1: https://www.mekiplaw.com/how-to-trademark-a-character-an-eas....
Trademark is intended to protect the holder from being impersonated, not from losing revenue from selling content.... So it's a lot easier to redistribute copies of trademarked work as long as you make it clear you are not affiliated with the trademark holder, in a manner which a reasonable person would heed.
So for example, if a piece of art is trademarked by Disney, and it is well known by the public, and I print a copy and put it in front of my shop, a reasonable observer might this my shop is owned, operated, or endorsed by Fisney. So that's not OK.
If instead I sell copies of that art in my shop, and make it clear to everyone I sell it to that I am in no way affiliated with Disney and this is totally unauthorized by Disney, I'm probably fine.
Trademarks are also industry specific. That's why Apple Records and Apple Computer both exist -- as long as a reasonable person could not confuse them, it's OK.
In short, trademarks are very very different from copyrights. They protect different activities.
In fact I should not have used the phrase trademarked work. A work (like an image or movie or novel or software program) does not get trademarked. The character, slogan, logo, product name, company name, brand name, color scheme, etc used therein to identify the brand, is what is trademarked. Very different.
I will add more examples, this time to illustrate copyright, which works basically the opposite : Suppose mickey mouse were not trademarked. Then while it would be illegal to redistribute verbatim copies of a recent Mickey mouse picture authored by Disney, as well as any modified remixed versions based on that verbatim picture, it would be perfectly legal to draw totally new art involving the same character as long as it was completely new without referring to the copyrighted work, because coypright protects the right of Disney to make money off distributing that picture they made, and they did not make or contribute to making your mickey drawing, and while you are using a character they came up with, in the absence of trademark, copyright isn't intended to protect the public from being confused as to who they are dealing with as trademark is.
IANAL this is based on decades of amateur interest in IP law.
Generally speaking, you are correct - unless there is a likelihood of consumer confusion, you are free to use a trademark already used by a senior user.
But marks like Apple and Mickey Mouse, from a trademark, are sufficiently famous that they get special protection. There is a concept called trademark dilution that only applies to sufficiently famous marks. With respect to such marks, a junior user can be liable for use of the mark even if there is no likelihood of confusion.
(BTW: By "senior" user, I means a user that gained trademark rights first and a "junior" user is one that started using the mark in commerce later.)
Aren't they? I thought it was just that the copyright holder has to be a recognized legal entity (so, the copyright would have to belong to the human operator or their employer, not to the ai model itself).
> Last September, the US Copyright Review Board decided that an image generated using Midjourney’s software could not be copyright due to how it was produced.
WotC's latest round of layoffs (within the past month or so) hit the art staff especially hard.
Some of the lists, such as on Reddit, appear to (erroneously?) list a few artists who advertise themselves as still employed by WotC. [2]
[1] https://twitter.com/CHofferCBus/status/1734947730491932929
[2] https://www.reddit.com/r/dndnext/comments/18ij198/list_of_kn...
https://www.federalregister.gov/documents/2023/03/16/2023-05...
Where this is more extreme here is that there is no human labor involved there is no invention. On the other hand LLMs make this extra tricky because in one way output they create is objectively unique but subjectively/culturally it's not.
To answer your question if you take AI generated image and change it enough for it to stand as it's own unique thing you could for sure claim it as your work and it would be eligible for copyright.
There's so many technicalities here that can be weaponized.
That would really be a great way to structure things.
There is another bright side of it. This images can be used for AI training without copyright violation. Does this apply to texts as well?
[0] In the most egregious case, they found that just asking ChatGPT to repeat a word over and over again will make it spit out training set data verbatim. OpenAI's response was to stop any conversation with over a page of repeated words, so you can't replicate this.
which is fine in my books. You could equivalently produce the same works from searching thru the digits of pi.
The only enforcement that's needed is on the end user of the model - if they choose to produce the training set, they are violating copyright. The creator of the model _does not_ violate copyright merely by creating and distributing the weights of a neural network, as long as the space of potential output is vastly larger than the training set.
in the same way you can brute force an MD5 Hash if you had a few centuries, I guess. I don't think "monkey's making Shakespeare" is a good metric for when to determine if a piece of art is unique enough to be scraped.
Tracing is looked down upon in the art community and this feels way too close to tracing if they just have these images in a database ready to reference. You can't store copyrighted movies nor music on such private databases (one of the few times I will ever utter the words "thank you DMCA"), I don't see why art pieces would be exempt.
In this context, I don't see a principled difference between the model weights and really good compression. If I send you a gzipped copy of the latest bestseller book it's still copyright infringement. And it would still be infringement if I shipped it inside a software program that can _also_ reshuffle the words in a bajillion different ways, if there's a "copy" of the original work in there.
OpenAI and other large language model developers are claiming they have a machine that can write books, but they also fed it shittons of books, and they can't account for where all that text went. At best they can say "well, it doesn't produce exact, verbatim copies of the training set all the time".
I don't think fair use saves us either, at least not the model authors, because the whole selling point of these models is to replace artists. Yes, artists can use them as tools, but that is far less lucrative. The valuations and hype being thrown around specifically come from, among other things, being able to cut the creative class out of their own business. No court is going to look at that and say "ok, yeah, sure, it's perfectly fine for you to be using other people's work to train on".
An individual using an AI art generator may still wind up getting novel output that isn't obviously infringing to anything in the dataset. If that's the case then they probably haven't infringed copyright. Or at least, it'd be difficult to make a case around it.
Regardless, even if these cases did happen often enough, it's erroneous to assume that this is something universal (i.e. can be applied to all generative AI models) or intentional. Said models are vastly smaller than the sizes of their training datasets, so it's more or less impossible for all the data to be stored verbatim. Some aspects may appear to look like memorization if the piece of data reappears many times in the dataset - in those cases, the algorithm has a really strong incentive to recite that data. This reduces the effectiveness and is considered an artifact that needs to be corrected, not some underlying idea of AI.
The way these algorithms work is public knowledge, there's not really any black boxes in the hands of OpenAI that would be relevant here. Considering that, I'm really doubtful that these claims can be easily supported.
You appear to be refuting a slightly different point, though. When OpenAI was making Dall-E 2, they found that duplicates in the training set would incentivize memorization of specific images, as you said. This is more like finding a secret cheat code that would make the model regurgitate everything it had remembered, regardless of how "incentivized" it is to do so or even if it had been aligned to not do that.
My personal argument is this: the primary metric that the training process attempts to minimize is perplexity. This is how good the model is at guessing the training set data. Base models are specifically being designed to compress huge amounts of text and we just so happen to accidentally get a decent word calculator out of it. The alignment fine-tuning that happens later adjusts the model to prefer answering questions, but the underlying memorized data is still there.
>The way these algorithms work is public knowledge, there's not really any black boxes in the hands of OpenAI that would be relevant here.
Nope. Modern GPT is entirely a black box. OpenAI stopped publishing model weights the moment Elon Musk stopped writing the checks. Hell, they don't even publish the model architecture anymore. How GPT-4 works, even on a basic "this is how many transformer layers and attention heads we're using" basis, is a trade secret.
Even if you have model weights, the actual meaning of the learned model parameters has never been known; there is active research on figuring them out. One particular problem is polysemanticity. If you look inside a particular hidden layer, you won't see a single "dog" or "cat" neuron in its hidden layers. You'll have a 512-dimension concept bouillabaisse with "dog", "cat", "bird", "guinea pig", "kangaroo", and so on all floating around whatever shape made sense at training time (even if it implies absurdities like "desk is the opposite of loin cloth"). To untangle this, you have to train another AI to pick out monosemantic clusters of neurons that can then be inspected, and that requires extreme amounts of GPU resources.
few of us are saving the world. And it turns out saving the world doesn't pay well either (just look at teacher salaries. I guess educating the future doesn't benefit society). Always found it such a weird angle to demean artists while literal billon dollar corporations profit off of exploiting them. So it's not like people don't value certain kinds of art.
I just want artists affected to be paid, and actually paid. Not "paid" the way spotify artists are. if you're making a billion off of 100 artists' work, you better be making each affected artist a millionaire in royalties.
With AI, the situation is far from being that clear-cut. For one, the majority of datasets aren't comprised of capital-A Art with specific authors and attributions - a ton of it is just random information from around the internet, or just downright junk data. A photo I posted for free could conceivably end up in that dataset. Then, when a model is trained on that dataset and that model is used to make other outputs, are those outputs actually derivative with a direct link to some origin? If I generate a landscape using an AI, to whom exactly would I even owe money? Equally to all contributors in a dataset?
Random means there can be copyrighted content in it. It shouldn't be "random". You can't do such things when making your own piece of media, so I don't see why it's okay when it's a bot scraping massive pieces of the internet for use in some for-profit venture. Scraping was already this gray area in the internet and practices like these argue against such techniques.
>Then, when a model is trained on that dataset and that model is used to make other outputs, are those outputs actually derivative with a direct link to some origin?
That's the million dollar question (literally. Or more like 10/100 million). I imagine like existing copyright it depends on how far out you derive. Arguments above that suggest the ability to regurgitate the exact or very close to piece of data would certainly lean towards "yes". In which case it implies they have that piece of data in their database, instead of just being a "reference".
That feels too close hosting off of copyrighted material in my eyes, but I don't have the full picture of how and what these LLMs contain.
>If I generate a landscape using an AI, to whom exactly would I even owe money?
Like any media creator/producer would tell you: depends on thr license. For example, I Google "mountain" and my first result is this :
https://unsplash.com/photos/aerial-photography-of-mountain-r...
License says "free to use under the unsplash license", which is surprisingly lax. But there are two stipulations
So we come back to the whole issue in the beginning with the first point. The 2nd point is a much more nebulous one to consider but not to be completely dismissed (maybe an AI service can be argued to generate a competitor).And that's one of the more generous results. Another result came from CNN but sourced from Shuttershock. Here's their license :
https://www.shutterstock.com/license
I'm not going to read nor summarize the license, but I'm sure CNN at the bare minimum needed to ask permission and likely negotiate to use it in their news post.
So yeah, an "ethical" training model will be doing this for every single image they use, if the aren't creating nor taking the pictures themselves.
>Equally to all contributors in a dataset?
Don't know. That would be for the courts to decide. If it's anything like streaming music and movies, you would be compensated proportionally to the amount of "usages" your work gets when generating pieces. Doesn't necessarily have to be equal, but it may surprisingly equalize out since algorithms are at the helm and not brands trying to stand out.
It's certainly be it's own rabbit hole to explore though. If we ever get that far.
The end does not justify the means.
https://docs.wto.org/dol2fe/Pages/FE_Search/FE_S_S006.aspx?M...
Adobe has not shown how they train the text encoders in Firefly, or what images were used for the text-based conditioning (i.e. "text to image") part of their image generation model. They are almost certainly using CLIP or T5, which are trained on LAION2b, an image dataset with the very problems they are trying to address, C4 (a text dataset similarly encumbered) and similar.
I welcome anyone who works at Adobe to simply answer this question of how they trained the text encoders for text conditioning and put it to rest. There is absolutely nothing sensitive about the issue, unless it exposes them in a lie.
So no chance. I think it's a big fat lie. They'd have to have made some other scientific breakthrough, which they didn't.
Using information from https://openai.com/research/clip and https://github.com/mlfoundations/open_clip, it's possible to investigate the likelihood that using just their stock image dataset, can they make a working text encoder?
It's certainly not impossible, but it's impracticable. On 248m images (roughly the size of Adobe Stock), CLIP gets 37% on ImageNet, and on the 2000m from LAION, it performs 71-80%. And even with 2000m images, CLIP is substantially worse performing than the approach that Imagen uses for "text comprehension," which relies on essentially many billions more images and text tokens.
The assumption is that these artists probably are in the Midjourney dataset, but Midjourney doesn't specifically go out and find these artists' art and steal it—they steal art indiscriminately.
[1] https://x.com/JonLamArt/status/1741545927435784424?s=20
https://web.archive.org/web/20231231203837/https://docs.goog...
After the model was already trained, they scraped this list of artist names off Wikipedia and did some optimizations so that prompts using them would render faster.
The actual model was trained on everything, whether it had a Wikipedia page or not.
Won't matter though, because in modern times ragebait is more important than truth.
And for the newer styles and for the general image "vocabulary", the model trainers can just buy enough samples legitimately. I'm surprised that Getty Images is not pushing their collection to trainers.
Artists are still needed to make input for AI, they aren't completely replaced by a machine.
And it's way harder to run a pirate AI datacenter than it is to set up a torrent tracker.
Laws will be written, license fees will be paid, and AI will continue to develop but with proper compensation.
Well not really, it's just freely available online.
Spotify eventually popped up, and "solved" a big problem in music streaming. But at a cost - the record labels have them by their balls.
You don't see any leading "pirate" Spotify clone dominating the market. Closest thing is YouTube, but monetization goes to the content owners.
So here's what you do, if you're a bunch of creative artists / content owners / etc.:
You find one company that's commercializing the tech, and sue them into oblivion. Force them into signing a deal that ensures they'll never be profitable.
Others will follow.
But, yes, obviously the models are out there. In the same way that people are pirating stuff like they've always done. You can't do much about individuals running models on their own rigs - but that's not the point.
The point is to make it as painful as possible for people that feel the urge to commercialize something at scale. The goal is "Fuck you, pay me" or get dragged through the courts.
I don't see how they reached the conclusion that this list is "the training foundation" of Midjourney. I feel like this headline is misleading, perhaps even outright false.
3rd Para: 24 page list.
Yeah, something is off here.
Why is OpenAI not being targeted? I find it hard to believe that they trained DALL-E exclusively off free material. Is it simply because they havent publicised/leaked their datasets?