Getty Images is suing the creators of Stable Diffusion
theverge.com
theverge.com
These words all provide a reference point to help humans understand the technology, but regardless of any apparent similarities, SD does not “see” the way humans see, or become “inspired” the way humans are inspired.
I suspect that if humans had the ability to operate on billions of images to systematically gain “inspiration”, the state of copyright law and rules around fair use would not look the way they do today. The fact that artists are all on a relatively level playing field, innate talents notwithstanding, seems at the core of the historical precedents set by artists drawing inspiration from each other.
I’d argue that generative AI isn’t adding a new player to that game, but altering that game entirely.
All of the more simplistic analogies seem similar to claiming that automobiles are really the same as bicycles, just faster. Except automobiles resulted in major shifts in both behavior and policy, and there’s a reason people don’t ride bicycles on the autobahn.
Generative AI is transformative and exciting tech, but we need to stop pretending it’s human, if only because it’s important that policy makers don’t see this technology as human-like intelligence.
That argument may benefit ethically questionable models in the short term, but seems legitimately dangerous when considering other emerging AI use cases and the possibility that AIs will be handling decision making for critical systems in the near future.
>claiming that automobiles are really the same as bicycles, just faster. Except automobiles resulted in major shifts in both behavior and policy, and there’s a reason people don’t ride bicycles on the autobahn.
But speed (and weight) is the key difference. If cars were just really fast and heavy bikes, would anything change? I don't think so. It's purely quantitative, rather than qualitative.
It might be that the quantitative differences make generative systems harmful (though I'm not convinced), but that's not substantiated by claims about it not actually being learning.
No, I think it’s neither copying nor learning in the traditional sense, but something fundamentally new, and that’s the crux of the issue.
My argument is that we need to treat this as what it is: a transformative piece of software, but software nonetheless. Not some entity that should be afforded human-like status when examining the ethical questions that naturally arise.
> But speed (and weight) is the key difference. If cars were just really fast and heavy bikes, would anything change? I don't think so. It's purely quantitative, rather than qualitative.
Those are two key differences, but there are a myriad more, each with implications that highlight the difference, e.g. theoretical range, fuel, safety, capacity, etc.
The point there is less about drawing a perfect parallel with AIs and more to point out that the category of automobile introduces a whole slew of considerations that must be accounted for to ensure safe operation alongside bicycles, horse drawn carriages and pedestrians.
As a thought experiment, I think it’s interesting to consider how one might justify photographing or photocopying original works using a similar line of thinking.
“This new scanner/copier sees a piece of artwork and re-draws it on your behalf, and is just a more efficient form of tracing.”
Most people would find this description problematic, but seem to accept a similar form of this when examining AI primarily because the AI software is also able to generate new output.
I’m not claiming that AI is similar to a photo copier, but I am advocating for a more honest conversation about what it is/is not.
I think that looking at past advancements that shattered all established norms at the time is an interesting way to test how one feels about such massive advances. Nuclear energy, firearms, radio, photography, etc. have all had similarly seismic impact.
Especially if the camera or photocopier has a distortion in it. Do Smartphone filters evade copyright? Does printing a digital photograph of a painting using a dot matrix printer evade copyright?
These aren't entirely new questions. https://huckleberryfineart.com/exploring-the-debate-over-and...
Or even moreso, say a one-off piece of copying equipment made by an artist that only ever copies the pages of a novel in a manner that makes it look like a Dali painting, substitutes the letters and number of the novel with emojis (simple substitution), and incorporates all of the pages and the machine itself into a display piece of art.
Where will the line be drawn? Who knows. But as I stated in another post the only thing the AI produces that can be copyrighted is the input of the human, whatever that is. So maybe the AI art isn't copyrightable, maybe the prompts to the AI are all that's copyrightable. Which would make a great deal of sense. But if someone can get the same image by using different prompts, or even by drawing it themselves, then they're free to do so. Or if the prompt is something simple like "Something like Dali, taller than it is wide", maybe none of it is copyrightable.
This entire thing is an interesting philosophical and legal question. And it's interesting to try to see the parameters that the decision will be made on.
Sorry, I just realized I went off on a tangent from the original posted story.
Your use of "just" is an issue, IMO. A significant problem of these systems is that they do copy. They do produce clearly derivative works of the data they're trained on. Copilot has been observed reproducing entire files, verbatim, from its training set. That they can be used to produce novel-seeming works is a legitimately interesting conversation to have, but as long as they're reproducing copyrighted works and clear derivatives thereof, that isn't the conversation the lawyers are interested in having.
And as a result, I feel very different about CoPilot than GPT-3 or especially Stable Diffusion.
Not that I have any idea what this is actually supposed to mean, but no I don't think the model "learned".
While the models themselves don't become inspired, you tell it what to take inspiration from in the same way that you'd decide "I want to make a drawing/photograph like ____ but different by doing ____."
> It's not a technological brain, but IMO it's the closest we've come.
This may be the case, but “the closest we’ve come” seems like a weak argument for anthropomorphizing that progress, especially when you look at how far the gap still is to human-like consciousness.
Hands and feet are probably due to intricacy, which to me seems more like the same issue that humans would have. Again, I'm not arguing that it's intelligent, just that it's...close. I'm not so sure that the gap to human-like consciousness is all that far anymore. If you were to somehow make one of these models more general instead of just text or images and instead "train" it the same way a human child gets trained I'm not entirely sure we'd be able to tell the difference.
Obviously we aren't there yet, but the rapid advances these past few years makes it feel like it might at least happen in my lifetime.
I don't think humans have any difficulties coming with the correct number of hand or digits, drawing them yes but not the number.
Lets say I have an idea for a picture I want. I would like a teddy bear that is in the style of H.R. Giger. I a human write the prompt "style: h.r. giger, a teddy bear" and plug it into Stable Diffusion. I then generate 1000 images, it takes say an hour, and then I pick the one I like.
Now the counter example for our experiment. I have the same idea. I go to my community college and learn to draw (lets say I'm decently talented). I goto a book store, buy a book of H.R. Giger's art. Study that book and try to replicate it, eventually can do a passable copy. I then find a teddy bear, possibly also copyrighted. Then finally draw my teddy bear in the style of H.R. Giger.
Now I dont think most people would say the second example violates any laws and is fair use. You cannot copyright a style (although some lawsuits are moving forward for rap flow) to my knowledge. Yes I would be ridiculed by the art community as derivative. Its unlikely to end up in the Louvre, but could probably be put on a shirt and sold or whatever with minimal legal risk.
So why does the first example feel different then the second? I think it comes back to the process. I would describe them both as the same process but everything after the 'idea' and writing the prompt is just 'steps'. Yes they are steps we attach value to, such as learning to draw, and talent. But ultimately, the real creative thing was the 'idea'.
Another parallel we've experienced has kind of shown us the way. When the first spreadsheet came out it was very much a similar process. However, we don't see it as so as numbers generally aren't seen as creative expression. Never the less there was a desire to calculate someones net profits or whatever. Previously, there was large amounts of accountants, I'm sure with tribal knowledge about how to get this done using a bunch of chalkboards and double checking their work. Most trained on this and knew how to do this fastidiously. Then spreadsheets were invented and it saved them a ton of time. The tribal knowledge was not needed. Even a lay person who had a need to calculate 'net profits' could do that with a fraction of the work prior.
Final thought, if producing derivative works based on previous art becomes trivial. Then wont the artists who can actually create original works be far more valuable, if immediately copied? Perhaps a lot of the concern of this moment is artists seeing how much of what they produce, and actually sells, is really derivative to begin with...
I still don't believe the examples are equivalent, despite some seeming similarities on the surface.
I agree that most people would say the second example is fair use, because the current definition of fair use was established before the existence of generative AI, and with natural limits built in and imposed by human capability.
What is currently deemed acceptable was deemed acceptable because of the friction involved in learning, and when examining what is "fair" about how one uses a piece of art, that evaluation did not at any point include a datacenter-sized AI slurping up every image indiscriminately and then enabling everyone, regardless of education or talent, to generate new works with minimal effort and with almost no time investment required.
If we achieved AGI tomorrow, and the AGI was capable of learning the way humans learn, and drawing inspiration the way humans draw inspiration, but could do so with superhuman efficiency, I think the same quandry exists. It's not about how closely the AI mimics human processes, but the fact that such an AI just fundamentally changes the game. It seems unlikely that the rules of fair use would look the way they do today, or that artists would be willing to allow their work to be ingested by such an AI, if they were given the opportunity to make that choice with a full understanding of the implications.
Agreeing to allow a human to look at my work and allowing a machine intelligence to "look" at it are just fundamentally different things, with drastically different implications. Comparing something to current norms of fair use falls down because the meaning of fair has almost certainly changed.
I struggle with the spreadsheet analogy, because this falls into a category of automating error-prone (when performed by humans) tasks guided by objective rules that look something like the laws of nature.
Art is inherently subjective, and with without human inspiration, the AI would not exist. Or at best, it would only be capable of producing outputs as mediocre as its inputs.
Unfortunately, in the art world there have been plenty of well known artists (even some masters) who made extensive use of assistants and in some cases never even touched a canvas. Or more to the point were incapable of actually producing the final product. However, they are still the artist of note for the piece as it was their 'idea' or they created the style, etc. So clearly its not as simple as 'it should be hard' or 'work' that makes someone an artist and their work of worth. See Warhol and many pop artists for pieces of extreme value, but of relative ease to produce. Again I fall back to the prompt being the 'idea' and the truly artistic thing, even if stable diffusion or whatever is doing the work. (https://www.worldhistory.org/article/1611/life-in-a-renaissa...)
It's not about "it should be hard", but rather "artists established norms about fair used based on the fact that it was hard".
Now that this technology exists, those assumptions no longer hold up.
Not in the literal sense, but in this scenario it would be like taking indistinguishable copies all of H.R. Geiger's art, painted from the same paint that H.R. Geiger used, washing the canvases clean, and then using those same paints to paint something new in H.R. Geiger's style from memory.
There is something indefinable that has been transgressed. The style, the ownership of the paints themselves are undeniably H.R. Geiger's.
Here's the receipt.
But, what was painted with them was not actually H.R. Geiger's. The originals were not destroyed or damaged in any way. H.R. Geiger lost nothing and if enough people saw the new art it might even profit the estate of the originator by increased attention and sales, heck, even the paint was free.
This result would have been impossible without H.R. Geiger's paints and how he used them. The question is, was this theft? Did this take something from someone? Are they damaged?
I think the answer is yes it is theft, but that nothing was taken and there is no damage, but that's just me.
In my opinion, under this rubric, the extent a computer-generated work can be copyrighted (or be seen as transformational enough to be fair use) would depend on the amount of effort a human put into iterating that specific piece of work.
And even then whether or not Andy Warhol's paintings are copyrightable, or are unauthorized copies of copyrighted work, is still going through the courts.
> When the first spreadsheet came out it was very much a similar process.
Processes have never been copyrightable. A particular implementation can be patented, but copyright is an entirely different form of intellectual property.
Sadly, I think artists have a better chance trying to limit their works inclusion in the datasets then making any claim as far as how easy it is to use AI. It likely, will fall along web scraping cases, which I'm not knowledgeable enough to say one way or another. However, it doesn't bode well that I can put an artists name into google image and see several examples of their work, this feels like public access and to my laymans mind would fall under fair use.
ref: https://www.thelegalartist.com/blog/you-cant-copyright-style
The problem here is... copying! AI models aren't just taking the "style" from works, they are directly copying the works because they do not have an understanding of what style is because they are computer algorithms that are designed to generate pixel data based upon other inputs.
AI is going to be a major competitive edge for corporations, and lawfare between Winners and Losers will be intense. I say corporations because I don't think normal people will have the capital to train AI's or even run them.
The difference between midjourney and stable diffusion is great example to look at these days. Midjourney is making waves things like 80's movie shrek and skyrim.
Curious to see where this goes; I feel like the UK venue is going to be relatively more friendly to Getty (in the US, I’d be comfortable with a transformative fair-use argument, but I don’t know what the analogues are in UK law).
"The data we licensed from Shutterstock was critical to the training of DALL-E,” said Sam Altman, OpenAI’s CEO."
- Stability AI pays Getty a fee, which they really should, to some degree/percentage, if they're making money off this.
- But the amount should be for a proper use license, rather than focus on being punitive. Perhaps they work out another kind of deal. This is what the execs figure out together.
- Then they re-train on the entire set, sans watermark, now with the full blessing of Getty, and the resulting Stable Diffusion v3.0 release ends up being mind blowing!
It is a company which has even been known to sue the original creators of images for images they never licensed to Getty.
They have been trawling the web for images and making spurious claims of rights over them and questionable legal threats.
The business model only makes sense in an environment where corporations lobby for laws enabling them to seek rents over resources which have been freely available and whose copyrightability is even questionable.
Frankly I put them up there with patent trolls and those evil corporations which claim patents over seeds which farmers round the world have developed and improved over generations.
But in a legal and commercial environment where rounded corners can be patented why should we be surprised?
Is this supposed to be a factual statement? Getty owns the rights to its photo library. How is that even in question?
It's important to note, this is not a faithful representation once you look at it closely. The text is always mangled into nonsense and things like brands or logos are never close enough to be even a passing likeness.
This is no different from a painter using a photograph of a subject to create a portrait, which is subsequently sold.
Or recording a song on the radio to learn to play the guitar, then using that song as inspiration to write a new song.
That being said, I suspect they are removing all watermarked images from training databases right now.
> Or recording a song on the radio to learn to play the guitar, then using that song as inspiration to write a new song.
I’d argue that it is very different from those things.
The AI software is not human, and the precedent set for painters and musicians is based on a foundation of human capability.
It’s no different in the same way a horse drawn carriage is no different from an F1 race car.
In all seriousness, I'm not surprised by this as I keep saying, this is quite obviously not fair use as Stability AI has a for profit service (Dreamstudio) and Stability AI didn't even seek a license from Getty to use their images in their training data.
All you have to do is this:
1. Just ask for permission.
2. Pay for a license from Getty Images?
That is all, it's really that simple.
Getty is not even against AI Art tools / generation:
> When asked what remedies against Getty Images would be seeking from Stability AI, Peters said the company was not interested in financial damages or stopping the development of AI art tools, but in creating a new legal status quo (presumably one with favorable licensing terms for Getty Images).
We will see more of this with other companies (perhaps bigger ones) seeking more lawsuits from Stability AI as it's evident to see the infringing images in the training data.
Should an artist pay Getty for leafing through their images or visiting a museum and closely studying?
Stability isn’t deriving images from Getty, they are getting “inspired.”
I fear that if this lawsuit succeeds that eventually just by existing, I’ll be forced to license the world I see in order to perceive it and incorporate it into my consciousness.
How will digital consciousnesses work with licensing? Will an AI be forced to pay a license fee because it uses copyrighted images as part of its consciousness? What happens if it can’t pay? Is it shut off? Is it forced into slavery?
Maybe the upside is that copyright can be revisited with the maximalists not in power. I think Getty benefits from unjust copyrights that last decades.
[There is a separate point around the value of the Getty database. Routinely, companies are sold with their customer list / database as a main selling point. Getty has undoubtable value through the sheer size of collected work. Nobody could expect that we just assign 'no' value to it, even if it can be accessed (with a watermark) for free by most.]
Getty has chosen the forum (UK courts) quite carefully. The precedents favor them substantially. In 2012 there was a case where a photo was found violating copyrights because the photographer independently took pictures of Westminster bridge and reproduced a similar photo.
Basically the case says you're violating copyrights if you're "inspired" by a photo and make a similar one intentionally. See eg. https://www.lexology.com/library/detail.aspx?g=57171429-4dea... for details.
Pretty sure that case would have been laughed out of court in the US, but historically the UK courts have been pretty strict in interpreting the scope of copyright protection. So even if you're successful in arguing for "the AI has a digital consciousness and is basically a human" analogy, case law in the UK still doesn't allow you to do that.
I personally think this is ridiculous though, but unfortunately that's the status quo (in the UK where SD is being sued).
An AI model isn't an artist leafing through their images or visiting a museum and closely studying, so what exactly is your point?
This is what most HNers miss when they invoke the 'copy art from the museum / gallery' argument.
Just because it's a scrambled watermark doesn't mean it's transformative.
Also, these images are copyrighted and it being generated by an AI (supposedly) turns it into a public domain image, which is incompatible with fair use.
And where do you draw the line: https://www.lexisnexis.com/community/casebrief/p/casebrief-a....
You're saying it can't be a 1:1 Getty Images watermark but it also can't be generated in a vacuum, so how did that 'recognisable watermark that apparently for some reason resembles a Getty Images watermark but it isn't' get in there?
Either way, under the law, derivative works are still under the copyright holder.
> Copyright protection for the owner of the original copyright extends to derivative works. This means that the copyright owner of the original work also owns the rights to derivative works. Therefore, the owner of the copyright to the original work may bring a copyright infringement lawsuit against someone who creates a derivative work without permission.
https://www.legalzoom.com/articles/what-are-derivative-works...
But here is the kicker, it all comes down to, permission.
https://www.gettyimages.co.uk/eula
> No Machine Learning, AI, or Biometric Technology Use. Unless explicitly authorised in a Getty Images invoice, sales order confirmation or licence agreement, you may not use content (including any caption information, keywords or other metadata associated with content) for any machine learning and/or artificial intelligence purposes, or for any technologies designed or intended for the identification of natural persons. Additionally, Getty Images does not represent or warrant that consent has been obtained for such uses with respect to model-released content.
https://petapixel.com/2016/11/22/1-billion-getty-images-laws...
I'd bet that Getty's is-this-our-image? technology and where-did-we-get-this-one? database can very easily figure out whether or not that is true. And already has.
In any case, adding the other guy's copyright notice to work that you claim is your original stuff smells to me like telling the judge to stick it where the sun don't shine.
If they had clearly copied Getty Images watermarks, sure... The problem is, they don't.
Still a strange question for me but will be very interesting to see the results of this. It was always going to happen eventually.
Stable Diffusion does not have a mental process.
See this for a clearer explanation - https://i.redd.it/2f00l6vsso6a1.jpg
But where exactly is the line in the sand? Reshape, resize, +1 on color map? When have I infringed on a copyright of a picture?
I would argue as soon as the diffusion model has comercial use, royalties are due for the original creators. In any case the data set has to be cited properly.
Otherwise it is piracy, which is very very cool, if you don't make money of it.
Just like OpenAI did for Shutterstock before they released DALL-E.
https://www.shutterstock.com/press/20435
As for individuals, once the licence fee is paid I expect Stability AI to remove infringing training material (with artists requesting to opt out), update their model significantly and publish a new an improved model that (perhaps not completely guaranteed) convinces individuals to switch.
This may be in the form of an updated hosted model on Dreamstudio. Although for the open source model, well this is in muddy waters.
Maybe this isn't going to happen with all individuals but with the mess that Stability AI created for themselves that is the consequence of prematurely releasing the model trained on copyrighted material and getting caught for it.
https://techcrunch.com/2022/08/12/a-startup-wants-to-democra...
It seems like it would be really challenging to improve the model for users while removing a lot of the training material. I'll be interested to see if and how it happens!
Opt out isn't how it works sir.
If we achieved quantum computing tomorrow, all of the norms around what is considered safe would change overnight. We would not retroactively tell people they had it coming if suddenly people were getting breached left and right because the assumptions they had relied on are no longer relevant.
TL;DC once you take apart the thing and put the lego pieces back in the box, there's no more thing, only pieces
I merely disassembled your car, at which point it ceased to exist and was no longer your car. And then I took the pieces of not-your-car for myself.
If this sounds silly to you, it should.
Clearly the issue is that these ML models are being trained on stuff that isn't theirs. You can argue that they never "stole" the images because it's computer data. So now software piracy is fine right?
All content on the internet, even uploaded to freely viewable image libraries, has some manner of license. I could just go on DeviantArt and grab some person's work and start making T-Shirts for sale, but that wouldn't be legal would it?
So why should it be just assumed that the license allows training on AI models? This needs to be tested in court.
Comparing theft to "intellectual" property that can be trivially copied often doesn't make sense and can lead one to wrong conclusions.
> So now software piracy is fine right?
Finally something almost on topic. But that's just begging the question. And it's not software we're talking about, but images.
Training neural networks on massive datasets is a pretty new phenomenon, I don't think anyone can say definitely whether it's copyright infringement or not, because...
> This needs to be tested in court.
Exactly. I'm glad we agree on at least something.
Historically the UK courts have taken a (overly IMHO) strict attitude (eg. https://www.lexology.com/library/detail.aspx?g=57171429-4dea... ) to copyrights so I'd guess the outcome won't be too favorable to Stable Diffusion, but it wouldn't it be funny to read your assertions in case they win.
As some background, Getty Images have partnered with BRIA, a similar company (https://www.theregister.com/2022/10/26/shutterstock_getty_op...)