Image generation ethics: Will you be an AI vegan?
simonwillison.net
simonwillison.net
I've thought about this comparison to human artists taking inspiration from each other a bit. The problem becomes the scale. A human artist can't look at 5bn images and use them to be able to instantly mimic the style of another artist, hundreds of times an hour.
Consider facial recognition technology. At an individual level, empowering people to quickly see all of the photos they have taken of their spouse through running facial recognition against their photos is useful and harmless. Allowing governments to run the exact same technology against every photo uploaded to Facebook is a massively harmful expansion of the surveillance state.
I don't think training models against 5bn unlicensed images is in the same scale of harm as running facial recognition against an entire country's worth of people. But this example does show that things that are fine on a small scale can be harmful at a big scale.
I feel it is insulting to veganism to try and coin this term. The moral arguments are not similar, and this one is frankly much weaker.
> I've thought about this comparison to human artists taking inspiration from each other a bit. The problem becomes the scale. A human artist can't look at 5bn images and use them to be able to instantly mimic the style of another artist, hundreds of times an hour.
A human can mimic something with a single reference. It might not be good, but that can hardly matter for a discussion of ethics. The morality cannot depend on whether or not humans can do something poorly or not.
> Consider facial recognition technology. At an individual level, empowering people to quickly see all of the photos they have taken of their spouse through running facial recognition against their photos is useful and harmless. Allowing governments to run the exact same technology against every photo uploaded to Facebook is a massively harmful expansion of the surveillance state.
Entirely different issue that happens to utilize a similar tech.
> But this example does show that things that are fine on a small scale can be harmful at a big scale.
It does not.
This is another example of an ethical choice that I made!
Is the harm here that is caused by being offensive to veganism outweighed by the value of sparking useful conversations about personal approaches to AI ethics?
I think that it is.
It's similar to the ethics of attaching a click-bait headline to a valuable piece of journalism. Is it OK to do that? That depends, both on the specific case and on the opinions of individuals who observe that decision.
And to be honest, if it hadn't worked I would have thought "well, it was worth a shot". The analogy genuinely works well for me.
I agree. This would more accurately be called "AI Luddism", opposing a new technology because of fears related to its immense productivity over humans. It has nothing to do with veganism
EDIT: that said, vegan is used in the sense of the article as "a really obvious and strong stance that people do not follow even though all the information about how it is bad is readily available"
https://theguardian.com/commentisfree/2015/jun/17/napoleon-d...
> In one conversation on St Helena, Napoleon was asked by his doctor what he would have done if he had managed to invade southern England in 1805. Napoleon’s reply was an absolute cracker: “I would have hastened over my flotilla with two hundred thousand men, landed as near Chatham as possible and proceeded direct to London, where I calculated to arrive in four days from the time of my landing. I would have proclaimed a republic and the abolition of the nobility and the House of Peers, the distribution of the property of such of the latter as opposed me amongst my partisans, liberty, equality and the sovereignty of the people.”
They smashed machines because that hurt the people oppressing them, not because they hated the machines.
Admittedly, the UK was warring againt the French Monach as well before, so it's not a clear cut thing, but things like Luddites and the 'Peterloo' show that Britain was not speeding towards greater democracy. They very slowly gave ground over centuries of struggle.
> The Tory Prime Minister in 1830, Arthur Wellesley, First Duke of Wellington, was resolutely opposed to parliamentary reform. However, there was growing support for limited change within his party, primarily because partially extending the franchise would allow the wealth and influence of Britain's growing middle class to be exploited.
That's 20 years later, a small extra step of democracy.
I was just trying to point out that smashing some stuff was a fairly reasonable thing to do when you don't have a vote and you're hungry, by providing some political context on the era.
At times this whole thing feels hypocritical. Generations of laborers grew up with the reality that their work was essentially a commodity. Now that it's our turn, we raise the age old arguments again.
Automation of labor has been an ongoing force for decades and a reality laborers have had to live with for years.
There are many, many, many people out there who believe it's fine to eat meat but not fine to e.g. feign attribution on intellectual work. Asserting that objections to this kind of tech are much weaker than various vegan arguments without explanation is odd.
> A human can mimic something with a single reference. It might not be good, but that can hardly matter for a discussion of ethics. The morality cannot depend on whether or not humans can do something poorly or not.
Asserting this without argument suggests you have a particular unstated idea of how ethical determinations are supposed to be made that you are projecting out onto the discussion. This would be important to lay out.
It's hard to imagine the mental gymnastics required to think the two are in the same league.
It's really not hard. You just have to care more for the harm to the well-being of human artists than for the harm for the well-being of farmed animals. It may surprise some people but speciesism is the default position for the vast majority of the population.
What you're saying is that you think intellectual property is more valuable, to you, than the life of other conscious non human beings.
I think for most people, this is a stretch.
The mental gymnastics are required to rationalize the act of eating meat and murder being the same thing.
But I don’t find either one hard to imagine.
The vegan argument is that factory farms are literal lifetime torture pins. This isn't much up to debate. You can shrug and choose to eat meat anyway, but co-opting their name for your argument on ip seems incredibly shallow. Even if you believe that ip violations are worse than animal cruelty. I'm not a vegetarian at all fwiw. Not too dissimilar from trying to frame this as the holocaust of artists or any other strongly held event / movement by many others.
> Asserting this without argument suggests you have a particular unstated idea of how ethical determinations are supposed to be made that you are projecting out onto the discussion. This would be important to lay out.
If your ethical framework suggests ip violations are immoral, unless you think the execution was bad that's just silly.
Can you say more about the scale argument? It reads as if you're saying that an entity that has an advantage shouldn't use that advantage.
If you reduce the comparison to human scale, do you still have a problem with the competition? I.e., if one artist has a working memory of 5 inspirational artworks (unlicensed) and another artist has a working memory of 3, is that wrong if they both compete to produce new artwork based on what they know of existing art?
If the answer to that question is no, then I don't see how the scale question is applicable.
One of the computer's "super powers" is its ability to run at scale.
Very few people are happy with that.
Therefore, if there was some way to eliminate this sort of advantage, especially for very large organisations like the government, that would be excellent.
To me this seems to similar to old fashioned luddites. The machines back then that improved the productivity of labourers mostly displaced the societal role of skilled craftsmen. Of course using the machine is different to doing it by hand, but we did have some benefits from improving our productivity using machines, as well as some downsides. The interesting thing is that in this case we are moving from using machines to improve productivity of material objects, into sociocultural objects (art), and it's not clear what that will "do" exactly.
> But this example does show that things that are fine on a small scale can be harmful at a big scale.
Right but of course, the words "can be" are there. I don't think anyone is claiming that scaling a given thing up is always harmless, but generally with these kinds of technological developments you really need a super strong, clear, and concrete (for this specific case) argument to stop them. I haven't seen any such arguments, except the traditional luddite one I posted above.
This idea of CAPTCHAs and “human only, no bots” networks is only for a decade or two, until bots just become more desirable members of every community, to humans and other bots.
What people don’t get is that their individual decisions in their own self interest doom the group. Dad works 10 hours a day to provide for family. Then mom does too. Now they both neglected their kids and each other.
Similarly, you choose a bot for its wittier comebacks and impeccably romantic style of talking. Your husband chooses a bot for its better sexual techniques and never complaining. Ultimately, neither of you needs the other and you’re one step away drom NO ONE IN SOCIETY needing you.
You would become the Borg. Except why have the drone bodies at all?
It's just another step on the path towards AI being better than humans at certain activities. We've already gone through this with things like math, chess, imitating voices, translating languages, and lately visual art seems to be coming close to a point where the AI will have an edge over humans. We will have to learn to cope, one way or another, because eventually, AI will be better at everything, as we are limited by our biological bodies.
Humanity will create its own synthetic god, and the consequences will be interesting.
For example, it's easy to construct hypothetical scenarios where a difference in scale seem to clearly constitute a difference in kind and where suppression of a new technology seems clearly reasonable. An easy hypothetical (but not implausible) example is some new kind of extremely deadly weapon that would be much easier to build than existing weapons if the "recipe" was made public.
But it's also easy to make the same argument for things like, say, the printing press. And indeed there has always been intense opposition from many to the printing press and subsequent technologies that make distributing written language much easier, with opponents often presenting arguments in this same form. Yet it feels wrong to me to apply the argument this way.
Is that just because I have existing opinions that generally oppose restrictions in the ability to distribute ideas, but I don't generally oppose restrictions on very deadly weapons? Or is there some other way to distinguish between valid and invalid instances of this form of argument, perhaps by attempting to estimate the overall risk versus reward?
"meat-based" vs "plant-based" is a spectrum we are all on; no reason to make it all-or-nothing
To the last question: No, but I wouldn't pursue art either way.
We should allow _everybody_ to run the exact same technology. Governments are not special. Power to the people.
Humans also are good at taking inspiration from ideas, where today’s AI is borrowing pixels. The AI is copying in a way that humans don’t, it’s not mimicking and taking inspiration - that is anthropomorphising the software that is trained and programmed to make automated mashups of images.
So in part it depends on what we do with AI images. There may be nothing wrong with training on copyrighted material if the resulting inferences are never distributed publicly nor used for commercial purposes. Of course that seems extremely unlikely, which is why it needs to be discussed carefully and debated in good faith, right?
Maybe in part it also depends on whether the AI software is guaranteed to produce something very different from any individual training image. If the outputs are guaranteed to be always a mashup, and never indistinguishable from any single input, that seems like it would be an easier pill to swallow. (There appears to be legal precedent along these lines for music sampling.)
Direct copying yes, but stealing ideas? No. Actually it is a well-known practice in the art industry to use several existing works as references, and creating new artwork that is bit of mixture of all of them.
That is exactly what DALL-E/StableDiffusion are trying to do.
AI isn't borrowing pixels, either. If you read the idea of diffusion models, quite the opposite, what it directly learns is how to destroy an image from what it is into gaussian noise. The trick here is we can reverse this process 'creating' arts from noises.
There is no copying pixels, not even in the most simplistic version of how this model works.
All this goes out the window with AI, we have to judge the entire model as a whole, and even though we see DALL-E etc spit out watermarks we can't call it a liar or a thief. "This part overfits but the whole model is transformative, we think" say the authors.
I disagree with this, we don't just hear about ideas we actually see the art. Both are shown pictures.
2) We still don't know exactly how much the models memorise. I'm sure you'll agree that something like Google image search retrieval doesn't qualify as original art, and copyright is still an issue. If you photoshop two images together, you probably also still have to give credit to the original images. We have to draw the line somewhere on the scale from "100% derivative to 100% original". It's not yet clear where AI image generation falls on this.
Ok so let's just feed the ai some dashcam footage and this argument is null.
> I'm sure you'll agree that something like Google image search retrieval doesn't qualify as original art,
Yes
> and copyright is still an issue
No
But for now when the training set is 100% others work and you can see the AI literally trying to inject a jumbled artist name/trademark into "original" art I can't see it as anything more than extreme photobashing
> Yes
>> and copyright is still an issue
> No
You don't think I'll get trouble with copyright infringement if I start using images I find on Google commercially? "But officer, I came up with the keyword for this image all by myself!!" Or do you just mean you don't think copyright should exist in the first place?
I think you’re wrong on this. Humans learn (anything) from teachers. Yes, you can learn much on your own, but the idea of a solo artist learning from books is exceedingly rare.
So in this case, you learn art by doing, critiquing your work and the works of others. You put forth effort.
While I am amazed at the generated images and think they truly are amazing, I can’t help but think they all feel a little cheap. Like someone took a shortcut that was never meant to be found.
I do think there are real ethical issues behind the training data for both image and code generation. Nothing that can’t be solved, but random images scraped from the web are not meant for training. First - it’s not necessarily a fair-use issue. And second, garbage in, garbage out. I don’t want my auto generated images to come with a Getty watermark.
Where I do think there is hope is for the use of these as tools for artists. Where there can still be a human behind the choices and curation, but using the algorithms as a means rather than an end.
If you take an art class where a teacher instructs you and tells you to replicate 12 paintings by Van Gogh, I would argue you've learned more from Van Gogh than the teacher, even if he's showing you some physical techniques.
The student/teacher interaction is where you learn. Even if you want to only paint in Van Gogh’s style, the back and forth with the teacher is where you actually learn how to apply that style to new works. To explore the meaning behind the stars, etc.
(You could get there solo, but you’d still need to go through the criticism/questioning phase somehow, which is probably best done with an IRL teacher).
And only some can boldly go, to were no person has gone before.
Which makes this the ultimate training goal for AI. Not AGI, but a synthesis AI, capable to produce "breakthrough" candidates for the field it is trained upon, by allowing noise and filtering for the criteria of great break throughs- explanation power, beauty, higher consistency, that puzzle piece fitting all gaps moment. If there is ever a creature out there, doing that, silicon or otherwise, humanity will own its continued existence to its existence.
We're still not sure whether humans are deterministic or not. So you can't really equate the human process of art creation to a computers. Humans may still be pulling from some external inspiration that will never be within the reach of computers (I like to think that's the case).
Or maybe you want to still argue that that's ultimately deterministic.
Okay, well, just have a person show up to the art computer, look deep within their soul, then type in a completely random, soulful, non-deterministic seed value into the art computer. Bam! Now the computer can create real art.
Thinking that this system, with such a small set or data and no natural language processing, is going to replace artists anytime soon, is, I think, incredibly eager to the point of foolishness.
My fear is copilot will be allowed to copy code, but I won't be.
This is not a reasonable model.
Art is not about form. It's about content. Meaning that can be generated and understand only by humans.
Would you say that artist did not produce something that could be called art, just because they used an AI model as part of their process?
Art is more deep. Has meaning. It doesn't matter how is it done. I mean... You wouldn't find something like that in art gallery on wall.
What if the image that person had drawn really impressed you on its artistic merits before you learned how they had made it?
These models take 5bn+ images and use them to influence the weights on a giant array of floating point numbers. Each input image has a truly tiny effect on the finished product.
The final model is on the order of 4.5GB - the compression ratio is unreal. Nothing of the original images remains.
So "stencil copy" doesn't work for me.
i suppose we'd all feel uneasy when AI eventually becomes better than humans at creative tasks which pay our bills or differentiate us within a profession.
> I myself remain a meat-eater.
It strikes me as off that one would consider themselves informed enough to have decided not to be a vegan but consider image generation AI unethical. As a former carnivore (literally sometimes going an entire day eating mostly or entirely meat), access to the information about just how horrific factory farming is and the willingness to open my eyes to it was the only thing that stopped me from being persuaded.
I only eat meat once or twice a week, and I try to consider the sources, but despite understanding the moral implications of doing so I have not gone vegan or vegetarian.
To my mind, this is similar to a situation in which I determine that using AI trained on unlicensed images is unethical but continue to chose to use those AIs.
Everyone cutting meat by 50% is more impactful than 5% of people cutting meat to 0%
"Aspiring vegan", to my ear, kinda sounds like one thinks about it, but performs zero action towards it, which is a bit of a disservice to those who do indeed consciously do not eat meat for the majority of their meals.
Furthermore, I think a lot of art generated by actual intelligence is made by those consuming tons of copyrighted material and putting a twist on things. Is it unethical to listen to The Monkees, since, to put it in the terms of the article, they were so clearly trained on the Beatles with a few tweaks here or there?
In music if you use a snipet of someone else's recording you have to pay them. It wasn't always clear that would be the case, but thats where they landed. (You end up with Led Zeppelin and Beatles Samples in some early rap).
Even "Sound alike " singers get litigated: https://en.wikipedia.org/wiki/Midler_v._Ford_Motor_Co.
But in visual art its a little different. Borrowing is more common and remixing is kinda allowed. When does it become "transformative?" (I always think of the "Hope" poster lawsuit, where the borrower sued the photographer as a strange one. ) https://www.law.columbia.edu/news/archive/obama-hope-poster-...
But what is the AI doing? I don't think its taking the input given and being inspired... Its kinda just sampling, in a way that makes it seem like its being original. Or is the unique training set/ annotations that are is giving the AI its unique output the art, in which case its more original.
At some point some AI is going to spit something out too close to something else and the courts will probably have to decide.
People have been spitting out stuff too close to decide and letting courts decide for well over a hundred years. This is nothing new.
If people make art that is in the style of someone else, by hand or by AI, it's generally fine to do so. If it gets too close, then courts can step in as always.
How is it different from how human artists train on copyrighted images?
We have no trouble to award them copyright on art which consists of elements, or is heavily inspired by elements of copyrighted works they've seen during their education?
Human imagination can't create anything really novel. Everything you create is just cutting, stitching and deforming what you already seen in semi-random ways until you get something interesting to somebody.
Try imagining how an alien might look.
That's an awfully low opinion of art and humanity.
The part where the company training the models makes $$$ on copyrighted work can definitely be a debate point.
Also the comparison to veganism and animal suffering is off putting.
Technology has continuously brought us easier and more immediate ways to create art, inspiring new generations of artists who hone their skills with the new tech. Meanwhile, older forms of art continue to be valued along side the new stuff.
I'm also having a hard time seeing the ethical crisis with these AIs being trained on copyrighted material. Styles are not (or at least should not be) copyrightable. An artist can be inspired by the works of another and go on to create something new. Many forms of derivative work are even specifically granted safety under existing copyright law.
Besides, it would be practically impossible to prove that a model was trained on copyrighted works. Even if we decided it was unethical, any law against it would be theoretical and practically unenforceable. Either way, artists will have to adapt.
I don't think the situation is near as dire as so many seem to believe. An AI can only reproduce a style that has been thoroughly explored by the content which it is trained on. New styles will continue to be rewarded. Digital artists will be encouraged to push boundaries. And for the time being, the AIs still have some pretty severe limitations so artists will be able to capitalize on those.
And one more thing I never see brought up when talking about these AI image generators: there's already precedent for how this will play out. AI music composers have been around for many years now, but Dua Lipa and The Weeknd appear to be doing just fine. Even the more classical composers and orchestras seem to be going just as strong as ever. If AI artists show no sign of toppling the music industry, why should we expect the fate of digital images to be so different?
All of those are tools which did not lead to potentially the same final product produced for a miniature fraction of the labor costs.
AI composition is generally pretty shit. If (more like when) it becomes better, we will be having the exact same argument regarding composers.
In what practical sense are paintings distinct from photography or digital art?
It's easy to see how they're technically different, but in terms of purpose paintings should have been supplanted by photography and digital art long ago. After all, the latter two can be reproduced "for a miniature fraction of the labor costs", provide the same utility, and can be reproduced at a much larger scale. Yet painting and similar physical art forms seem to be as economically viable as ever - possibly more than ever, but I'm having trouble finding hard numbers to confidently back that up.
“The fact that it can compress such an enormous quantity of visual information into such a small space is itself a fascinating detail.”
This is not a detail: it’s the principle mechanism. The ability to compress something is conferred by the identification and exploitation of structure, conversely the scarcity or absence of structure inhibits or prohibits compression. You can eyeball check an RNG with compression techniques.
This has counter-intuitive consequences that you can test on your laptop! Even using off-the-shelf codecs it only takes a modest corpus to see that pop music compresses better than eclectic jazz, which compresses better than white noise. The same thing holds for headshots of people: a big pile of headshots drawn from a reasonably broad corpus of humans will enjoy a noticeably lower compression ratio than a subset selected by any plausible “conventional attractiveness” filter. “Conventional attractiveness” (defined any common-sense way) correlates sharply with bilateral symmetry, with obvious implications for storage space.
Information theory is the thread that ties together all this AI craze stuff!
I am not saying that attribution doesn’t have an ethical dimension. Only that attribution is morally/ethically independent of the legal construct of copyright.
Not that I think it matters much because I don’t expect AI art will have much impact as art in the long run…which is not the same as saying it won’t have much impact in the realm of disposable images.
The reason I think AI art won’t be terribly important is that there do not appear to be very many intellectually interesting things to say about it as art beyond is-it-art-because-I-hung-it-on-the-wall?
And none of this is to say that artists won’t make interesting AI art. Because artists and their processes are intellectually interesting in ways that training sets are not. I mean David Hockney’s iPad art is interesting because of David Hockney not because of iPads.
YMMV.
I have been playing with Stable Diffusion for a week. I think it is going to have a significant impact. The so-called disposable art probably gives lots of artists their first paid gigs that gives them paid time to practice and build a portfolio, while learning to work with customers. I think a lot of that work will shift to prolific prompt-engineers who know their way around the software.
I'm very pro-automaton (I'm a robotics engineer) but I don't like the idea of taking work from a bunch of artists without their consent to build a model that depresses their wages. I would rather these companies worked on sample-efficiency (an important area of research) and stuck to the millions of openly licensed images available on the web, then worked with twitter and instagram to add image license options to the photo widgets so that we can all build up a consent-based model. And artists that want to keep their work for themselves have that choice.
I don't think it will happen, but I do wish we actually honored the consent of the artists. I think it's a shame we're not.
I similarly understand and sympathise with the apprehension of generating images from this massive harvesting of data with little regard of what should and should not be in the dataset.
I think by using these models as pre-training weights and fine-tuning on data which one believes they have the right to use or (even better) has created themselves, you can (IMHO) greatly minimise the harm of your model's output.
I also like this from a conceptual stand-point. We have the right to learn and be inspired by others, but when it comes to putting the paint brush to the canvas or the ink to the page, it should be our own experiences that we draw primarily from.
I'd love to be able to run GitHub Copilot weighted heavily towards code that I've written myself, as an additional training layer.
If the outputs were sufficiently different from the most similar training examples would it be exempt from copyright liabilities? On the other extreme, a single copyrighted image in the training set could spoil the whole model, it would be impossible to make sure the training set is 100% clean.
This is a great kind of question to ask, and can point us to the right areas of non-deep-learning copyright to think about – but not in a way that lends itself to a quick settled answer in a comment section, I suspect. https://law.marquette.edu/facultyblog/tag/fairey-v-ap/
> On the other extreme, a single copyrighted image in the training set could spoil the whole model, it would be impossible to make sure the training set is 100% clean.
Where possessing stolen goods is illegal, it takes quite a lot of effort to determine that an e.g. antiquities dealer has met the reasonable standard of inquiry for 100% of their inventory – but that doesn't mean that isn't the legal requirement. That's an area where we know the legal standard, and this is much newer ground, so it very well may be that tech that depends on datasets too big to take responsibility for may not end up being a great foundation for anything with legal liability attached.
Digital technology enables such massive economies of scale, doing away with almost any kind of friction you'd encounter when doing something in the real world, that they are qualitatively different from their analog, offline counterparts. Quantity has a quality of its own, as the saying goes.
I'm undecided on the morals of this, so don't understand it as an argument against doing it per se, but "I'm doing it perhaps a couple of hundred times daily if I really try hard, so it's clearly okay to do it billions of times a second" isn't a good argument for it.
If you believe in copyright then you don't believe in ethics^1 ;).
[1] Assuming you've actually thought about copyright from first principles—I think many people are too busy to have given it enough thought.
Art is typically quite style-promiscuous, and mostly protected by copyright. It’s also quite analog/continuous. Code is more discrete; you can diff it and easily see if an AI copied a particular block that doesn’t appear in other projects.
Personally I don’t think there is a strong argument that training an AI on art is problematic, as long as the AI doesn’t then commit copyright infringement. I think this process is equivalent to art students consuming art and synthesizing their own styles.
I think it’s more problematic if AI starts regurgitating discrete chunks of FOSS code, since that’s probably a license violation.
For digital artists, lots of skills are going to become irrelevant; the raw skill of pixel pushing and line work may well become obsolete. But the hardest part was always having good taste, composition, ideation. These are not going away any time soon.
And if you enjoy pushing pixels around, nothing to stop you doing that; we still play chess for fun even though computers dominate humans at the task. Indeed, computers make very scalable and accessible coaches; one could argue that chess is more fun, and easier to learn, now that computers have superseded humans. Perhaps the same will be true about art, and later, coding.
Why would you treat the ML model (I very much hesitate to call it AI) differently in these two cases?
To be clear I don’t think learning from (but not cloning) a GPL project is problematic. That is just what normal students do.
What if the code is 99% FOSS code, but with variable names changed say? I'm curious because for humans I think that would still count as copyright infringement, but Github seems to only filter out 100% code matches.
I suppose copyright law is not precise enough that you could algorithmically decide of two pieces of code are too close? The only way to be sure seems to be "clean room design", where the AI hasn't seen the FOSS code in the first place?
That clause would make their software not Free and Open Source by definition.
> The license must allow modifications and derived works, and must allow them to be distributed under the same terms as the license of the original software.
So AI training would be allowed, but all code subsequently generated by the AI would have to be under a compatible license.
It will take very very large ethical issues, direct death or suffering, for me to reconsider. It's just such an amazing opportunity that I am not willing to just give it away for nothing.
Similar to how I still eat meat.
Artists have been trained on hundreds of copyrighted images in magazines, museums and on the web.
I feel both that the time has come for fundamental cultural change within the tech/hacker community, and that expressions like "AI Veganism" are indeed appropriate and useful on many levels.
They enable the discussion of lifestyle choice, rights, health, ecology and many other salient factors in relation to digital technology.
At this point, those still clinging to a rejection of modern tech-critique as in any sense anti-progressive or "Luddite" are in fact the ones woefully out of date and out of step with how the world is turning.
FWIW, so do I.
But much like we say of capitalism and democracy, it's the worst apart from all the others.
I have yet to find a better one that captures the attendant issues.
Might be a textbook example of technological progress: Lots of people have their jobs eliminated (which suuucks), but many more benefit from having access to professional-level custom artwork that they didn't have before.
AI/ML is coming for everything. High-paying jobs like doctors, lawyers, and developers are definitely not immune.
These aren't constraints on AI developed, controlled, and accessible only to enormous corporate entities; even if they deign to provide limited access to the public at times.
If you’re a professional artist today, you need to lean into this and adapt somehow. Current AI is pretty limited, but that’s not going to last.
I wish I had an answer for the “how to adapt” question, but I don’t. All I know is that this isn’t the first time people have tried to futilely resist a technological revolution.
I do think though that copyright law needs to be updated to deal with this AI stuff. Copyright was created to protect creative works, since that is a very important thing to protect both for the economy and culture. I just hope that the people and companies creating these AI are on the right side of that fight.
I doubt it though.
technological progress*, it's very debatable if that can be called "progress" in the general sense of the term
Several periods of known human history, starting from the late bronze age, disagree.
Now, could someone choose to use it in a completely derivative way to essentially rip-off a well known artist? Sure. That would be really crappy to do to someone. And also a very brain dead use of something so nearly infinite in its capabilities for generating new things.
Could art directors replace illustrators, filmmakers replace concept artists? Sure. But I think something different is likely to happen which is that artists will use these tools to go further. The "bicycle of the mind" Steve Jobs always used to talk about.
As someone who put in the proverbial ten thousand hours in art and technology, I'm completely excited about what this tech enables.
Now, that doesn't mean people's fears aren't valid. Desktop publishing decimated the traditional printmaking industry. Many artisans lost their living. But many small independent publishers were suddenly able to put out their own material to a vast audience. Similar with the internet.
I think it ultimately comes back around to things like universal basic income, or what have you. Rather than taking away amazing tools, we should solve the problem of economic fear that automation and displacement has and continues to bring.
Anyway, this is all so new that these are just my roughly formed personal opinions. I could very well be wrong! But I feel like a lot of the debate is superficial. That there's a lot of misconception about how these models really work. And it also reveals a lot of holes in our own understanding of what it means to think, be original, etc.
What would be brain dead about making lots of money? The popular artist has already found market fit and potentially has an easy corpus of work on which to train.
But even if something IS completely derivative, some would say it's fair game. I'm not trying to judge — just trying to point out that if someone chooses to blatantly just copy another living artist, it seems like that's more on them than the model itself.
And I personally love remixing styles! But I find that adding even one additional artist to a prompt is more interesting — you immediately get somewhere completely novel.
For “machine learning”, though, it seems the big players aren't even going to pay artists for the work they use, let alone try to avoid hurting their livelihoods.
Let me answer it in what I think is the most generous reading:
Copyright is not meant to be used a bludgeon to inhibit the sharing of ideas, it's meant to incentivize the creation of works while still providing a mechanism to enhance the commons [0].
The limitations of copyright are:
* You cannot copyright ideas, only realizations [1]
* Copyrighted works can be copied without the authors consent if it falls under fair use [2]. In particular, the amount republished is considered in addition to whether it harms the original copyrighted author
* To show copyright infringement, you have to show a "substantial similarity" [3] and have the challenged work in question not be protected under other non-infringing uses [4] such as being an "aggregator" or other such service
In particular, if we're having a conversation about copyright but we don't talk about scraping for search (Google, Bing, Yahoo, etc.) or other aggregation services, then I'd ask why Google is allowed to aggregate information but everyone else isn't.
In terms of 'transformative work', I'd point out that many artists copy the style of other artists, sometimes even using code, and while there are many "ethical" dilemmas that artists bring up about this point, this very clearly falls outside of the scope of copyright.
Also, to not talk about the over reach of copyright, the nearly 100+ year copyright terms and the level of automation to demonitize and bully smaller artists using copyrighted work fairly is negligent.
Is this harming artists more than it's helping? If so, why is the argument different now than when it was used in the past for similar issues?
Let's say an artist is mad that their work got used in one of these AI models. How would they prove it? What if it turned out that the model was actually using work from another, older, artist that the angry artist used as an influence and didn't actually use any of the "angry" artists work at all. Should we start suing the angry artist because they made their art "in the style of" the older artist?
I'm all for having these conversations but hot takes like this are simplistic and counter productive.
[0] https://en.wikipedia.org/wiki/Copyright_law_of_the_United_St...
[1] https://en.wikipedia.org/wiki/Copyright_law_of_the_United_St...
[2] https://en.wikipedia.org/wiki/Fair_use
[3] https://en.wikipedia.org/wiki/Substantial_similarity
[4] https://en.wikipedia.org/wiki/Copyright_infringement#Limitat...
"Fair Use" is a uniquely US concept and does not exist in copyright law elsewhere, so for any kind of global release this isn't correct.
> In particular, if we're having a conversation about copyright but we don't talk about scraping for search (Google, Bing, Yahoo, etc.) or other aggregation services, then I'd ask why Google is allowed to aggregate information but everyone else isn't.
Author's Guild v Google is probably the closest we have to a tried case, but Fair Use rulings are both limited to the US and specific in their nature, so I'd be wary of generalising one domain (text) to another (images). Eg. Fair use takes into account harm, and the potential harm to human artists is nontrivial.
> Also, to not talk about the over reach of copyright, the nearly 100+ year copyright terms and the level of automation to demonitize and bully smaller artists using copyrighted work fairly is negligent.
I agree with you - Copyright needs an overhaul and it's offensively lopsided in favor of the powerful as it currently stands.
This is why the Creative Commons organization created CC0 in the first place to get around the differing and/or vague implementations of "public domain" [0].
As the author is most likely American and talking about Americans, I think my critique still mostly holds but you're absolutely right that once we go internationally it gets much more complex.
I only know of the Berne Convention [1] but doing some superficial research, it looks like their "fair use" doctrines are overly vague [2].
[0] https://wiki.creativecommons.org/wiki/CC0_FAQ#What_is_CC0.3F
TL;DR: the models behind these things objects are more like human minds than databases. They perform integration and reproduction as do our minds (within very narrow constraints of course).
These attempts to reason about ethics around them predicated on pearl-clutching about IP and originality miss the point and significant and nature of such models and the systems around them in a forest-trees way.
Attempting to describe what they do and what relationship it bears to prior work is only sensible when you start with the premise that they are more like human minds, than "tools for replication" in any of the senses we have had to reason about as a society before.
A worse problem is certain: our legal system is certain to lag even further behind in both comprehension and ability to respond.
Oh well!
Considering this, I find it incredibly obnoxious to use the term vegan in this fashion, somehow equating training a software on copyrighted material to literally torturing living, feeling and thinking animals for their short lives. It is so completely out of touch with reality that I'm just speechless.
avoider?
teetotaler?
conscientious objector?
On the other hand, government is what can end factory farms. No market mechanism can do it since it actually makes a factory farm more competitive the more it cuts corners and costs on housing animals and fattens them with antibiotics.
The people who don’t care about animal welfare might be persuaded if they realized just how much factory farms contribute to the rise of antibiotic resistant MRSA bacteria and other superbugs. They could make the past pandemic look like a cakewalk.
- Artists: you’re comparing us to animals raised for meat now?
- Animals: you’re saying killing and eating us is equivalent to stealing someone’s picture?
Honestly though, the more I think about this particular analogy the more I like it.
The key to the analogy isn't meant to be about eating animals compared to ripping off art though - it's about personal ethical choices. From the article:
> ... my current mental model is to think about this in terms of veganism, as an analogy for people making their own personal ethical decisions.
how about AI Cannibal? apt, i feel...by your logic. and better than "AI vegan".
everyone has access to human flesh and sadly only a few indulge in this opportunity. how about we leave vegans out of this?
I believe that harming to animals is bad. And yet I still chose to eat them. I am not a vegan, even though I understand the moral reasoning for being one.
I am uncomfortable that generative AIs are trained on images without the artists consent. And yet I still chose to use those models.
If I _was_ a vegan with respect to animal products but chose to use AI models the analogy would not be as fitting for me.
Note that I am not saying that either of my moral choices here are the correct ones!
1. Do you believe harming humans is bad?
2. Would you choose to eat them if they taste just like chicken?
Perhaps the reason you choose not to eat humans is because it is illegal and is punishable to kill humans for food whereas killing animals for food is still acceptable in society. And hence it is an easy low hanging moralistic fruit to be vegan.
It is also likely that you might be harmed if you consume human flesh.(Hello, prions!) There are multiple deterrents to becoming a cannibal vs eating like a non vegan.
You haven’t made a moral choice at all when you consume animal flesh while still believing that harming animals is bad. Just like you haven’t made a moral choice at all when you appropriate other artists’ work without their consent. There is no ‘moral choice’ made here because you chose not to choose when there is no stick to be afraid of..
The only reason why it would be easy for AI art to exist is because it is not punishable. There is no law against it.
In other words, people will continue indulging in their whims as long as there is no stick to punish them.
Your thought?
You seem to be making my argument for me here: this is why I used the vegan analogy.
Note that I am not saying that I have made the moral choice here, in either case!
In both cases here I have decided to prioritize my own convenience over what can be argued as the morally "right" thing to do.
You keep bringing up ‘moral choice’ or the lack of it. That’s very confusing. Is this a morality issue or a legality issue?
Why bring up veganism as the base construct when you don’t identify with it?
Make a better argument.
The question at hand is: Is it ethical for an AI to produce images from images of questionable source. All the current AIs are trained from tons of images from the web, and there's no way to guarantee they weren't polluted with images that they didn't have permission for.
We've decided that it is ethical for humans to do that. But that's at least partly because it's impossible to create art otherwise. You're always exposed to other people's art. An AI doesn't have that problem, and additionally has a much, much better memory... And the actual creation process is different.
I think it's a really interesting ethical dilemma and haven't decided what side I'm on yet.
But like the author, a "vegan" image AI would be very welcome. I think it wouldn't be nearly as useful because a lot of modern concepts would be missing. But it'd still be welcome.
Say artist A is afraid of AI and doesn't want his paintings be used in training. The model will never generate anything like A unless A is copying someone else.
In time the amount of AI generated images is going to grow, it will flood the web and social networks, maybe evolve in new forms, and A will have no impact, no influence, won't be quoted. Deleted from the AI mind means deleted from the social awareness.
It wasn't copied - it was viewed, which is perfectly legal.
The AI does not have a copy of any of the the images it learned from. It merely learned styles and concepts of many images with associated text for those styles and concepts.
I'm wondering if you could erase an artist from a trained model by adding gradients in the opposite direction, erased from the sum of gradients as if it wasn't part of training.
Weights are usually randomly initialized, defining a function. Many images (which are randomly batched on each epoch) are passed through, and a loss function tallied. Backprop gets a gradient, and a training schedule tweaks the weight a little bit. No image has its "gradients" somehow added. The overall function has gradients, but they are not smooth, due to many things such as ReLU being non-smooth. These are not invertible in any sense.
Th networks have many, many irreversible steps in them that truncate, that max pool, that perform dropout during steps, many of these choices done randomly. There are gradient normalizations to clip or stretch to deal with vanishing or exploding gradients, and SOOO many lossy things done that are completely irreversible.
So none of this can possibly work. A network is not simply a sum of images in some format.
The model doesn't save the image itself, but eats the gradients. The question is how we see this process, does it entail a copyright penalty?
No, that would be incredibly slow. The loss is summed over items in a batch, then the gradient is computed once based on the loss, then backpropagation is done. It's done this way by default both in TensorFlow and in Pytorch.
It is possible to sum per entry (see here [1] in pytorch docs) but is is extremely rare. The majority of training loops (this one included) is the standard "do a batch" then call "loss.backward()" which computes only once per batch.
Look at the "backward()" call in pytorch docs [2]: This "Computes the gradient of current tensor w.r.t. graph leaves." If this were done per item and being summed, there would be no need for this in the training loop, and nearly every training loop I've seen has this.
Here's pytorch explaining all this in detail [3].
And in any case, since the functions in the pipeline contain many non-smooth terms, the information required to restore an image, even if only one were done, is not in the gradient, any more than the slope of a line tells you the x-intercept.
Another way to think of it: even for perfectly smooth functions, derivatives lose information that cannot be recovered.
[1] https://pytorch.org/functorch/stable/notebooks/per_sample_gr...
[2] https://pytorch.org/docs/stable/generated/torch.Tensor.backw...
[3] https://pytorch.org/tutorials/beginner/blitz/autograd_tutori...
There is no "Should" that makes sense here either way.
> Are the AI generated images copyrighted?
No, this is already established in the US.
> Are they protected behind any form of paywall?
You can if you want.
> Are those using the images generated by the AI responsible for derived works?
Derived works of the images? Like I generate a picture of a dog, and you draw a dick on it? Obviously not.
> No, this is already established in the US.
The case you're likely thinking of [1] is not at all this case. Thaler tried to have the AI assigned the copyright, then transfer it to himself. USPTO denied that the AI could have the copyright, so it could not be transferred.
This is completely the opposite of what this case is. He wanted AI to be the sole copyright holder, which the court ruled cannot be. Any creative human input to make the AI create the image (text prompts, post selection from among any images to find one that is good) is fully copyrightable by the person.
The text that an artist uses and then selects final images from are both considered creative enough to make the final copyrightable, just as if they entered numbers into Photoshop tools to make an image the way they wanted it. Plenty of these are creative enough and take significant tweaking by the authors via the text they fiddle with to be copyrightable, the same as the text they created to generate the work.
[1] https://www.copyright.gov/rulings-filings/review-board/docs/...
>> Are the AI generated images copyrighted?
you wrote
> No, this is already established in the US.
What cases are you then using to conclude this?
>It is not copyrighted by default anymore than any other trivial non de minimis effort
I doubt de minimus applies, since the images coming out of this AI are no where near any originals that I've seen - they may similar style, but they're so completely different that they are almost all certainly copyrighted by the author.
For example, the simplest, littlest effort to snap a picture gives it full copyright. This AI takes more work to use than snapping a picture.
And de minimis would be on the infringer to prove - the owner still has a copyright until proven otherwise. So the images are most likely copyrighted.
Anything copyrightable in the US is automatically copyrighted by default upon creation. It's why any photo taken is immediately copyrighted by the photographer. Any painting is automatically copyrighted by the painter. Any computer art is automatically copyrighted by the computer operator.
I'd expect most of these works are copyrighted just as much as if someone painted them, or used blender to make them, or used a collage tool to merge self owned photos.
But not all pictures that a camera takes are copyrighted. The same applies to AI generated content.
To me asking if AI generated images are copyrighted implies the latter issue. If you're trying to answer a different question it's just an uninteresting semantic debate.
What case are you using to claim AI images are not copyrighted? The one I cited? Or another? The one I cited, the only one I can find, is not relevant here.
Why are you avoiding citing a case to back up your claim?
No, again. The AI can be human operated, and thus it can produce copyrightable works. But it can also be operated by other things, so it is not by default copyrighted. Your own case as well as the monkey selfie are just fine for this.
> Why are you avoiding citing a case to back up your claim?
Our debate here is not defending premises with evidence, it's you not understanding what the premise is.
And which work in the entire thread up to you moving those goalposts was not human operated?
Blog post: "Stable Diffusion is a new “text-to-image diffusion model”". "you can try..." "type in a text prompt..." "and added a prompt..." and on and on.
In every example in the post, the images were created by human inputs.
Top post in this chain: "Are the AI generated images copyrighted?" Notice the "the" instead of "Are AI generated images copyrighted?" Did you miss that word? What "the AI" do you think this refers to? Some other AI that you made up to split hairs over, or "the" AI under discussion that is creating the art in this thread?
Did you notice no one except you in the replies to that comment misunderstood it? Not one person misread it, except you.
>it's you not understanding what the premise is.
Not a single item in the chain above you was about anything other than humans using AI to make images. You made up a premise to split hairs over, but it looks like that is your modus operandi judging from past posts.
I think you need to read the thread carefully before claiming others don't understand.
Have a good day.
>To me asking if AI generated images are copyrighted implies the latter issue. If you're trying to answer a different question it's just an uninteresting semantic debate.
I think it is important to note that works by the ai are not inherently copyrighted, and they certainly aren't copyrighted to the ai. It is different to by copyrightable.
It's ok if you think that's not a meaningful distinction. But I explained it multiple times and you didn't get it.
> it looks like that is your modus operandi judging from past posts.
Looks like yours is get tilted over nothing