17 ms·

Copyright infringement in AI art

139 points by AndrewDucker 4y ago

gedy 4y ago
Independent of whether one thinks AI generated images are really art, I feel like they are getting very good at simulating "inspiration" and other indirect ways of copying other artists and their styles. I think the last thing we need to do is start incorporating copyright and other filters into these models to filter out content. This just empowers others to control and monopolize content, worse than today.
- revicon 4y ago
  Agreed, per the article “ style and a ‘look and feel’ are not copyrightable”
  - Hemospectrum 4y ago
    I wonder if they can be trademarked. If not now, maybe later.
    - nl 4y ago
      Yes they probably can be trademarked. But a big question is why? Trademarks are very expensive. It makes sense for a company logo that was AI generated, but spending 100K+ for very limited protection of a piece of art seems questionable.
      - verdverm 4y ago
        Trademarks are actually much cheaper than that, I paid a few thousand dollars for mine. The price range your quoting is more typical of patents
        nl 4y ago
        Was it international? And in how many categories? It adds up very quickly.
        clcaev 4y ago
        You can lose your trademark if you do not actively enforce violations. Hence, the full cost of a trademark should include its legal maintenance.
        verdverm 4y ago
        that is an unknown quantity which one cannot estimate
    - surfacedetail 4y ago
      It's more difficult to trademark stuff like style, there are certain rules of what can be trademarked; long story short, non-graphical representations need to be clear, precise, self-contained, easily accessible, intelligible, durable and objective. Trademark also gives a more limited protection in some ways.
Imnimo 4y ago
I'm not copyright expert, but I would have imagined the answer is relatively straightforward: if an image would be infringing if drawn by a human, it's also infringing if drawn by a model, and similarly for non-infringing. Does the manner in which the image is generated play into whether it infringes a copyright?
- revicon 4y ago
  I’m also not a copyright expert, so I wonder about things like if I draw a (crappy) picture of Spider-Man and slap it on a tshirt, can I sell it without violating copyright? I’m sure there’s established law about this kind of thing.
  - eru 4y ago
    My guess is that your badly drawn Spider-Man T-shirt would violate trademark, but not copyright.
  - egypturnash 4y ago
    If your shitty Spidey is a drawing you did yourself, you are not violating copyright. You are however blatantly violating Disney’s collection of trademarks on “Spider-Man”, the distinctive appearance of the character, and a host of associated marks.
- kube-system 4y ago
  The general consensus in the IP space (in the US at least) is to treat AI no differently than any other tool.
- eru 4y ago
  > Does the manner in which the image is generated play into whether it infringes a copyright? Yes, it does. Let's leave AI out for a moment. If you lock yourself into your room with no Internet access, and you draw (or write etc) something independently that just happens to look exactly like some already existing copyright-ed work, you are still not infringing any copyrights. If you produced the same work by making a copy, you would infringe copyright. See https://en.wikipedia.org/wiki/Clean_room_design https://en.wikipedia.org/wiki/Clean_room_design for how that's relevant in real life. However you can infringe on patents and trademarks with independent work, ie even if you don't just make a copy.
olladecarne 4y ago
ML revolutionized translation a long time ago and the demand and pay for translators went down. It also used the works of thousands of humans who translated text and they never saw a dime. Copyright applies to translations so we've already gone through something similar. The same will happen with art and every other medium that ML touches. I have a love/hate relationship with ML because of this. It seems that there will be a painful transition period as many humans are displaced, probably even average coders like me. In a way, it's no different than what a human does but humans can't scale like computers. I hope someone can engineer a new economic system that works for what's coming.
- TaylorAlexander 4y ago
  Community ownership. We need everyone to be part owner in the economy that they rely on. Then productivity gains are good for people even if there is less work to do.
- matheusmoreira 4y ago
  > I hope someone can engineer a new economic system that works for what's coming. What's coming is post-scarcity for all kinds of intellectual work. There will be nothing to economize: supply of both labor and product will be infinite. I don't want corporations engineering a made up economy where there is none.
  - soulofmischief 4y ago
    I predict a new system arises in which styles themselves are copyrightable and eligible for royalties. Human-sourced styles will become trend, as people seek originality to establish identity in an otherwise indistinguishable world of machine-driven experiences.
    - lioeters 4y ago
      I think your idea is prescient. I can imagine a market of proprietary "styles", I suppose in the form of pre-trained networks/datasets, that can generate any number of concrete instances of images, texts, and other media.
thriftwy 4y ago
What I'm dreaming of is an AI which can make any song covered by any performer or band in history of music. Totally doable but would likely put you on the most wanted list above all the drug lords.
- stemlord 4y ago
  I'm dreaming of an AI that can determine whether a song was AI-generated or is truly one degree from human hands
  - cgrealy 4y ago
    I highly doubt that’s even theoretically possible, given the vast range of variety and quality of music created by humans.
    - angusturner 4y ago
      Feel like people would have said the same with images 6 months ago tbh. (I mean sure, you could get a fake photograph of a person, but the text control that clip has unlocked is just wild…)
      - cgrealy 4y ago
        I’m not sure I get your argument. I’m saying that it’s impossible to distinguish between machine written music and human written music. You could maybe make the argument for distinguishing between a human and a machine playing a piece of music. But if I write a song that just goes A, D, C, G on a loop (plenty of songs like that), how would you determine that a machine song that goes A, D, G, C is not human written?
        angusturner 4y ago
        My bad, I misunderstood you! I agree with you.
  - thriftwy 4y ago
taylorius 4y ago
Copyright will need to be rethought in this new world, where artistic interpretations of copyrighted works can be mechanically produced. Personally I don't see how it can be made coherent.
wesleywt 4y ago
If you learn painting by copying other painters then paint a new painting, copyright infringement? Unfortunately many small artists making a living by designing logos will dissappear with this technology.
- patengineer 4y ago
  It depends on what the painter does with the 'new painting'. If it is solely that the painter(considered a student) is learning how to paint, then copyright infringement would not apply as it would be covered under 'fair dealings'. This would apply in the UK, perhaps Europe. I'm not familiar with US law. Quoting the Act - "Fair dealing with a artistic work for the purposes of private study does not infringe any copyright in the work." If the painter then goes on to sell this painting, well, that's copyright infringement because they are making a copy of a painting available to the public.(assuming no licencing or authorisation from the legal copyright owner was previously obtained)
  - sgift 4y ago
    Either you or I misread the question. I think it wasn't about selling a copy. It was about using copying to learn how to paint (which is fine, as you've stated) and then produce your own painting which you then sell. The sold painting isn't a copy. So .. no copyright infringement?
    - patengineer 4y ago
      Ah, I think it depends how you interpreted the question. I took 'new painting' to mean a copy of other painter's work. Looking at it from your perspective(which on reflection is the right interpretation), I would say no copyright infringement assuming the 'new painting' is an original creation i.e. not a duplicate of an existing piece of art by another artist. Plenty of people, myself included, learn creative arts by mimicking others and then developing our own style which might be similar to the artists we got our inspiration from. I think as long as it's clear that there is no duplication of the original works(for sale), then all is okay. So if I learnt my art by mimicking Van Gogh and my new paintings are of a similar Van Gogh style, as long as they are original creations, no copyright infringement imo.
whywhywhywhy 4y ago
>Moreover, the developers of these tools are aware of the potential pitfalls of producing exact replicas of art in their training datasets. OpenAI admitted that this was a problem in some of the earlier iterations of the program, and they now filter out specific instances of this happening. This is the most telling thing. They have a system, which might not even be part of the "AI" it might simply be they are comparing outputs for close matches to the OG dataset and then hiding them from the end user. I mean we've seen OpenAI's solution to their un-diverse data set so we know what standards they work to.
- kxq 4y ago
  they addressed duplicates during pre-training (by removing duplicate examples in the dataset), as documented here https://openai.com/blog/dall-e-2-pre-training-mitigations/ https://openai.com/blog/dall-e-2-pre-training-mitigations/
jefftk 4y ago
The article gives several examples of AI-regenerated Starry Night, and says: It looks like it, but it’s not the exact same thing, it’s almost as if the AI is drawing it from memory, which in some way it is, it’s re-constructing what Starry Night looks like. This seems to ignore that those are pretty clearly derivative works of Starry Night? Famous UK example which really stretched the definition of a derivative work ("London bus crossing a bridge, black and white") https://en.wikipedia.org/wiki/Temple_Island_Collections_Ltd_v_New_English_Teas_Ltd https://en.wikipedia.org/wiki/Temple_Island_Collections_Ltd_... (Not a lawyer)
_trampeltier 4y ago
I searched just quick, and for ex. Getty does not allow to use pictures for ML. Anyway, I think is pretty clear, usually it is not allowed, exept. user upload a picture to a platform and give all rights away because the platform say it in the EULA.
angusturner 4y ago
The question that fascinates me about this whole debate is on what basis (if any) one draws the line between a machine learning from copyrighted data vs a human learning from copyrighted data. And I do think “learning” is the correct metaphor here. There are good reasons to believe that the brain is doing something conceptually similar to generative statistical models. i.e compressing inputs (/percepts) into some kind of abstract representation space. Even if we don’t understand the specific details yet. As far as I’m concerned, it makes no sense to prevent training on copyright data. If I use a trained model to substantially reproduce parts of a copyrighted work, feels like the issue is largely with me and what I do with the content. These models are just tools after all.
- anankaie 4y ago
  I suspect that the difference is in knowing versus unknowing use, and the deliberate choice of including learned elements or not on the part of the user. In the case of the ML model, it can reproduce a part of a training example perfectly, without the user knowing this happened (see the early Copilot analysis wherein it was possibly to get it to quote well known blocks of code verbatim). This is much less likely to happen when a user is consulting her own memories, though obviously never say "never".
- trention 4y ago
  On the basis that the "machine" is not a human and there is nothing that necessitates allowing neural networks to do all the things humans are permitted to do.
  - angusturner 4y ago
    I don’t think it makes sense to talk about these systems as though they have any agency (and I think I was careful not to). Of course there is nothing that necessitates letting these systems do what humans do. What I am saying is that I am allowed as a human to ingest copyrighted material, learn from it, form mental abstractions and use those learnings to generate new art. Why then, would I be forbidden from externalising some of that process into an artificial system that operates on similar principles. The trained model obviously does not have agency. It is a creative and computational tool, much like any other digital tool or product.
JohnHaugeland 4y ago
What's fascinating to me is how many of the people decrying what's happening as unfair to artists also brag about being software pirates, where actual human labor is going unpaid for
- jazzyjackson 4y ago
  > unfair to artists > actual human labor say what?
6gvONxR4sf7o 4y ago
I think it’s clear that what’s happening is (mostly) legal, but the better question is whether it should continue to be. It will be a bad outcome if a few people profit off of a huge group of people’s labor without their consent and without compensating them, in order to put that large group out of work. Technological advances change the labor market all the time, but it’s absolutely different to exploit your own labor in this kind of way to put you out of a job. I think if someone wants to build a research ML model, it should be fine to use whatever data you can get your hands on. But if you want to make a commercial model, you should have to have the content creators’ permission to use their data to train your product.
- visarga 4y ago
  It's not just the model publishers that profit from generative models, the users gain something as well. And the access barrier being lowered, now many more people can benefit. Even the artists can benefit. Let's not forget the contribution of the prompt to the final result. The users have a finger in there too. That means they are directly involved in the creative process, they deserve a claim to the copyright.
  - 6gvONxR4sf7o 4y ago
    If the artists benefiting too is a reason to allow this, then the artists having a say in how their data is used shouldn’t be a problem.
    - visarga 4y ago
      Training neural nets is a novel transformative process - it only takes gradients for each training example, it doesn't save any image, just take the gradients and sums them up overlapping the influence of all the examples, mixing up everything in a way we can't tell which feature comes from which training example. More like inverting the dataset into abstract principles and then generating top-down novel examples. It's like learning, but differs in that humans can't possibly see so much data. It's not like copying, but it can generate copies of training examples as an edge case. It feels like stealing, but you can't assign who is the victim, maybe it draws a bit from multiple sources of inspiration. It's very creative and inspires all sorts of idea juxtapositions, but some people consider it's only parroting known styles. In the end I think we will want to use the new toys so we will allow training on anything. Maybe a "Do-not-use-for-training-neural-nets" flag will be observed by the conscientious labs.
amelius 4y ago
Copyright wars will start when somebody enters the prompt "... in the style of Disney" into Dall-E.
6stringmerc 4y ago
Okay I dislike the US copyright system immensely - virulent yearning for reasonable terms and ending sampling payments forever more… But let’s get practical here for a minute - if an ape can’t own the copyright to a selfie, then nothing generated by a machine - no matter how sentient - will legally be protected. Just keep that in mind, because the ML results will not abide by human creation protections.
chucklenorris 4y ago
I have a different but related question. Let's say i use a dall-e like ai to generate character models for my next game. Are they copyrightable? What would stop someone to outright rip off the models from my game and use them in other works?
xor99 4y ago
If this all worked in a fair way it would be treated similarly to sampling in music where the original artists get a royalty each time a track is used. If that source is clear like where it's "x in the style of y" using these elements of x explicitly. However, the complexity is often too great and how can you prove fractional sampling of your work? Very tricky as the article points out! It's also blocking many new forms of art if you enforce the fractional sampling idea in law. On the other hand, small startups selling to customers who need to produce proprietary content could incur a risk with AI generated products. Aka the customer won't want to risk their product by including potentially infringing items (like with Co-Pilot's potential use of unpaid labour or open source code and other licences).
jheriko 4y ago
i never expect anyone in silly valley to do due diligence... it was a pretty obvious problem from the outset. if you are making a product, then using copyrighted material is far from fair use, at least outside of the US.
fleddr 4y ago
I have to admit that I did not see it coming. Ignorant of AI's capabilities, I always imagined it to automate away our boring, mechanical, repetitive stuff. Which in some hopeful utopian scenario would mean that we could focus on the humanities: family, health/sports, culture and anything creative. For AI to come for art, at the very core of our humanity (the oldest cave painting is 64K y/o) is quite the confrontation indeed. One that will rapidly escalate more broadly (poetry/writing, photography, cinema, music) and more deeply. The amount of AI art will explode, in turn serving as input for even more AI art. The speed at which this will happen suggests to me we need to be far more alarmist in coming up with answers to many questions. Philosophical, legal, socio-economical, etc. I really think the attitude "I'm not an artist so this will only benefit me" is shortsighted, and I'm saying that as a non-artist myself. It's like we're at the beginning of a very short and steep exponential curve of total disruption.
- ramoz 4y ago
  > The amount of AI art will explode, in turn serving as input for even more AI art ... The speed at which this will happen suggests to me we need to be far more alarmist in coming up with answers to many questions ... I agree. I don't want to live in a AI-hamster-wheel future. Human creativity is driven by emotion, curiosity, intuition... etc. AI creativity is basically an advanced form of copy-paste/blend using context -- all of which is learned from human content. We should be working harder to empower human-creativity... which is an argument AI artists often make in favor of generative creativity... however this narrative falters when money gets cheap and automation disrupts whatever economies exist for creatives (esp writers and artists).
  - trention 4y ago
    Brynjolfsson and McAfee made a simular argument a decade ago - we will use automation to replace rather than augment human labor, the result being a local optimum far from the global one (that's in terms of results, the dystopian disempowerment of humanity is another story.
- LampDrewNear 4y ago
  This is how I feel as well. I'm a software engineer by trade, but I'm really much more interested in writing music and stories. I always imagined myself retiring the day I have saved up enough money so I could focus on art and creativity. My life certainly has felt much less meaningful since I saw what DALL-E 2 can produce. And my main concern is not that of professional artists' financial situations. I'm mainly worried about how the massive influx of computer-generated content will inflate away the meaning of human-created art. And I'm afraid that when art becomes so good and so customized to each individual consumer, we won't have much common culture left. I think this Penny Arcade strip well puts into words how I feel: https://www.penny-arcade.com/comic/2022/06/06/dalliance https://www.penny-arcade.com/comic/2022/06/06/dalliance In a way, I'm so mad at my fellow engineers. There's so much good we could do, is really killing the creative arts the right thing to do? And I don't believe for one second this talk about AI just being a tool, and prompt-engineering being a new craft for artists to learn. I think this kind of interface will quickly go away, and soon enough AI apps will look much more similar to TikTok or YouTube than they do today.
fleddr 4y ago
It seems the more interesting legal question is not regarding input, but whom owns the copyrights of the output, if any. Interesting article: https://www.technollama.co.uk/dall%C2%B7e-goes-commercial-but-what-about-copyright https://www.technollama.co.uk/dall%C2%B7e-goes-commercial-bu... If I interpret that article correctly, at least Dall-e does not claim copyright on the generation nor does it grant it to the person writing the prompt. The copyright question seems entirely unanswered. Perhaps that should be codified in law explicitly: AI generated art cannot be copyrighted. If it was created from the collective input of humanity (unasked), its output may as well benefit all of humanity as well. It's hard to make the case that somebody writing a prompt deserves full copyright protection. Even if so, just add one word and one can claim it's once again an "original". Not respecting AI copyrights would also help protect the vast area of commercial art, as it lowers its value if it cannot be legally protected. This could get hairy though, because what happens if you take AI output as input and then start manually tweaking/editing it. It would then be a mixed AI/human works.
- archontes 4y ago
  It's not entirely unanswered. https://www.smithsonianmag.com/smart-news/us-copyright-office-rules-ai-art-cant-be-copyrighted-180979808/ https://www.smithsonianmag.com/smart-news/us-copyright-offic...
  - anonymousab 4y ago
    I wonder if that applies to AI generated code as well.
  - fleddr 4y ago
    Thanks, did not know that. I consider it to be good news.
  - Pulcinella 4y ago
    This just says that the AI can’t own the copyright, just like how Photoshop, the software program, can’t be granted a copyright but a person using Photoshop can.
donkarma 4y ago
only thing that should worry people is the fact that it is cheap and bad much like machine translation which would make people rely on it, downgrading the baseline of art for companies forever
dahart 4y ago
> Assuming that the input phase is fine and the datasets used are legitimate, then most infringement lawsuits may end up taking place in the output phase. And it is here that I do not think that there will be substantive reproduction to warrant copyright infringement. This might be true, but feels like maybe the most practical answer to this whole question is to focus on the inputs and to establish ways for artists to control whether their art is used as input. Don’t let legally questionable content into training, and you automatically prevent getting legally questionable output. Why would we do anything else, really? It’s potentially a large assumption to make that the datasets are currently legitimate under the lens of copyright. Pulling images off the internet just because they’re accessible is not valid, and humans aren’t currently allowed to do that under the law. But it might be comforting to artists and practical for both artists and AI researchers if artists could specify a form of copyright or a license that allows ML training provided that their work is not recognizably reproduced. Certainly datasets like Flickr with Creative Commons search tools allows a wide variety of public use that is entirely legal. What reasons do we have to not establish something similar for machine learning?
kmeisthax 4y ago
> It looks like it, but it’s not the exact same thing, it’s almost as if the AI is drawing it from memory, which in some way it is, it’s re-constructing what Starry Night looks like. The blog post also has a picture of four slightly-different recreations of a Van Gogh painting. The thing is, under current copyright law in the US, all four of those recreations would be legally considered the same image. The standard for copyright infringement is access plus substantial similarity. It would be easy to construe access to a training image set through ML software that was trained on that set; and thus any regurgitated output would be infringing.
janef0421 4y ago
I don't know if "copyright infringement" is really a useful legal construct here. The issue is not so much that a person's work is being duplicated, but that a person's work is being used to reduce the value of their labour without consent or compensation. It's probably more under unfair competition. Ultimately, though, this issue probably requires new constructs and laws, which is unlikely to happen. The law is the ultimate expression of the power of the dominant class, and these sorts of AI/ML systems benefit the socially dominant class of business owners and investors.
mysterydip 4y ago
Isn't this the other half of the equation for "I host all my pictures/art on site xyz because it's free" and somewhere buried in the accepted terms of use was "you give us permission to use it for anything we want"?
antioppressor 4y ago
Yeah, but what happens if you explicitly state that your art/images must not be included in AI training?
- jacobsievers 4y ago
  If I've legitimately obtained a copy of your work, then I'm entitled to use it, read it, even make my own copies for personal use. Any new work I attempt to distribute would be measured against your original to determine infringement.
  - antioppressor 4y ago
    So if you use part of my image, which has watermark embedded via noise and it is found in your image, then you are fucked. Good.
kazinator 4y ago
> I prompted DALL·E for “Crumbling ruined city, sci-fi digital art by Simon Stålenhag”, and it produced a few images that were very much in the style of Stålenhag. The image shown is nothing has some concepts in common with Stålenhag, but it has a lot more depth. I can't find any Stålenhag which evokes the same way. In Stålenhag's work, or what of it is easily accessible online, there are some recurring gimmicks, which are absent in the DALL-E work, or present in a disguised form. DALL-E doesn't juxtapose any mint condition 1970's Volvos or Saabs against a background showing a menacing superstructure from the distant future. It doesn't invoke any cheesy mechanical monsters or characters inspired by popular Sci-Fi. For instance, " Stålenhag lays out everything plainly. In about a millisecond, you see what you're supposed to see: the easily recognizable elements like ordinary people or retro objects, against fantasy objects, set in some nice lighting: atmospheric effects like haze, sunset or whatever that completes the mood. In the DALL-E there are some old objects, but they are not instantly seen, and they are in old condition. For instance what looks like the wreck of an airliner appears in the rubble in the foreground, but you might not see it right away. That makes the scene plausible. It is more of a spooky glimpse at something that could actually someday be, somewhere. It has the DALL-E weird elements in it. Like why do the cables on what looks like a telephone pole seem to connect to elements in the distant background; DALL-E seems to have perpetrated an https://en.wikipedia.org/wiki/Impossible_object https://en.wikipedia.org/wiki/Impossible_object there, but it's not wham-in-your-face; not everyone will spot it right away. There is what looks like a channel of water through the debris, but what looks like a reflection in it is so weirdly wrong that you can't make up your mind whether it's a weird local reflection or a glimpse through an aperture in the bottom of a floating sky structure, where a distant reflection is shining up from some ocean below. Stålenhag's work doesn't have any of these sorts of demented optical labyrinths in store for the viewer. He's like a Norman Rockwel*, targeting the nostalgia of that viewer who grew up on a diet of 70's, 80's and 90's anime and sci-fi, and is longing for the 20th century. -- * E.g. check out the cheesy fromage called "The Long Way Home". Rockwell plus Storm Troopers. This just screams "Hey, you look transparently like a Generation X geek; would you be interested in buying a limited edition print of me?"
andybak 4y ago
Does anyone have any thoughts on how this might relate to the Marvin Gaye estate vs Robin Thicke and Pharrell Williams "Blurred Lines" case? I know music has very different copyright rules to visual images but that case springs to mind as it was a situation where infringement was judged to have occurred on much looser grounds than previously required. Could a similar expansive judgement happen in the case of AI art and shift the goalposts so to speak?

139 points by AndrewDucker 4y ago

gedy 4y ago

Independent of whether one thinks AI generated images are really art, I feel like they are getting very good at simulating "inspiration" and other indirect ways of copying other artists and their styles. I think the last thing we need to do is start incorporating copyright and other filters into these models to filter out content. This just empowers others to control and monopolize content, worse than today.

revicon 4y ago
Agreed, per the article “ style and a ‘look and feel’ are not copyrightable”
- Hemospectrum 4y ago
  I wonder if they can be trademarked. If not now, maybe later.
  - nl 4y ago
    Yes they probably can be trademarked. But a big question is why? Trademarks are very expensive. It makes sense for a company logo that was AI generated, but spending 100K+ for very limited protection of a piece of art seems questionable.
    - verdverm 4y ago
      Trademarks are actually much cheaper than that, I paid a few thousand dollars for mine. The price range your quoting is more typical of patents
      - nl 4y ago
        Was it international? And in how many categories? It adds up very quickly.
      - clcaev 4y ago
        You can lose your trademark if you do not actively enforce violations. Hence, the full cost of a trademark should include its legal maintenance.
        verdverm 4y ago
        that is an unknown quantity which one cannot estimate
  - surfacedetail 4y ago
    It's more difficult to trademark stuff like style, there are certain rules of what can be trademarked; long story short, non-graphical representations need to be clear, precise, self-contained, easily accessible, intelligible, durable and objective. Trademark also gives a more limited protection in some ways.

Imnimo 4y ago

I'm not copyright expert, but I would have imagined the answer is relatively straightforward: if an image would be infringing if drawn by a human, it's also infringing if drawn by a model, and similarly for non-infringing. Does the manner in which the image is generated play into whether it infringes a copyright?

revicon 4y ago
I’m also not a copyright expert, so I wonder about things like if I draw a (crappy) picture of Spider-Man and slap it on a tshirt, can I sell it without violating copyright? I’m sure there’s established law about this kind of thing.
- eru 4y ago
  My guess is that your badly drawn Spider-Man T-shirt would violate trademark, but not copyright.
- egypturnash 4y ago
  If your shitty Spidey is a drawing you did yourself, you are not violating copyright. You are however blatantly violating Disney’s collection of trademarks on “Spider-Man”, the distinctive appearance of the character, and a host of associated marks.
kube-system 4y ago
The general consensus in the IP space (in the US at least) is to treat AI no differently than any other tool.
eru 4y ago
> Does the manner in which the image is generated play into whether it infringes a copyright? Yes, it does. Let's leave AI out for a moment. If you lock yourself into your room with no Internet access, and you draw (or write etc) something independently that just happens to look exactly like some already existing copyright-ed work, you are still not infringing any copyrights. If you produced the same work by making a copy, you would infringe copyright. See https://en.wikipedia.org/wiki/Clean_room_design https://en.wikipedia.org/wiki/Clean_room_design for how that's relevant in real life. However you can infringe on patents and trademarks with independent work, ie even if you don't just make a copy.

olladecarne 4y ago

ML revolutionized translation a long time ago and the demand and pay for translators went down. It also used the works of thousands of humans who translated text and they never saw a dime. Copyright applies to translations so we've already gone through something similar. The same will happen with art and every other medium that ML touches. I have a love/hate relationship with ML because of this. It seems that there will be a painful transition period as many humans are displaced, probably even average coders like me. In a way, it's no different than what a human does but humans can't scale like computers. I hope someone can engineer a new economic system that works for what's coming.

TaylorAlexander 4y ago
Community ownership. We need everyone to be part owner in the economy that they rely on. Then productivity gains are good for people even if there is less work to do.
matheusmoreira 4y ago
> I hope someone can engineer a new economic system that works for what's coming. What's coming is post-scarcity for all kinds of intellectual work. There will be nothing to economize: supply of both labor and product will be infinite. I don't want corporations engineering a made up economy where there is none.
- soulofmischief 4y ago
  I predict a new system arises in which styles themselves are copyrightable and eligible for royalties. Human-sourced styles will become trend, as people seek originality to establish identity in an otherwise indistinguishable world of machine-driven experiences.
  - lioeters 4y ago
    I think your idea is prescient. I can imagine a market of proprietary "styles", I suppose in the form of pre-trained networks/datasets, that can generate any number of concrete instances of images, texts, and other media.

thriftwy 4y ago

What I'm dreaming of is an AI which can make any song covered by any performer or band in history of music. Totally doable but would likely put you on the most wanted list above all the drug lords.

stemlord 4y ago
I'm dreaming of an AI that can determine whether a song was AI-generated or is truly one degree from human hands
- cgrealy 4y ago
  I highly doubt that’s even theoretically possible, given the vast range of variety and quality of music created by humans.
  - angusturner 4y ago
    Feel like people would have said the same with images 6 months ago tbh. (I mean sure, you could get a fake photograph of a person, but the text control that clip has unlocked is just wild…)
    - cgrealy 4y ago
      I’m not sure I get your argument. I’m saying that it’s impossible to distinguish between machine written music and human written music. You could maybe make the argument for distinguishing between a human and a machine playing a piece of music. But if I write a song that just goes A, D, C, G on a loop (plenty of songs like that), how would you determine that a machine song that goes A, D, G, C is not human written?
      - angusturner 4y ago
        My bad, I misunderstood you! I agree with you.
- thriftwy 4y ago

taylorius 4y ago

Copyright will need to be rethought in this new world, where artistic interpretations of copyrighted works can be mechanically produced. Personally I don't see how it can be made coherent.

wesleywt 4y ago

If you learn painting by copying other painters then paint a new painting, copyright infringement? Unfortunately many small artists making a living by designing logos will dissappear with this technology.

patengineer 4y ago
It depends on what the painter does with the 'new painting'. If it is solely that the painter(considered a student) is learning how to paint, then copyright infringement would not apply as it would be covered under 'fair dealings'. This would apply in the UK, perhaps Europe. I'm not familiar with US law. Quoting the Act - "Fair dealing with a artistic work for the purposes of private study does not infringe any copyright in the work." If the painter then goes on to sell this painting, well, that's copyright infringement because they are making a copy of a painting available to the public.(assuming no licencing or authorisation from the legal copyright owner was previously obtained)
- sgift 4y ago
  Either you or I misread the question. I think it wasn't about selling a copy. It was about using copying to learn how to paint (which is fine, as you've stated) and then produce your own painting which you then sell. The sold painting isn't a copy. So .. no copyright infringement?
  - patengineer 4y ago
    Ah, I think it depends how you interpreted the question. I took 'new painting' to mean a copy of other painter's work. Looking at it from your perspective(which on reflection is the right interpretation), I would say no copyright infringement assuming the 'new painting' is an original creation i.e. not a duplicate of an existing piece of art by another artist. Plenty of people, myself included, learn creative arts by mimicking others and then developing our own style which might be similar to the artists we got our inspiration from. I think as long as it's clear that there is no duplication of the original works(for sale), then all is okay. So if I learnt my art by mimicking Van Gogh and my new paintings are of a similar Van Gogh style, as long as they are original creations, no copyright infringement imo.

whywhywhywhy 4y ago

>Moreover, the developers of these tools are aware of the potential pitfalls of producing exact replicas of art in their training datasets. OpenAI admitted that this was a problem in some of the earlier iterations of the program, and they now filter out specific instances of this happening. This is the most telling thing. They have a system, which might not even be part of the "AI" it might simply be they are comparing outputs for close matches to the OG dataset and then hiding them from the end user. I mean we've seen OpenAI's solution to their un-diverse data set so we know what standards they work to.

kxq 4y ago
they addressed duplicates during pre-training (by removing duplicate examples in the dataset), as documented here https://openai.com/blog/dall-e-2-pre-training-mitigations/ https://openai.com/blog/dall-e-2-pre-training-mitigations/

jefftk 4y ago

The article gives several examples of AI-regenerated Starry Night, and says: It looks like it, but it’s not the exact same thing, it’s almost as if the AI is drawing it from memory, which in some way it is, it’s re-constructing what Starry Night looks like. This seems to ignore that those are pretty clearly derivative works of Starry Night? Famous UK example which really stretched the definition of a derivative work ("London bus crossing a bridge, black and white") https://en.wikipedia.org/wiki/Temple_Island_Collections_Ltd_v_New_English_Teas_Ltd https://en.wikipedia.org/wiki/Temple_Island_Collections_Ltd_... (Not a lawyer)

_trampeltier 4y ago

I searched just quick, and for ex. Getty does not allow to use pictures for ML. Anyway, I think is pretty clear, usually it is not allowed, exept. user upload a picture to a platform and give all rights away because the platform say it in the EULA.

angusturner 4y ago

The question that fascinates me about this whole debate is on what basis (if any) one draws the line between a machine learning from copyrighted data vs a human learning from copyrighted data. And I do think “learning” is the correct metaphor here. There are good reasons to believe that the brain is doing something conceptually similar to generative statistical models. i.e compressing inputs (/percepts) into some kind of abstract representation space. Even if we don’t understand the specific details yet. As far as I’m concerned, it makes no sense to prevent training on copyright data. If I use a trained model to substantially reproduce parts of a copyrighted work, feels like the issue is largely with me and what I do with the content. These models are just tools after all.

anankaie 4y ago
I suspect that the difference is in knowing versus unknowing use, and the deliberate choice of including learned elements or not on the part of the user. In the case of the ML model, it can reproduce a part of a training example perfectly, without the user knowing this happened (see the early Copilot analysis wherein it was possibly to get it to quote well known blocks of code verbatim). This is much less likely to happen when a user is consulting her own memories, though obviously never say "never".
trention 4y ago
On the basis that the "machine" is not a human and there is nothing that necessitates allowing neural networks to do all the things humans are permitted to do.
- angusturner 4y ago
  I don’t think it makes sense to talk about these systems as though they have any agency (and I think I was careful not to). Of course there is nothing that necessitates letting these systems do what humans do. What I am saying is that I am allowed as a human to ingest copyrighted material, learn from it, form mental abstractions and use those learnings to generate new art. Why then, would I be forbidden from externalising some of that process into an artificial system that operates on similar principles. The trained model obviously does not have agency. It is a creative and computational tool, much like any other digital tool or product.

JohnHaugeland 4y ago

What's fascinating to me is how many of the people decrying what's happening as unfair to artists also brag about being software pirates, where actual human labor is going unpaid for

jazzyjackson 4y ago
> unfair to artists > actual human labor say what?

6gvONxR4sf7o 4y ago

I think it’s clear that what’s happening is (mostly) legal, but the better question is whether it should continue to be. It will be a bad outcome if a few people profit off of a huge group of people’s labor without their consent and without compensating them, in order to put that large group out of work. Technological advances change the labor market all the time, but it’s absolutely different to exploit your own labor in this kind of way to put you out of a job. I think if someone wants to build a research ML model, it should be fine to use whatever data you can get your hands on. But if you want to make a commercial model, you should have to have the content creators’ permission to use their data to train your product.

visarga 4y ago
It's not just the model publishers that profit from generative models, the users gain something as well. And the access barrier being lowered, now many more people can benefit. Even the artists can benefit. Let's not forget the contribution of the prompt to the final result. The users have a finger in there too. That means they are directly involved in the creative process, they deserve a claim to the copyright.
- 6gvONxR4sf7o 4y ago
  If the artists benefiting too is a reason to allow this, then the artists having a say in how their data is used shouldn’t be a problem.
  - visarga 4y ago
    Training neural nets is a novel transformative process - it only takes gradients for each training example, it doesn't save any image, just take the gradients and sums them up overlapping the influence of all the examples, mixing up everything in a way we can't tell which feature comes from which training example. More like inverting the dataset into abstract principles and then generating top-down novel examples. It's like learning, but differs in that humans can't possibly see so much data. It's not like copying, but it can generate copies of training examples as an edge case. It feels like stealing, but you can't assign who is the victim, maybe it draws a bit from multiple sources of inspiration. It's very creative and inspires all sorts of idea juxtapositions, but some people consider it's only parroting known styles. In the end I think we will want to use the new toys so we will allow training on anything. Maybe a "Do-not-use-for-training-neural-nets" flag will be observed by the conscientious labs.

amelius 4y ago

6stringmerc 4y ago

Okay I dislike the US copyright system immensely - virulent yearning for reasonable terms and ending sampling payments forever more… But let’s get practical here for a minute - if an ape can’t own the copyright to a selfie, then nothing generated by a machine - no matter how sentient - will legally be protected. Just keep that in mind, because the ML results will not abide by human creation protections.

chucklenorris 4y ago

I have a different but related question. Let's say i use a dall-e like ai to generate character models for my next game. Are they copyrightable? What would stop someone to outright rip off the models from my game and use them in other works?

xor99 4y ago

If this all worked in a fair way it would be treated similarly to sampling in music where the original artists get a royalty each time a track is used. If that source is clear like where it's "x in the style of y" using these elements of x explicitly. However, the complexity is often too great and how can you prove fractional sampling of your work? Very tricky as the article points out! It's also blocking many new forms of art if you enforce the fractional sampling idea in law. On the other hand, small startups selling to customers who need to produce proprietary content could incur a risk with AI generated products. Aka the customer won't want to risk their product by including potentially infringing items (like with Co-Pilot's potential use of unpaid labour or open source code and other licences).

jheriko 4y ago

i never expect anyone in silly valley to do due diligence... it was a pretty obvious problem from the outset. if you are making a product, then using copyrighted material is far from fair use, at least outside of the US.

fleddr 4y ago

I have to admit that I did not see it coming. Ignorant of AI's capabilities, I always imagined it to automate away our boring, mechanical, repetitive stuff. Which in some hopeful utopian scenario would mean that we could focus on the humanities: family, health/sports, culture and anything creative. For AI to come for art, at the very core of our humanity (the oldest cave painting is 64K y/o) is quite the confrontation indeed. One that will rapidly escalate more broadly (poetry/writing, photography, cinema, music) and more deeply. The amount of AI art will explode, in turn serving as input for even more AI art. The speed at which this will happen suggests to me we need to be far more alarmist in coming up with answers to many questions. Philosophical, legal, socio-economical, etc. I really think the attitude "I'm not an artist so this will only benefit me" is shortsighted, and I'm saying that as a non-artist myself. It's like we're at the beginning of a very short and steep exponential curve of total disruption.

ramoz 4y ago
> The amount of AI art will explode, in turn serving as input for even more AI art ... The speed at which this will happen suggests to me we need to be far more alarmist in coming up with answers to many questions ... I agree. I don't want to live in a AI-hamster-wheel future. Human creativity is driven by emotion, curiosity, intuition... etc. AI creativity is basically an advanced form of copy-paste/blend using context -- all of which is learned from human content. We should be working harder to empower human-creativity... which is an argument AI artists often make in favor of generative creativity... however this narrative falters when money gets cheap and automation disrupts whatever economies exist for creatives (esp writers and artists).
- trention 4y ago
  Brynjolfsson and McAfee made a simular argument a decade ago - we will use automation to replace rather than augment human labor, the result being a local optimum far from the global one (that's in terms of results, the dystopian disempowerment of humanity is another story.
LampDrewNear 4y ago
This is how I feel as well. I'm a software engineer by trade, but I'm really much more interested in writing music and stories. I always imagined myself retiring the day I have saved up enough money so I could focus on art and creativity. My life certainly has felt much less meaningful since I saw what DALL-E 2 can produce. And my main concern is not that of professional artists' financial situations. I'm mainly worried about how the massive influx of computer-generated content will inflate away the meaning of human-created art. And I'm afraid that when art becomes so good and so customized to each individual consumer, we won't have much common culture left. I think this Penny Arcade strip well puts into words how I feel: https://www.penny-arcade.com/comic/2022/06/06/dalliance https://www.penny-arcade.com/comic/2022/06/06/dalliance In a way, I'm so mad at my fellow engineers. There's so much good we could do, is really killing the creative arts the right thing to do? And I don't believe for one second this talk about AI just being a tool, and prompt-engineering being a new craft for artists to learn. I think this kind of interface will quickly go away, and soon enough AI apps will look much more similar to TikTok or YouTube than they do today.

fleddr 4y ago

It seems the more interesting legal question is not regarding input, but whom owns the copyrights of the output, if any. Interesting article: https://www.technollama.co.uk/dall%C2%B7e-goes-commercial-but-what-about-copyright https://www.technollama.co.uk/dall%C2%B7e-goes-commercial-bu... If I interpret that article correctly, at least Dall-e does not claim copyright on the generation nor does it grant it to the person writing the prompt. The copyright question seems entirely unanswered. Perhaps that should be codified in law explicitly: AI generated art cannot be copyrighted. If it was created from the collective input of humanity (unasked), its output may as well benefit all of humanity as well. It's hard to make the case that somebody writing a prompt deserves full copyright protection. Even if so, just add one word and one can claim it's once again an "original". Not respecting AI copyrights would also help protect the vast area of commercial art, as it lowers its value if it cannot be legally protected. This could get hairy though, because what happens if you take AI output as input and then start manually tweaking/editing it. It would then be a mixed AI/human works.

archontes 4y ago
It's not entirely unanswered. https://www.smithsonianmag.com/smart-news/us-copyright-office-rules-ai-art-cant-be-copyrighted-180979808/ https://www.smithsonianmag.com/smart-news/us-copyright-offic...
- anonymousab 4y ago
  I wonder if that applies to AI generated code as well.
- fleddr 4y ago
  Thanks, did not know that. I consider it to be good news.
- Pulcinella 4y ago
  This just says that the AI can’t own the copyright, just like how Photoshop, the software program, can’t be granted a copyright but a person using Photoshop can.

donkarma 4y ago

only thing that should worry people is the fact that it is cheap and bad much like machine translation which would make people rely on it, downgrading the baseline of art for companies forever

dahart 4y ago

> Assuming that the input phase is fine and the datasets used are legitimate, then most infringement lawsuits may end up taking place in the output phase. And it is here that I do not think that there will be substantive reproduction to warrant copyright infringement. This might be true, but feels like maybe the most practical answer to this whole question is to focus on the inputs and to establish ways for artists to control whether their art is used as input. Don’t let legally questionable content into training, and you automatically prevent getting legally questionable output. Why would we do anything else, really? It’s potentially a large assumption to make that the datasets are currently legitimate under the lens of copyright. Pulling images off the internet just because they’re accessible is not valid, and humans aren’t currently allowed to do that under the law. But it might be comforting to artists and practical for both artists and AI researchers if artists could specify a form of copyright or a license that allows ML training provided that their work is not recognizably reproduced. Certainly datasets like Flickr with Creative Commons search tools allows a wide variety of public use that is entirely legal. What reasons do we have to not establish something similar for machine learning?

kmeisthax 4y ago

> It looks like it, but it’s not the exact same thing, it’s almost as if the AI is drawing it from memory, which in some way it is, it’s re-constructing what Starry Night looks like. The blog post also has a picture of four slightly-different recreations of a Van Gogh painting. The thing is, under current copyright law in the US, all four of those recreations would be legally considered the same image. The standard for copyright infringement is access plus substantial similarity. It would be easy to construe access to a training image set through ML software that was trained on that set; and thus any regurgitated output would be infringing.

janef0421 4y ago

I don't know if "copyright infringement" is really a useful legal construct here. The issue is not so much that a person's work is being duplicated, but that a person's work is being used to reduce the value of their labour without consent or compensation. It's probably more under unfair competition. Ultimately, though, this issue probably requires new constructs and laws, which is unlikely to happen. The law is the ultimate expression of the power of the dominant class, and these sorts of AI/ML systems benefit the socially dominant class of business owners and investors.

mysterydip 4y ago

Isn't this the other half of the equation for "I host all my pictures/art on site xyz because it's free" and somewhere buried in the accepted terms of use was "you give us permission to use it for anything we want"?

antioppressor 4y ago

Yeah, but what happens if you explicitly state that your art/images must not be included in AI training?

jacobsievers 4y ago
If I've legitimately obtained a copy of your work, then I'm entitled to use it, read it, even make my own copies for personal use. Any new work I attempt to distribute would be measured against your original to determine infringement.
- antioppressor 4y ago
  So if you use part of my image, which has watermark embedded via noise and it is found in your image, then you are fucked. Good.

kazinator 4y ago

> I prompted DALL·E for “Crumbling ruined city, sci-fi digital art by Simon Stålenhag”, and it produced a few images that were very much in the style of Stålenhag. The image shown is nothing has some concepts in common with Stålenhag, but it has a lot more depth. I can't find any Stålenhag which evokes the same way. In Stålenhag's work, or what of it is easily accessible online, there are some recurring gimmicks, which are absent in the DALL-E work, or present in a disguised form. DALL-E doesn't juxtapose any mint condition 1970's Volvos or Saabs against a background showing a menacing superstructure from the distant future. It doesn't invoke any cheesy mechanical monsters or characters inspired by popular Sci-Fi. For instance, " Stålenhag lays out everything plainly. In about a millisecond, you see what you're supposed to see: the easily recognizable elements like ordinary people or retro objects, against fantasy objects, set in some nice lighting: atmospheric effects like haze, sunset or whatever that completes the mood. In the DALL-E there are some old objects, but they are not instantly seen, and they are in old condition. For instance what looks like the wreck of an airliner appears in the rubble in the foreground, but you might not see it right away. That makes the scene plausible. It is more of a spooky glimpse at something that could actually someday be, somewhere. It has the DALL-E weird elements in it. Like why do the cables on what looks like a telephone pole seem to connect to elements in the distant background; DALL-E seems to have perpetrated an https://en.wikipedia.org/wiki/Impossible_object https://en.wikipedia.org/wiki/Impossible_object there, but it's not wham-in-your-face; not everyone will spot it right away. There is what looks like a channel of water through the debris, but what looks like a reflection in it is so weirdly wrong that you can't make up your mind whether it's a weird local reflection or a glimpse through an aperture in the bottom of a floating sky structure, where a distant reflection is shining up from some ocean below. Stålenhag's work doesn't have any of these sorts of demented optical labyrinths in store for the viewer. He's like a Norman Rockwel*, targeting the nostalgia of that viewer who grew up on a diet of 70's, 80's and 90's anime and sci-fi, and is longing for the 20th century. -- * E.g. check out the cheesy fromage called "The Long Way Home". Rockwell plus Storm Troopers. This just screams "Hey, you look transparently like a Generation X geek; would you be interested in buying a limited edition print of me?"

andybak 4y ago

Does anyone have any thoughts on how this might relate to the Marvin Gaye estate vs Robin Thicke and Pharrell Williams "Blurred Lines" case? I know music has very different copyright rules to visual images but that case springs to mind as it was a situation where infringement was judged to have occurred on much looser grounds than previously required. Could a similar expansive judgement happen in the case of AI art and shift the goalposts so to speak?