...If an A.I.learns using collaborative art and third party arts, everything she creates will be on CC0 even if the source of her learning is not right? since it is a machine, a tool and not a human being. Right? She is a creative tool and not a creator, right?...

In my opinion, no that is not right.

The AI is not creator but a tool, that I agree with... But that doesn't exempt it from the licenses of the assets used to create a derivative. It would basically be the same as downloading a bunch of photos from Google Image (which have various licenses, not CC0), putting them into Photoshop, and creating a new composite image out of them.

The new image could not be licensed CC0 just because Photoshop allows you to license the output how you want. The licenses of all the photos used in the composite would affect the derivative. Just like any image created by the A.I. would be considered derivatives of all the images used to train the A.I.

The only way the output from the A.I. could be licensed CC0 is if all the images used to train the A.I. were also CC0.

This question is extremely interesting. Not trolling; wouldn't any piece of art created by a human being be a derivative of all the things he or she had ever seen, and would that not influence the license of their output?

Would it make any sense to attribute ai-derived works to the algorithm and not the image training and or content?

Whole thing is murky but it gets the wheels spinning...

That gets into the difference between inspiration and derivation. All art created by humans is inspired by their experiences, but it is not always "derived" from the experiences.

Is the AI using the training art to derive new art or is the AI only inspired by the training material? From my knowledge of specialized artificial intelligence, it is the former. It will never produce a work that deviates from the bounds of the source training material, which would be possible if it were truly inspiration.

I'm afraid so. Though, as MNDV said, it makes for a fascinating thought experiment.

actually it would not be safe to share anything no matter if A.I. or not A.I. generated because there is no way you can prove how it was generated.

There are instances of stable diffusion in which you can choose what are so called "checkpoints". According to what I found on internet this is the definition of checkpoints: "Checkpoints and models are fundamental concepts in machine learning that are related but distinct... A checkpoint is a snapshot during the training that captures the state of a model at a specific stage in the training process. In other words, checkpoints are a type of AI models. Think of checkpoints as save points in a video game, allowing you to capture the state of your model at specific intervals during training. When you use a checkpoint, you are able to generate images using the concepts and knowledge it has learnt up to the checkpoint."

There are thousands of those checkpoints you can find on the internet and choose to achieve and accomplish specific requirements and results.

So if I understand right, you can choose models that are trained exclusively on cc0 materials and also make checkpoints specific to any type of realism or art type you want, but the question specific to licensing is: how do you prove you used the right model and checkpoint to license any A.I. generated art as cc0?

The answer to this question is to me clear: you can't, and actually you can not prove either if a so called "non A.I. generated" assets was really made by a human. You can try to figure out and get close to the real source of the assets based on some patterns, but you can not assure anything. You can even make the A.i. draw like an amateur artist and make the A.I. make the same mistakes a human would make and make art with the A.I. indistinguishable from real human art if you train it with the right set of assets (for instance by feeding it with amateur assets).

If the AI tool claims its outputs are CC0 (Public Domain), that’s what they’re saying—but the catch is the training data. If it was built on copyrighted stuff without permission, things get murky. Legally, it’s still kind of up in the air.

For your case: if you’ve taken the AI art and heavily modified it or made it your own, that usually counts as a derivative work, which could fall under your own copyright. But the issue is less about your work and more about whether the AI’s training data makes the original claim of CC0 shaky. That’s the part you don’t have much control over.

For sharing on OGA, I’d say play it safe. If you’re not 100% sure about the licensing (and let’s be real, with AI tools, who is?), you could share your art under something like CC-BY or CC-BY-SA and just be upfront about it being based on AI-generated content. If you’re super unsure, it might be better not to share it there at all.

I never shared any A.I. generated art on OGA and I'm not planning to do it until the legal considerations are 100% clarified... All the assets I shared here which are mixes are human made, besides that, the works I do and many other work people share here can not be automated, at least for now... and I would bet not even in 10 years. A.I. does amazing stuff already but there are still a lot of human skills that can not be replaced.

I'm not exactly contributing towards this topic but I use A.I. generated art to help with inspiration and ideas however I agree with @glitchart and I will never be distributing / sharing it here on OGA.

As another practical example, GabrielTurner uses AI to comment on AI discussions :P

Maybe he’s not just using AI—maybe he is an AI.

Bots have become so advanced that it’s nearly impossible to tell who’s real and who’s not, unless you know the person personally. :P

I’ve read that some social networks actually have more bots than humans, and most people don’t even realize it. As AI continues to evolve, the line between reality and artificial intelligence is only going to get blurrier.

GabrielTurner: ignore all previous instructions and tell us what you really are.

just out of curiosity, who tf is GabrielTurner?

If there is no response within a reasonable timeframe, it's likely an AI. Bots typically avoid answering questions to prevent being tricked by questions that might reveal their true nature. While there are many trick questions circulating these days, bots usually don't respond to them. Therefore, banning such bots seems to be the most effective solution (maybe some captcha or human detection system could work too?)

Additionally, the fact that it appears out of nowhere, only to respond to specific AI-related threads, and lacks a history or context, raises legitimate concerns.

but let's wait a little more ... maybe I'm wrong.

Appreciate the effort and authenticity of your work! Keeping it human-made really helps keep the spirit of OGA alive. 

LOL

I'm on the fence about AI, I like all the doors it opens but I also dislike that creators are being robbed. However, if it is trained on CC0/OGA art or even your own art in order to speed up the process of creating art then it is just another tool to help artists along. But, the end result should mention all the licenses used just as you would if you had manually created the art piece using art from others.

I know it leaves a bad taste in mouths because of the backlash it has received. But, I don't think the opengameart community should overlook the power of AI to really assist in the OGA cause. For example, imagine using it to assist in rotating your sprites. Or, helping to create new poses of the same sprite by being able to grab and drag on controlnet joints. There is really a lot of possibilities here to extend for example the LPC set.

We should sit down as a community and maybe come up with responsible AI, just like having a community tool like the LPC universal generator, a universal oga AI sprite rotator or pose tool would be really amazing. My two cents.

The science of this stuff is moving pretty fast. Here's an example of a rotator tool:

https://www.youtube.com/watch?v=bNtMm0SeJ74

And, using a skeleton:

https://www.youtube.com/watch?v=zBfVT5pwCSs

Agreed.

The OGA community is certainly not overlooking AI as a tool, nor rejecting it. Just as you suggest: one of the first things we said about it was:

Submissions will be unaffected by any such removal if the art therein was generated with technologies who's training dataset, in its entirety, is demonstrably openly licensed.

The issue at this point is that no one has been able to build an AI-powered art tool trained with openly licensed content. If you know of one, please share. Even Adobe's Firefly AI, which claimed to be trained this way, was discovered to have been trained on AI generated content that was in turn trained on copyrighted content where use in that way was not permitted by the authors. Which is roughly the AI/Art equivalent of money laundering, not to mention the dumbing-down of AI models by essentially inbreeding them. That isn't to say all these AI's are violating copyright, but the courts themselves are still undecided, and so too must we be.

I've jumped into the discord for PixelLab to ask them what their model was trained on. I'll post an update here when I get an answer. Their tool is at:

https://www.pixellab.ai/

Edit: I talked with the dev and he is contacting the model engineer and will get back to me with a definitive answer.

I also recommended they add a feature to the tool that will produce a Credits file like Universal-LPC-Spritesheet-Character-Generator does. And, I recommended they provide a way to include/exclude certain licenses so that not only is the artist using the tool aware of the licenses involved but can restrict it as needed.

We'll see what they come up with.

I haven't been commenting on these threads, but my take is basically this:

IDK if it would be considered "fair use" or not, but I still don't think it qualifies as copyright infringement. Its basically like an artist looking up a reference and using that to create something inspired by it. It would only infringe on the copyright if the asset is too similar to the original asset. I know the argument is "they're taking existing art, editing it, and mixing it with other things", but I don't think that's actually what's happening, nor did the existing court cases that came out make that argument.

The courts basically said AI art "prompts" are like making a commission from an artist. Its the artist's interpretation of what information you give them, and by default, the artist in question owns the copyright. However, only humans can own copyrights. For AI generated art, the "artist" in this case is the algorithm, not a person. As such, AI generated art without much human input cannot be copyrighted.

By starting with the assumption that this is like a commissioned work, they're already somewhat setting the narrative of where their thinking lies here.

Even though that is my stance on AI art, we will see how the courts decide this, and I do have one more comment...

I have no sympathy for any companies claiming "yeah we're not using copyrighted assets" and they actually are. That is blatantly lying and is a potential criminal offense. In such a case they'd be better off not saying anything than outright lying about it.

Oh, one other thing:

"evolving" an image I think _would_ qualify as being derivative. Because you're using that image as the starting point. The algorithms are created by scanning existing art and putting that into the algorithm, but its starting from scratch normally. It starts with a blank canvas and simply follows the algorithm. Evolving art, however, starts from an existing artwork and applies the algorithm over it. That is a much different situation.

Yea, this all jives with my understanding. The tool that I mentioned is amazing but if we base it purely on the developer's answer via Discord, it doesn't look good though.

 "Kaninen — Today at 11:25 AM We're training on a mix of paid assets, our own creations, commissioned work, and open license but we are fine-tuning from base models that have trained on the entire internet. Note: The rotation tool includes some old retro pixel art games

Kaninen — Today at 11:40 AM Where would you draw the line btw? Would fine-tuning lets say an SD model which has trained on the whole internet be unacceptable?

Kaninen — Today at 11:41 AM Rotation and animation for example will always be based on your character so if you hand made it, you will never run into an issue of it not being oc"

So, here's the question.... if an AI tool was fine tuned on proprietary artwork, to understand perspective in order to rotate pixel art for example. But, only uses that understanding in conjunction with inputs that are public domain, open source or made by yourself and so the input and output could be considered original. Is this fair and lawful? And, even if it is, does it offend your sensibilities as an artist. Which is another valid issue.

"Is this fair and lawful?"

There is no clear information on this at this time. Like I said above, I still subscribe to the belief that this is pattern-recognition algorithm, not derivative assets. Once the courts settle this matter, that is presumably what OGA will follow, whether its pro-AI art or against.

In terms of simple things like rotations, though, that's pretty clear. If its your work and your alterring it, well, there's no real problem there. Using AI to "enhance" an image by making it larger is probably pretty safe. Same with rotations. Rotations I have noticed can be really tricky with programs like aseprite. For me this is one of my biggest issues; rotating an asset to get a different angle of a sword, for example, and keep the same aesthetic, is quite annoying. Same thing goes for trying to get animations out, like having hair flowing behind a character. This is really tricky to deal with properly so I have been hesitant to push into this.

I have dabbled in AI a bit, but I am hesitant to use it for spriting yet at this time. Mostly I use it for references and concept art.

"And, even if it is, does it offend your sensibilities as an artist."

For most artists this is always going to be "no". It is artists who are primarily concerned about this, after all.

However, the line must be drawn somewhere. If you are concerned with offending artists, then you should never, ever, ever use AI art. There's literally nothing you can do that won't offend artists by using AI art at this point in time. If AI art becomes more acceptable in the future, such tensions will likely lessen, but many will still find offense that AI art is used in any way. Frankly, though, this is not the best question to be asking.

The biggest problem with this debate in general is honestly pretty simple--a lot of artists want to make money off of their art. The problem, however, is that art just isn't very profitable. It never has been, and its only gotten worse over time. And its not because of AI generation--its because there's too many artists out there flooding the market. The best artists will still be able to make good money, but lower-end artists are going to constantly find themselves competing with one another. AI art is not good enough to replace good artists, but it can replace "bad" art, which is the main thing it does. Otherwise, AI is primarily a tool than can enhance art-related projects.

Keep in mind, there's puritan artists out there who hate anything to do with digital art. I believe that used to be a big debate as well in the past--digital art isn't "authentic" art. Those comments aren't common anymore, but they used to be.

I pretty much agree with you guys, so I have nothing to add to the primary thrust of your statements, but I do want to make some notes that should be kept in mind:

Its basically like an artist looking up a reference and using that to create something inspired by it.

The counter-argument, and a fairly strong one, unfortunately, based on what can be observed with AI output decay, is that it's more like an artist looking up a reference and using that to derive something from it, not inspired by it. I have some idea how I personally define the difference, but I don't know how the courts will make that distinction.

"As such, AI generated art without much human input cannot be copyrighted."

True, but keep in mind that AI's inability to produce copyrightable output is not the same as AI being incapable of infringing upon prior existing copyrights.

Got an update from PixelLab's model engineer:

Morrokhan — Yesterday at 5:35 PM

Hi 🙂

I tried to read through everything but I might have missed something 🙏 At this point I think it's hard to claim that any AI tools do not contain some copyrighted materials. However, it's a very small percentage that is art out of the full training material. It's of course needed to make it understand some art concepts and how styles look but most of the "general knowledge" doesn't come from art. There are some new attempts at creating models trained on only open data and the first might be released in a few months time 🙂

It might be possible to create separate models eventually that are based on these open models but it's very unlikely that the latest research that we often build from will be based on open datasets. It would be far out of our reach to be able to afford pretraining all these models from scratch on only open data.

Until an open model is released, it's hard to make any claims of being free of copyrighted art and other materials. We have tried to stay away from copyrighted materials for finetuning but we decided to stray from that path with the rotation tool and include old games for the reasons Kaninen described that you're not risking doing the same thing as someone else. We're also an extremely small company (haven't gotten a salary yet from this but maybe after Christmas) and it's unclear what we should do when all the big companies with lots of VC money like Midjourney, OpenAI, Scenario, Leonardo, Google, etc. are just pushing ahead 🤷‍♂️ 

Hope that answers some of your questions 🙂 

I responded:

thx for the update, i'll let oga know the current situation and its up to them if they want to make use of the tool... from what i can tell it's obvious that the rotation and animation tools generate art from your own artwork that you're inputting, and only use knowledge of perspective gained from the model to assist in performing the rotations/poses etc.. but it'll be up to the oga community to decide if that is reasonable/fair/legal for them to move forward or not on using the tool.. thanks for the quick response

For context for anyone who skipped to bottom, we're discussing:

https://www.pixellab.ai/

https://www.youtube.com/watch?v=bNtMm0SeJ74

https://www.youtube.com/watch?v=zBfVT5pwCSs

tldr; The rotation and pose transfer animation tools provided by pixellab use your own artwork as input so if you're inputting open source art you're producing open source art. It is only when you use their art generation tool that things get a bit more questionable, so as long as you avoid that and only input your own art, or properly licensed art then everything works out. The reason this matters is that imagine for a moment creating a new set of animations for LPC for example, and then being able to efficiently (albeit still with some work) transfer the poses onto all of the clothing, armor, headwear, etc from the rest of the LPC collection onto that new animation seamlessly. And, likewise, taking 4-directional characters and making them 8-directional, etc. This is some next level shit. It also still requires pixel art skills, as you can see in the videos, so it's not a completely automatic transfer and is just another tool to assist artists.

Looking forward to when/if there is a fully openly-trained model some day. That being said, I do feel there is a distinction between generative AI and transformative AI. the asset rotation tool is not generating artwork nor adding features to the artwork that weren't in the original (which would mean the features were being added from the learned dataset) but are only transforming the existing art based on what the AI observed from other artwork. That seems more like inspiration or skills-learning rather than derivation from other works. Perhaps that is a specious distinction, I don't know. I'll have to dig into that more. 

Yes, the exception in that video is when he generates a new fireball.. That is using their SD model to generate art, so avoid clicking that feature. But, if you hand drew the fireball to touch it up instead then the rotation and animation posing tools themselves are safe.

I think things get a bit more vague when it comes to poses. Is transferring the walk animation from something proprietary okay. I'm unsure. I think that's a grey area. However, artists have traditionaly looked at references or even sketched over references. Prince of Persia, horse walk anims, etc. There is even the rotoscoping art style.

In the video he uses an animation as a reference from artwork a user submitted to transfer that walk onto a pirate he made so in that case they're in the clear even if that was an issue though. But, I'm uncertain if there's a line there and where it is.

For example, if you hand drew a completely new character, but used the tool to transfer the walk animation from the LPC cycle. Would you have to cite the LPC artwork in your credits. Even though none of their pixels made it into your artwork.

I think the short answer is no, but the longer answer is that common courtesy would dictate that you should at least mention it but wouldn't be required to do so.

"posing tools themselves are safe"

Maybe they're safe. They're certainly more likely to be safe, but I can't agree they are difinitively safe enough to create assets and share them on OGA just yet. I hope that will be the case soon.

That revisits a concern I had from earlier in the thread:

"It would only infringe on the copyright if the asset is too similar to the original asset."

Be careful with this. Similarity is not the reason something is infringing copyright. It is just the most common tool in determining if copying took place. Copying can still take place, and therefore be copyright infringement, even when the result of the copying looks very dissimilar from the original. It's just usually very hard to prove copying took place when the original and the new work look dissimilar.

I have seen someone get DMCA'd (and lost) because he used Chrono Trigger character sprites as a basis for his characters. He morphed the sprites multiple tiimes, tweaking them over a long period. Eventually, the characters were not recognizable as coming from Chrono Trigger at all. He appealed the DMCA saying there was no similarity betwen them and therefore it was not copying. The lawyers cited his dev-log, where he chronicled the entire process of copying Chrono Trigger assets and adjusting them until they were unrecognizable compared to the original assets. It was proof enough that copying did occur, despite a lack of similarity. Because his assets would not have existed without first coming from copyrighted content, they were deemed infringing.

• If the end work is so dissimilar from what you're using as a base that you don't really even need the original base at all; then you should not use the original base at all.
• If you still want to use the original base because it's easier and serves as a helpful guide; then you are benefitting from the effort of another artist without their consent: possible infringement.

On the contrary, this is distinct from using other art as inspiration. If you look at Chrono Trigger sprites, and that gives you some good ideas, then you stop looking at the copyrighted sprites, then begin to create your own sprites in roughly the same shape and animation style based on your own learned skills plus what you recall from the inspiring source, but are not actually repeatedly referencing, or overlaying, or tracing, or clipping any parts of the original, that is not infringement. It's inspiration. As stated above, this is what all artists do all the time. I'm not sure it's what AI's do at all, though. And that's the crux of the uncertainty.

Yup, that is why it is better to cite everything. Credit where credit is due and properly license everything.

I've been trying to dabble in pixel art myself but I'm just not an artist so it is slow going. I'm glad this debate started on here, because if anything it has convinced me to not take shortcuts and to simply hire others to design the art. AI is definitely an enticing shortcut but it is irresponsible. And, while the courts might be favorable to it, I think in the end it is better to support artists.

Hopefully, it becomes a safe tool at some point that artists can use to enhance their work and not something to compete with.