Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Orochi]

Alright so with the controversy about this stuff going on, I think it would be nice to get some official policy on:

1. Submitting AI-Generated content to the forums
2. The use of vanilla assets or other content to train models
3. The use of AI-Generated content resulting from said models in mods

Considering there are already a couple of posts with this stuff on the forum and how popular kit-bashing is, imo that points to it being approved of, but getting it spelled out in a more centralized and visible topic (maybe in FAQ?) would be nice.

However there are a few more controversial policies related to mod content. Specifically on:

1. Submittinf AI-Generated content not trained on vanilla assets
2. Training models on forum content such as sprite repositories
3. The use of mod content to train models
3. The use of any of the resulting content from said models in mods

This is more controversial, as a lot of content creators consider the use of their art to train models plagiarism. Hell, there are quite a few that consider just about any use of AI-Generated content at all unethical in some way. Yet using mod content to train an AI model means being able to compile a more robust model. It's the difference between a generic model based on "starsector ships" and being able to specifically create models for low-tech, midline, and high-tech ships. Or maybe even models dedicated to nothing but low-tech cruisers.

It also drastically lowers the barrier to create add on content for mods with distinctive art styles. Someone could say, create a model based on the ORA and double the ship count in around a week at most- and the required amount of time will only go down.

So it would be nice to get some official declarations of what is and is not allowed. Also, feel free to speak your mind on these issues. Any potential issues, favored solutions, questions I missed, so on.

[BaBosa]

I personally think it is fine as most of what the AI will make will still need some work to be actually good and as for using other people's work to train the model, all humans base what they make on what they've seen which is basically the same.
Though if what you get happens to look a lot like someone else's work specifically then ask them permission just like you would if you were making similar art from scratch.

[David]

....as for using other people's work to train the model, all humans base what they make on what they've seen which is basically the same.

With the intent of civil discussion which is nonetheless in direct contradiction with your position: it is in no way the same.

The unsaid assumption behind your statement is that an algorithm is replicating the process of consciousness and therefore due the rights of such a conscious, creative entity. None of this is proven or even plausible.

To expand: the theoretical basis for procedural image generation does not suggest that consciousness is possible within its scope, nor do the creators of such algorithms claim that it's what they're doing (except in the most romantic, aspirational sense of a long-term goal which has [I would argue] little to nothing to do with the computer science they're performing), nor has it even been demonstrated, except in the most facile and technically incorrect sense. Therefore the same legal and ethical standards which apply to the work of human artists cannot be applied to the output of these algorithms when judged according to similar rationale.


That said, we'll talk over the issues raised by the OP and come up with an answer.

[Grievous69]

You say it isn't the same, and then never explained why. I reread the rest of the post three times and not sure what is being said there, English isn't my first language but from my long time experience with it, it reads like mumbo jumbo to me.

What does consciousness have to do with anything on this topic? Just because a human doesn't work off of an algorithm, one can be "massively inspired" by something and it's a-okay. But AI uses a tiny fraction of someone's work, mixing things into a cocktail and suddenly that's wrong.

Let me be clear, I completely understand the legal and ethical problems that arrived with it. The problem I have is weak arguments from people being mad for the sake of being mad.

Ok some artists will lose their jobs, so what? Tons of jobs have became obsolete with the technological advances. Are we obligated to stop researching and using new tech because one profession is more important than the other one? Just as old jobs became not needed, new ones could be made available. The problem is laws being slow to define so there are a bunch of grey areas, and people abuse that. But that happens with everything.

I'm glad this post was made since it's smart to set the rules sooner rather than later. To avoid any hypothetical retroactive removal of mods that were then ok, but then became off limits.

[David]

You say it isn't the same, and then never explained why.

I did, however: I said the assumptions behind a statement were false because they rely on propositions of fact which are undesigned, unproven, and unintended.

I reread the rest of the post three times and not sure what is being said there, English isn't my first language but from my long time experience with it, it reads like mumbo jumbo to me.

I apologize; the language has to be very precise because, in my opinion, the core of the current wave of AI hype is based on sloppy use of language which inserts unproven and wildly irresponsible conclusions into the discussion.

Here's the statement I believe is incorrect: "AI art using image input then outputting other images" is the same as "an artist looking at images and creating work inspired by them".

My reasoning:

• an algorithm is not the same as a human artist; it literally has no brain, no mind, no consciousness, and no creativity
• when an algorithm "makes new work based on what it's seen", the underlying process is not the same as a human because it cannot "see", it cannot "learn", it cannot "remember", it cannot "be creative", etc.
• The result of these differences means that an image created by an algorithm and a human artist are legally and ethically distinct.

What I have seen is a lot of use of language which conflates - treats as the same - entities and processes that are fundamentally different. This is done in a way that precludes real and important questions about the results of those differences.

For example, does a human "see" in the same way as a camera? Yes and no.
Does a human learn through seeing in the same way a camera records images? Yes and no.
Can a human reproduce an image previously seen the same way a camera records images? Yes and no.

You can answer "yes" to these questions, but your answer is imprecise, and does not account for distinctions that may be important when given specific situations. To answer "yes" and use that answer as the basis for statements used to cover specific situations in which precise details become important is to make a statement based on false premises.

What does consciousness have to do with anything on this topic? Just because a human doesn't work off of an algorithm, one can be "massively inspired" by something and it's a-okay. But AI uses a tiny fraction of someone's work, mixing things into a cocktail and suddenly that's wrong.

I hope the preceding paragraph explains why consciousness is important.

To your comment: I didn't say anything was "wrong"; I said they were not the same. Human beings and machines have different rights. The output derived from a human being's mind versus the output of data collected by a machine are treated differently in both ethics and law. The laws are, admittedly, pretty bad and written by people who don't understand them half the time. Nonetheless, this is a real distinction.

To return to the argument I was making: Using the process of a human artist's work as an argument/example for how to treat the process of an image generating algorithm is an ethical/legal statement not based on correct assumptions.

To use a metaphor: A bird and an airplane both fly. But birds and airplanes are different.

• They use different processes to fly, which have very different inputs and outputs (bird seed vs. jet fuel) with very different effects on the world
• They originate from entirely different contexts and processes (birds are wild animals, jets are created machines owned by someone)
• Therefore, they should not, and are not treated the same ethically or legally, eg. government regulation of air travel vs. wildlife preserves

My proposal is that images created by artists are different from those created by algorithms using large datasets. The fact that both produce images is true, but it elides context which positions each image-producer into a huge number of distinct categories (ethically, legally, socially, artistically, etc).

For the 'pro AI art' crowd, I think a more productive argument - and one based in both fact and precedent - is to argue for image algorithms as a tool, like a camera. A camera does not itself commit acts of creativity, but it does record. The image-recordings produced by a camera can be changed creatively via human artistic input. Cameras and photographs also have a distinct ethical and legal position within our society - there are rules that apply to photographs that do not apply to images created solely by human effort.

Ok some artists will lose their jobs, so what? ...

You're introducing an argument speculating about ethical outcomes. I think it's better to stick to unpacking the assumptions of proposals through use of reference to to the existing status quo because it doesn't result in talk about ideal political outcomes (which tend to get unproductively contentious).

[Alex]

You say it isn't the same, and then never explained why.

I think he did, very clearly. Maybe the term "AI" is making it confusing; I believe that's actually the intent behind its use for these algorithms even though they have just about nothing to do with intelligence. It's basically automated kitbashing on a larger scale, somewhat better covered up. (*Usually; theses algorithms can also occasionally just reproduce some of the input.) It's not doing anything qualitatively different. (Edit: David covered this in a lot more detail, but, leaving this in.)



We've talked about this a bit, and think "same as kitbashing" a good provisional basis for treating these models. The law has obviously not caught up with this yet and who knows what this'll look like when it does, but for the moment, the forum rules/stance on this will be "treat this exactly as kitbashing". So for example:

1) You can train on vanilla assets for the purpose of making Starsector mods (anything beyond that requires explicit permission, same as e.g. using Starsector assets for mods for another game)
2) You need permission from the copyright owners to use mod assets or *anything else*, so e.g. you can't use the output of a model trained on random internet data or some such**

And so on. Basically, just replace "used an AI image generator" with "kitbashed using these images" and go from there, copyright-wise.

For clarity: "kitbashing" is the practice of taking pieces of sprites and making new sprites using these and some original work to make the pieces fit together.

**Some of the mods posted now may be in violation of that? I'm honestly not sure.  If this is the case, the mod authors will need to address that.

[Grievous69]

Thanks for explaining in detail David, I have understood your point much much better now. I think the topic is just convoluted enough where the language barrier might cause some confusions if not carefully described. And also specific word usage may either help or make things ambiguous.

Analogies (like yours) are bloody great since everyone can read and comprehend them in the intended way. Sorry to you both for making you write such long posts but I bet it'll help others as well.

[Void Ganymede]

Fitting for a game with AI cores!

"AI" is a crap term. It's worth breaking the tech down into actual categories:

Diffusion Models: common case and a no-brainer. All the agency is in the human. Most of the creative process (prompting what concepts to interpolate and how) is in the human. Fine-tuning some variant of Stable Diffusion using LoRA, on several hundred example images, seems a very close equivalent to kitbashing those images. Another close analogy is 3D modeling then rendered down into sprites, or making ships with vector graphics - it doesn't match Starsector's style and is arguably a crutch to compensate for a lack of painting skills by using other skills.

If someone throws a pen at you, you blame them and don't fistfight the pen. If someone generates SD environmental shots and portraits they're responsible for curation choices made.

Large Language Models: get weird. Language capabilities are fancy enough the user's intent is diluted. Is something like AI Dungeon the result of the user's agency driving it down particular paths, or a labyrinth all its own? Top of the line models like Bing and GPT4 can do storytelling with their own distinct character and prompting fails to replace that character with the user's own.

examples

Where it gets interesting is LLMs can write prompts for themselves and for SD. Curation of those prompts is still in the hands of the user, but if you ask Bing to write a story about a space pirate from Volta then generate a prompt for a profile picture, using Starsector writing and rough style descriptions for context as part of the prompt... Well, Bing's too much a Goody Two-Shoes to really get into character. So inside the pirate there'll be a Lawful Good soul trying to get out even though that's not in the user's prompt and probably not what they're curating for.

As another example, Bing and GPT4's RLHF tunings resonate with certain character archetypes. A really weird one is Judge Holden from Blood Meridian. You ask GPT4 to write a speech as the Judge, and it gets into character so hard it has trouble getting out. Based on that, with the right prompting, there's some great potential for a force-of-nature type character that'd slot well into the Sector. Is the result ripping off Cormac McCarthy, doing transformative fanfic, or a novel work with obvious inspirations? How much of the credit goes to human agency of the prompter/curator, and how much to GPT4's downright uncanny death-god impression?

[close]

There's a common counter-argument that LLMs have zero agency or qualia - they are machines, we are divine beings! It's not strictly wrong, but the character simulacra they roleplay act as if they do - and if you treat them as if they do, the simulation gets more effective.

Multi-Modal Models: No public access yet, but they're coming. While the improvements from GPT3.5 to GPT4 aren't really a step change but rather a gradual curve, they certainly *feel* like a step change to most users. Low parameter count free models will be easy to notice, but if a model is advanced enough that the only distinction from human work is suspiciously high quality, stylistic tells, and a certain lack of global coherence for tasks too large to fit into a context window...

Personally I'm excited to see what kind of fanart or fanfiction GPT5 can produce for some of my favorite imaginary worlds.

[Alex]

There's a common counter-argument that LLMs have zero agency or qualia - they are machines, we are divine beings! It's not strictly wrong, but the character simulacra they roleplay act as if they do - and if you treat them as if they do, the simulation gets more effective.

Just a comment on this - I've made some very basic Markov chain name generators for Starsector - for planets etc. And with an absolutely trivial depth - like, "what letter is most likely given the previous two" - it produces *some* things that are extremely eerie. And it's dumb as a box of rocks; it's a trivial lookup in a few arrays and not much else. GPT/whatever is certainly much more involved and effective, but fundamentally the same thing; any feelings it produces in us that there's "someone there" is all in our heads.

And, yeah, AI sure is a crap term! Or rather - it's entirely a marketing-oriented, non-technical term.

[Void Ganymede]

Right, that's why the argument's technically not wrong - just unconvincing to people who view sufficiently advanced technology as magic. All that wetware full of pattern recognition that sees faces in clouds, tells stories about weather, gives pets more credit and inanimate objects more blame than either deserves...

"LLM storyteller is more likely to tell good stories given the prompt treats it with the encouragement and respect befitting a skilled storyteller" has a simple probabilistic interpretation, but most people will see the anthropomorphic one.

[BaBosa]

Spoiler

....as for using other people's work to train the model, all humans base what they make on what they've seen which is basically the same.

With the intent of civil discussion which is nonetheless in direct contradiction with your position: it is in no way the same.

The unsaid assumption behind your statement is that an algorithm is replicating the process of consciousness and therefore due the rights of such a conscious, creative entity. None of this is proven or even plausible.

To expand: the theoretical basis for procedural image generation does not suggest that consciousness is possible within its scope, nor do the creators of such algorithms claim that it's what they're doing (except in the most romantic, aspirational sense of a long-term goal which has [I would argue] little to nothing to do with the computer science they're performing), nor has it even been demonstrated, except in the most facile and technically incorrect sense. Therefore the same legal and ethical standards which apply to the work of human artists cannot be applied to the output of these algorithms when judged according to similar rationale.

That said, we'll talk over the issues raised by the OP and come up with an answer.

You say it isn't the same, and then never explained why.

I did, however: I said the assumptions behind a statement were false because they rely on propositions of fact which are undesigned, unproven, and unintended.

I reread the rest of the post three times and not sure what is being said there, English isn't my first language but from my long time experience with it, it reads like mumbo jumbo to me.

I apologize; the language has to be very precise because, in my opinion, the core of the current wave of AI hype is based on sloppy use of language which inserts unproven and wildly irresponsible conclusions into the discussion.

Here's the statement I believe is incorrect: "AI art using image input then outputting other images" is the same as "an artist looking at images and creating work inspired by them".

My reasoning:

• an algorithm is not the same as a human artist; it literally has no brain, no mind, no consciousness, and no creativity
• when an algorithm "makes new work based on what it's seen", the underlying process is not the same as a human because it cannot "see", it cannot "learn", it cannot "remember", it cannot "be creative", etc.
• The result of these differences means that an image created by an algorithm and a human artist are legally and ethically distinct.

What I have seen is a lot of use of language which conflates - treats as the same - entities and processes that are fundamentally different. This is done in a way that precludes real and important questions about the results of those differences.

For example, does a human "see" in the same way as a camera? Yes and no.
Does a human learn through seeing in the same way a camera records images? Yes and no.
Can a human reproduce an image previously seen the same way a camera records images? Yes and no.

You can answer "yes" to these questions, but your answer is imprecise, and does not account for distinctions that may be important when given specific situations. To answer "yes" and use that answer as the basis for statements used to cover specific situations in which precise details become important is to make a statement based on false premises.

What does consciousness have to do with anything on this topic? Just because a human doesn't work off of an algorithm, one can be "massively inspired" by something and it's a-okay. But AI uses a tiny fraction of someone's work, mixing things into a cocktail and suddenly that's wrong.

I hope the preceding paragraph explains why consciousness is important.

To your comment: I didn't say anything was "wrong"; I said they were not the same. Human beings and machines have different rights. The output derived from a human being's mind versus the output of data collected by a machine are treated differently in both ethics and law. The laws are, admittedly, pretty bad and written by people who don't understand them half the time. Nonetheless, this is a real distinction.

To return to the argument I was making: Using the process of a human artist's work as an argument/example for how to treat the process of an image generating algorithm is an ethical/legal statement not based on correct assumptions.

To use a metaphor: A bird and an airplane both fly. But birds and airplanes are different.

• They use different processes to fly, which have very different inputs and outputs (bird seed vs. jet fuel) with very different effects on the world
• They originate from entirely different contexts and processes (birds are wild animals, jets are created machines owned by someone)
• Therefore, they should not, and are not treated the same ethically or legally, eg. government regulation of air travel vs. wildlife preserves

My proposal is that images created by artists are different from those created by algorithms using large datasets. The fact that both produce images is true, but it elides context which positions each image-producer into a huge number of distinct categories (ethically, legally, socially, artistically, etc).

For the 'pro AI art' crowd, I think a more productive argument - and one based in both fact and precedent - is to argue for image algorithms as a tool, like a camera. A camera does not itself commit acts of creativity, but it does record. The image-recordings produced by a camera can be changed creatively via human artistic input. Cameras and photographs also have a distinct ethical and legal position within our society - there are rules that apply to photographs that do not apply to images created solely by human effort.

Ok some artists will lose their jobs, so what? ...

You're introducing an argument speculating about ethical outcomes. I think it's better to stick to unpacking the assumptions of proposals through use of reference to to the existing status quo because it doesn't result in talk about ideal political outcomes (which tend to get unproductively contentious).

[close]

Fair enough. I personally don't like making decisions based on the specialness of conscience and rights because they're quite subjective and impossible to get a consensus on as everyone has different opinions.

My two cents is that these models are based on our brains and unlike your bird and plane example, diffusion models don't take in any inputs that humans don't though humans take in other stuff as well. That said I am not arguing that they should be treated as humans, they are still tools. Just because they mimic the brain instead of arms, ears, eyes, etc doesn't change anything there.

While I understand requiring permission to train models on other mod makers' assets, why the ban on using open domain images from elsewhere on the internet?

[Doctorhealsgood]

So... gamma cores when? I want one.

[Buggie]

While I understand requiring permission to train models on other mod makers' assets, why the ban on using open domain images from elsewhere on the internet?

How would they know if the person in question genned an image using stuff from the public domain or not? I think its just one of those cases where its better to be safe than sorry.

[Alex]

My two cents is that these models are based on our brains

(I can't overstate this enough: they are not, and it's not even close.

Based on what you're saying, you might be seeing these models as something fairly nebulous that "learns" or whatever? It's not, it doesn't, and the idea of granting it rights is equivalent to granting rights to a "hello world" program. It's just a bunch of code running a well-defined algorithm and producing a bunch of numbers. These numbers - a few gigs of them iirc (edit: for a larger model? I could be way off here, though) - are then used to produce the image output based on prompts etc in, again, a well-defined multi-step process; and a fair bit of that data is some alternate representation of some form of the training data.

And, side point: it takes the images in as rgb values, nothing like a human would. Which is, incidentally, part of why adversarial-image attacks are possible... but anyway. That's a really minor point given the nature of what it is. Btw, I believe the point of calling it "AI" and the marketing around that is to make it unclear what this actually is and create exactly this sort of confusion about it. It *is* a powerful tool, though.)

While I understand requiring permission to train models on other mod makers' assets, why the ban on using open domain images from elsewhere on the internet?

If it's open domain/the license allows it, then that would not be banned. The point is that a lot of these models seem to be trained on data that the people training it do not have rights to, and that's what would not be allowed.

[BaBosa]

My two cents is that these models are based on our brains

(I can't overstate this enough: they are not, and it's not even close.

Based on what you're saying, you might be seeing these models as something fairly nebulous that "learns" or whatever? It's not, it doesn't, and the idea of granting it rights is equivalent to granting rights to a "hello world" program. It's just a bunch of code running a well-defined algorithm and producing a bunch of numbers. These numbers - a few gigs of them iirc (edit: for a larger model? I could be way off here, though) - are then used to produce the image output based on prompts etc in, again, a well-defined multi-step process; and a fair bit of that data is some alternate representation of some form of the training data.

And, side point: it takes the images in as rgb values, nothing like a human would. Which is, incidentally, part of why adversarial-image attacks are possible... but anyway. That's a really minor point given the nature of what it is. Btw, I believe the point of calling it "AI" and the marketing around that is to make it unclear what this actually is and create exactly this sort of confusion about it. It *is* a powerful tool, though.)

While I understand requiring permission to train models on other mod makers' assets, why the ban on using open domain images from elsewhere on the internet?

If it's open domain/the license allows it, then that would not be banned. The point is that a lot of these models seem to be trained on data that the people training it do not have rights to, and that's what would not be allowed.

I did specifically mention that I am not arguing that they should be treated like humans or be given rights or anything else. Just because it can perform tasks that humans can do (make art) does not make them people, in my opinion at least. It is still a tool even if what it is doing are things some other people might consider fundamentally human.

The models are based on neural networks right? Which are based on the neurons in our brains? I'll admit I am not an expert in the slightest so I am genuinely curious.

Yea I definitely agree that once you actually start talking about specifics, calling it AI is uselessly vague. Which is why I'm mostly avoiding using that word.

Ah, so just no scrapping the whole internet sort of thing. Picking out allowed stuff is algood.