Comment by throwaway1851 | Hacker News Reader

Comment by throwaway1851 | Hacker News Reader

Parent Full thread

There are really significant, novel copyright issues implicated by these large generative models trained on other people’s IP.

If you take a step back, you can see that there are different ways to frame what is happening. One frame is: “Defendant built an algorithm that memorized features of Plaintiff’s IP. Defendant’s algorithm recombines parts of those features in order to produce works in the same domain that compete with Plaintiff’s work, all without Plaintiff’s consent.”

Bear in mind that copyright holders are among the most litigious out there. If generative art becomes as big a deal as some people expect, they will have every incentive to use their huge litigation budgets to claim a piece of the action.

Because the law develops very slowly, the legal process has not yet had the occasion to really evaluate what transformative use means in this novel context. I’m personally interested in seeing where things go, but it’s going to be a while before we know where the law is headed.

> “Defendant built an algorithm that memorized features of Plaintiff’s IP. Defendant’s algorithm recombines parts of those features in order to produce works in the same domain that compete with Plaintiff’s work, all without Plaintiff’s consent.”

The fun part is, this is how human artists learn too.

They absolutely do. What they don’t do is mechanistically clone compressed mathematical representations of input data. The human part of the creative process could very well be a distinguishing feature, legally.

Tracing and collage both mechanistically clone elements of the input. Trendfollowing, mimicry, and copying others is standard in any creative area.

I don't think computer-generated works will be easily distinguishable from human unless they're desired to be or shipped with metadata. It's already hard enough to distinguish human artists from other human artists without having names attached up front.

> Tracing and collage both mechanistically clone elements of the input.

Yes, and tracing counts as art fraud.

Collage is a bit different, because you are mixing many clones of many other objects such that you create a new object; additionally the way you assemble the clones may transform them (a photo of the mona lisa has different surface texture than a painted version, even more different if it is clipped from newsprint), but while the borders of this are not clear, it is clear when people are far enough over the border. Think of hip-hop, sampling, and remixing music, and some of the legal battles which have come out of that.

> Yes, and tracing counts as art fraud.

I just spit out my coffee, turns out I'm not a bad artist, I'm actually a fraudster!

They can (and do) clone compressed electrical representations of everything they see. Everything is just stored in memories in their brains instead of memory on the cloud. In the case of these complex AI, you cannot really extract the initial works, right? They have all been atomised, mashed together, and very imperfectly encoded as weights and parameters.

Yeah I am sure a lot of lawyers are going to have a lot of fun arguing every way imaginable.

I see this type of reduction again and again to advocate for one position or another relating to ML. Human consciousness and thought processes aren’t “just” anything (math, electrical impulses, etc.) — the fact is, we don’t know what the brain really does and how it’s connected to our conscious experience, or even what that is!

Deep learning is very powerful and impressive in its applications to date. However, it’s so saturated with hype (and humans are so prone to anthropomorphizing things) that it’s often viewed as something much more profound than it actually is. Neural networks, despite their name, don’t model the brain. And they lack a whole array of “intelligence” features that humans possess and use constantly.

All of this is to say that there are very significant differences between computer algorithms and human cognition, and I tend to think the legal system will be unpersuaded by arguments that ignore those differences.

Also, this is to say nothing of the public policy interests that shape the law. Regardless of what’s “under the hood,” the law can simply treat human and machine output differently. I’m not a copyright lawyer, of course, so I can’t speak to the norms or technicalities of copyright law itself.

> Human consciousness and thought processes aren’t “just” anything (math, electrical impulses, etc.) — the fact is, we don’t know what the brain really does and how it’s connected to our conscious experience, or even what that is!

There are a lot of things we don’t know, but it is not magic. There is no discussion that artificial neurones don’t have much in common with the real ones, which are very non-linear and much more connected. But in the end it’s all electrochemistry.

> Neural networks, despite their name, don’t model the brain.

But that’s not directly related to my point. My point that even in the case of a ML model, you cannot get an exact reproduction any more than you can get from a human’s memory. In one case it’s scrambled somewhere in someone’s brain, in the other on a hard drive but the difference is not really relevant. Subjecting an AI’s production to the copyrights of all the things it’s been exposed to is very similar to subjecting a painter’s production to the copyrights of all the painting they have seen.

> “There are a lot of things we don’t know, but it is not magic. . . . [I]n the end it’s all electrochemistry.”

That’s an element of a belief system you may choose to subscribe to, not a fact. It can’t be a fact because there’s no way to prove or disprove it.

> “My point that even in the case of a ML model, you cannot get an exact reproduction any more than you can get from a human’s memory.”

Not quite. ML models can and do memorize and regurgitate near-exact features of the inputs. It’s not the goal, but it happens.

It will be hard to tell whether an image was created by a “human creative process” or AI.

Will the same image be legal if a human made it but not if it was created by Stable Diffusion? How will someone even know, short of a legal discovery process?

A legal discovery process is a perfectly reasonable (and common) way to determine if something is legal, and yes, it's completely unsurprising that the exact same outcome might be legal or illegal depending on how it was obtained (or even with what intent the actions were done), legality takes these things into account.

I wouldn’t be too surprised if tools are created to quantify the amount of a source image that’s gone into an output image.

Someone will create some AI that detects if the image contains stuff from their copyrighted portfolio

The watermark.

What watermark?

The model is open source and free for people to make their own modifications. I don't see how any watermark can survive in those conditions.

https://news.ycombinator.com/item?id=32573523

All this is super, super interesting to me. I don't think anybody really knows how it will play out, there's a lot of good arguments floating around.

All images from stable diffusion have an invisible watermark unless you modify the source to remove it.

It's extremely trivial to remove the watermark.

I fail to see how the brain isn't doing a very similar thing?

Is my brain's floating point calculations subject to round off error?

In the (actual) neurons, is there a representation of real numbers? Where are the numbers in the brain stored?

I feel like people who assert this neither understand what neural networks are and how brains work.

> Is my brain's floating point calculations subject to round off error? > > In the (actual) neurons, is there a representation of real numbers? Where are the numbers in the brain stored?

None of these details are relevant to the bigger picture similarities of non-hardcoded learning from training data. None of these details change the ethics of what's being discussed here.

> I feel like people who assert this neither understand what neural networks are and how brains work.

That's just your bias showing.

> What they don’t do is mechanistically clone compressed mathematical representations of input data.

Don’t they, though? What else do you think is behind the curtain if not mathematics?

> The fun part is, this is how human artists learn too.

We don't actually know exactly how human artists learn, and human artists are capable of innovation, nobody knew pointillism or Bauhaus before they were invented.

A little know fact is that for humans it takes a long long time to learn, while they learn, they develop a style, if they don't they are not "real" artists, but merely executors, artists evolve, sometimes dramatically, in unexpected ways [1] [2].

So for us humans learning is an experience, not just recombining parts of features of other things.

We are also highly influenced by feelings, unfortunately, so sometimes we do things a certain way because we felt that way, not because we wanted to paint that thing that way, or because we are not good enough to do exactly what we wanted to do.

Is Mona Lisa happy? Who can tell?

Was Leonardo happy when he painted it?

What was Leonardo thinking when he painted it?

What was happening in his life?

Is that the best smile Leonardo could paint or it's an enigma he put there for future generations?

These questions are more important for an artist than the mere features of the painting.

The philosophical question is: is art discovered or invented?

If it's discovered, then SD can generate art, if it's invented, than SD it's not even generative work, because to invent something from something else, you need inventiveness.

[1] Picasso 1896 https://mymodernmet.com/wp/wp-content/uploads/2018/01/pablo-...

[2] Picasso 1946 https://www.photo.rmn.fr/CorexDoc/RMN/Media/TR1/MS4GY/16-515...

Artist today use the exact same method of learning from other peoples artwork to generate new artwork and styles. These models are learning just like any artist learns and then producing new content.

This is patently, obviously _wrong_ for anyone who has tried learning any artistic skill in their life. Sorry to be this straightforward, but it gets on my nerves every time I read it.

If you tried learning, let's say, the chiaroscuro technique from Caravaggio you'd be analyzing the way the painter simulated volumetric space by using white and dark tones in place of natural lighting and shadows. You wouldn't even think of splitting the whole painting into puzzle size pieces while checking how many how those look similar when put close one another.

Given somewhat decent painting skills, you'd be able to steadily apply this technique for the rest of your life just by looking at a very small sample of Caravaggio's corpus.

On the other hand if you tried removing even just a single work from the original Stable Diffusion data set you used to generate your painting, it would be absolutely impossible to recreate a similar enough picture even by starting from the same prompt and seed values.

Given how smart some of the people working on this are, I'm starting to believe they're intentionally playing dumb to make sure nobody is going ask them to prove this during a copyright infringement case.

>This is patently, obviously _wrong_ for anyone who has tried learning any artistic skill in their life. Sorry to be this straightforward, but it gets on my nerves every time I read it.

Both my parents (though retired now) were commercial artists. I was trying to be an artist at one point in my life before moving in Engineering and Science. All my parents friends are artists so I grew up around artists.

Ask any artist here who is using illustrator, Photoshop, Krita etc. How often do they google image search for textures, or reference images that gets incorporated into their artwork? The final artwork is their own but it may incorporate many elements from others artwork.

>If you tried learning, let's say, the chiaroscuro technique from Caravaggio.. You wouldn't even think of splitting the whole painting into puzzle size pieces while checking how many how those look similar when put close one another.

Ever seen hyperealistic pointillism?

Who are you to be the arbitrator of how an artists creates their work? Have you ever gone to a modern art gallery and seen all the different methods people use to create artwork?

Art is boundless and unique to each who creates it.

If a Artist uses a tool to create art, everyone agrees that is art. It could be a paint brush, clay, software on a computer etc etc. If an artist uses AI as a tool to create art then suddenly it's not art.

Even if that's true (which it isn't), it doesn't matter. It's absolutely logically consistent to allow humans to do it while forbidding AI to do it.

> It's absolutely logically consistent to allow humans to do it while forbidding AI to do it.

Why would you, though? If art is for the sake of art, then all art is valuable regardless of origin. If art is for the sake of providing human employment, AI being better in no way stops performative make-work from existing. If art is for the sake of copyright trolls to troll harder, then fuck art, feed it to the AI!

Art is for the sake of humans producing said art. I don't give a flying f*ck about "art" (it's not art) generated by a neural net using statistical patterns.

This is such a strange position to hold. At some point in the future may be confronted with an image that is emotionally stimulating, and you will have no idea whether a human or AI created it. Are you suddenly going to dismiss and discard it just because you subsequently learn it was AI generated? You must see how silly that is.

Your "logic" for making AI art illegal is basically "don't like it". Your personal and subjective opinion is that it's not art by definition.. This is like refusing to eat artificially grown meat because you have some strange idea about what food "should" be. Even if the meat was made MORE delicious you would still claim it wasn't food and turn it away. There's no logical consistency to your position, it's purely reactionary.

A good chunk of art is just that, to my understanding. People go to art shows with a knowledge by whom, when and “how” the art was done. They will be very confused if asked to tell two pics apart if you time-travel to Picasso, ask him to paint a new unique pic and then generate another one with AI. They can even find an idea, symbolism and what he felt/thought of behind an AI version.

All this boils down to a simple fact that egos like to think of themselves (and of artistic interaction) much more than there actually is.

I remember a story when a literature teacher insisted on a definite symbolism of some minor detail in a novel. People contacted the author about it and he said no, there is nothing behind it. It was just a filler without any second thought. Makes you think how much symbolism is far-fetched in classics, where you cannot simply email an author.

My position is perfectly logically consistent. Art is distilled human experience and human emotions emerging in a particular context after a chain of events. None of this is true of AI "art".

What is not logically consistent is to claim that a black box utilizing statistical relationships between pixels in a giant dataset is an "artist" and that its products create "value".

The compiler is not a programmer, AI can never be an artist.

The chain of events:

Step 1: Tweak settings and type text

Step 2: Look at the result

Step 3: If you like the result, go to step 4, otherwise go back to step 1.

Step 4: Save and share the result

Feels like art to me. Ultimately it's still a human using a tool to create art. For me AI art is just the name of the art style.

By that logic, a google image search + sharing results you like is also art. But it really isn't. At best it's curation.

The problem is forbidding it won’t do much.

- It appears people can train AIs from scratch or at least fine-tune them at home.

- Even if your art isn’t in “the training set”, that does not prevent the AI from learning its style. (Someone can decode it to CLIP embeddings. It could have a really good text model trained on vivid art museum descriptions of your art.)

- The ability of an image model to generate your art means it could also be trained in reverse to recognize it, producing a caption model, which would give vision to the blind. And surely you’d feel bad about that.

If you forbid it, the investment in developing those models disappears. They will be stuck at ~what we have now at best.

You can also require cloud providers to enforce a ban on training (and deploying) such models, it's doable. Good luck training it in your basement, it will probably take you a decade.

If this is banned, it will become a lot like piracy - yes, it's available, no, most people (at least in the West) don't do it, practically no businesses do it.

Training these models is much much cheaper than you think it is, and there’s good data for it already.

Either use a CC0 set like Wikimedia/Flickr and throw in some dead artists like Brueghel, or train on data from a country we don’t respect the IP of. Lots of Taobao product photos out there. It’s enough.

You are getting ridiculous now. Training this on "Taobao product photos" will lead to a useless model that is unable to produce practically all of the "cool" demos posted here in the last week.

A few months ago this task was virtually impossible. Then it was possible, but extremely expensive and pay-walled behind the "Open"AI' website.

As of about a week ago this tech runs on consumer GPUs. The weights have been downloaded 100s of thousands of times, and fine-tuning / modifying is possible.

Training from scratch is about $500k still, but it will only get cheaper and easier.

This doesn't contradict anything I have written. The average technical user will be unable to train this exact model (not to mention the supposed future more powerful ones) in their basement in this decade.

That just feels like such a pessimistic forecast to me. Of course, the current trajectory of improvements in model efficiency and better commercial GPUs / ML-accelerators may hit a wall.

But I would not be surprised if this was trainable on a commercial GPU at home within that time. But I think another important trend that we are seeing is that you don't need to train these models from scratch.

Open-source "foundation models" means that you can usually get away with the much easier task of fine-tuning, as to not throw away / re-learn everything that these large models have already fit.

Edit: I initially said 2-5 years, but on more reflection this does seem optimistic (for training from scratch).

If things go that way, the 'legit' models will continue dev, just using licensed content (along with public domain works). It will be more expensive for the end user, but that cost will shrink over time for general work. Tools that mimic working artist though might not be available (or will expensive). This all seems pretty ideal, so the pessimist in me guesses it's fairly unlikely.

I am not sure it will be possible to get enough training data that way.

I don't know enough about diffusion models but if LLMs (of current size) have to use only public domain, they will be undertrained and we will see significant degradation in performance. Not to mention that Codex will be effectively dead.

Human artists can copy an art using transparent tracing paper on a lightbox just like that, and that will be plagiarism. No reasons AI output shouldn’t be.

How about if I use Stable Diffusion, running 24/7 on a massive set of parallel clusters, and file for IP protection on all of the resulting images?

What potentially human-creatable images have I just taken ownership of?

Let's extend: what if I claim IP ownership of every image which StableDiffusion could produce?

The main reason is that you can’t, because only a work produced by a human is eligible for copyright.

For example, it has previously been litigated that nobody has IP ownership of an image taken with a camera by an animal.

Reply on news.ycombinator.com