Hacker News Reader

Comment by somewhereoutth | Hacker News Reader

Depending on the outcome of the NYT case, I.P. may in fact be the death of (profitable) AI.

One problem with AI (vs e.g. Google search) is that there is no mechanism for attribution. So it is impossible for rights holders to understand whether their works are being monetised (unlike, say, Spotify).

Should NYT win, then either AI services will have to pay into a (very large) common pot for rights holders, which likely will make AI uneconomic (even if it does turn out to have some business benefit), or (paid) AI services will simply be banned.

View on HN

There’s no way AI services are going anywhere. If the US kneecaps itself for parasitic rent seekers then China will win and benefit from the productivity these tools produce.

So the proposal of sociopathic thieves is that we turn into china and steal people’s property? What an odd take.

It’s a hard world and we need to play to win.

Win what? This is about making pretty pictures or words, not a resource war. You think Chinese Hollywood will suddenly pop up and topple the US's?

Yeah by stealing things from our own people? Thats a dumb way to win.

> there is no mechanism for attribution.

What if there was?

What if I hand-painted an image (without AI) and God himself came down from heaven and explained to the world that my painting was derived 10% from my observations of the collected works of Lisa Frank, 3% from looking at Davinci paintings, and 87% a long tail of various sources. And then I sell my painting for $100.

OK. Now what?

> OK. Now what? Do I owe Lisa Frank $10?

No, because copyright law says that if you hand-paint an image (without AI), then as long as the images isn’t deemed (by a court, if necessary) to be a derivative work, then you, and nobody else, own the copyright. The word of God is not necessarily considered by the law. And even if the court takes the word of God as gospel, a 10% derivative work might not exceed the necessary threshold of infringement.

Very good!

Alternatively, if I created a large body of works that were 90% Lisa Frank and brought it to a large market distribution, making significant revenue, then she could sue me for damaging her market and her brand. Which she has done to others in the past and rightly won. Go Lisa Frank!

But, if I used an AI tool in the process of making an image that came out 10% Lisa Frank, she wouldn't have a strong case against me. Nor does she have a case against the million random teenagers hand-copying her style but not making significant revenue from their copies.

Either way, AI doesn't really factor in here. The results are the results and the markets are the markets regardless of AI or paint brushes.

Do you think it's reasonable for artists to want a way to forbid major AI corporations from training on their work? Because some people will stop producing/stop making available their work if they don't get access to a different way of doing so.

Much like how some artists are uncomfortable with making high-quality scans of their work available online, because it becomes prohibitively difficult to prevent people from selling prints. Forget the current legal framework, it seems fair to me to say "I don't want that happening to my art" when "that" is a corporation finding a way to monetize it. Even if the AI is using less than a pixel of your work, I think it's fair to ask that it not.

I get that the concept is upsetting. And, I'm not going to tell people to not be upset or not state their will. Ex: DeviantArt has instated an opt-in checkbox for "It's OK to train on my art" and I think that's great.

But, everything I've heard and learned about how AI works leads me to the conclusion that such a move is almost entirely symbolic. Counter to what our egos would like to believe, the vast majority of the training images and the vast majority of the learning value for AI comes from random, crappy photographs of various things. Without a million photos of dogs and trees, the AI can't make a stylized dog distinguishable from a stylized tree. And, it's turning out that, from that base of crappy photographs, it doesn't take many high-aesthetic images to make a high-aesthetic AI.

Quality aside, art still has copyright for these exact reasons (and photographs are no different, long settled case law). So you can't just say "well no one cares about this blurry dog photo". Courts need more precise language than "care".

If it needs millions of photos of dogs and trees you

1. Take them you self

2. Pay someone with a db of dog and tree photos soem agreed upon amount (Aka someone did the work for you already and you value their work)

3. You scrape only on photos you are sure are CC0 or any other kind of permissive license. I'm sure you can find a few million dog photos this way.

But clearly they value #2 without payment so lawsuits were inevitable. Maybe because #3 took too much time and they decide #2 is worth the legal fees to be first to the race (something else I'd want fixed, but that's another rabbit hole).

I'm not sure I like the direction AI is going anyway.

Right now we're on a path towards just a handful of megacorps with AI agents that will make large numbers of white collar workers redundant. Imagine something like closed-source software only much worse.

In the last six months model size has come way down and quality has gone way up through different training styles/methods. There is an incentive to release separate models capable of running on 8/16/32gb ram. On my laptop I have something like six free/open source models installed. LLMs are getting so easy to make that people are doing this in their spare time as a hobby. This is starting to feel like a small stepping stone towards AGI but only time will tell.

It’s not impossible. AI developers can 100% trace node that contribute to a result during the processing of a prompt, therefore they can be attributed.

They’ll never do this, of course, cause they’re riding high on wins of stupid people pushing “you can’t work backward from a result to find the nodes” (which is probably somewhat grounded in reality).

Attribution is absolutely not impossible, they just want to think it’s not possible, because enabling such a thing would show just how egregious the copying actually is.

> they just want to think it’s not possible

As far as I'm aware, what you propose really isn't a solved issue. Most naive approaches you'd think might work are either infeasible or give chaotic nonsense results.

Partial matching between the output text and the training data is possible as a step after generation (though doesn't determine whether it's coincidence).

>> One problem with AI (vs e.g. Google search) is that there is no mechanism for attribution.

Sorry if this is a silly question, but is that really the case? Can we not train an LLM on successively larger training sets, each will uniquely ID's model and associated better performance?

Are there other examples where the law crushed fundamental technological shifts?

(Not just delayed briefly by starving investment or something.)

That assumes that the current crop of LLMs are a fundamental technological shift.

Furthermore, rights holders did it for file sharing, why not AI too?

Crypto is the most recent one.

Stem cells

So essentially ai is profitable only if it resells stolen property. Almost as if it’s not intelligent and it only procedurally generates output. The more the better.

How is it 'stolen' if it was part of a public dataset? Is the argument that OpenAI was scraping data that should have been behind a paywall?

If I create a sculpture and put it on my front lawn, I think it's fair for me to complain if you scan it and start 3d printing and selling copies, even if I intended it to be in public view. Just because I want the public to be able to enjoy my version of something, doesn't mean I want people to feel free to ingest it and monetize it for personal profit.

Right, but if I look at your sculpture and get ideas on how to make a better sculpture, then make an entirely different sculpture using yours as inspiration, and you get upset that I'm 'copying your style' we generally don't give you much sympathy.

Sure, I was commenting on the "public" aspect, not the "fair use" debate.

Right but you are not a 3D scanner are you?

And if I'm not, then neither is an Ai?..

Just because it's publicly viewable does not mean it's free for reuse. Imagine an author who uploads chapters of their book to their own website. You cant take that and claim it's yours.

Similarly with all the publically viewable source codes on GitHub/Lab. Just because you can view the code doesn't mean there aren't licenses in play for how you are able to use it.

Further, what happens if stolen information is posted publicly? (E.g. wholesale copies of full books on html pages?

Just because it's public doesn't mean the viewer has freedom to use it as they see fit.

That 'usage' is taking statistical notes about the work (creating factual statements about the work) and imputing those notes into a database, averaged with a few billion other notes about other works. That is a usage that copyright under current law simply doesn't cover or protect for. It doesn't even need the analysis if 'fair use' because, there's no copying or public performance happing in the creation/training of the model.

Where infringement arguably can happen is when that model is used in the generation of content - and if the user is prompting regenerate a protected work, then that is where the infringement happens. But not before. Maybe the various ai services can adequately guard against that illicit usage. Maybe not. And if not, its those live services that would need to be shut down.

But the creation and training of a model, and even distributing that model for people to use with their own computers in private does not engage in copyright infringement.

> creating factual statements about the work

All digital content is statements about the work. Pictures are pixels with facts about colours. Books are characters describing the words.

AI models are compressed versions of those.

Downloading, ingesting into models, and redistributing them is copyright infringement.

Even if they weren’t compressed; downloading them without permission is not permitted. Every now and then you can read about ai criminals discussing techniques for illegally acquiring books, images and other protected content even here on HN. It wouldnt be an issue if it were licensed content, or a dudette training her local ai. These are companies that want to monetise people’s work. Somehow, just like thieves, they feel they are entitled to snatching someone else’s property.

Reply on news.ycombinator.com