Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

> If artists can look at each others copyright-in-full-force works and then imitate it, why can't AI?

AI can reconstruct many training images in surprising detail. Show me an artist who can draw you an image they practiced on 10 years ago including the rough shape of the watermark.

> Of course generating pictures is different from using them, let's say in ads.

Microsoft's Copilot is a paid service, same with OpenAI's upcoming ChatGPT Plus subscription. Nobody here claims that it's illegal to draw Mickey Mouse and hang it on your fridge, but those AI models aren't trained just for fun. If Microsoft is allowed to make money from ignoring countless license terms I should be allowed to make my own short film trained with Disney cartoons and sell it. But we all know how well that would turn out.



> Show me an artist who can draw you an image they practiced on 10 years ago including the rough shape of the watermark.

I can't but that doesn't mean there are none. For music we have a lot of examples of artists who can copy perfectly. For visual arts we have people who take a helicopter tour of a city and then draw it in amazing detail.

Yes, time discounts everything for humans. That doesn't mean human artists don't cargo-cult watermark-like things. (One example is putting text/symbols on clothes that the designer doesn't bother to understand.)

> Copilot and MickeyMouse

This thread started by someone commenting about paintings in the Alte Pinakotheke, then about Disney.

Now it's language models.

Obviously they are different, but what's more important is that copying is more clear-cut in code, less degrees of freedom compared to visuals.

That said, at this point it's important to discuss how copyright works. You can look at someone's code. You can copy the style. Let's say you never saw functional code, you see a lot of it, then you learn functional style. Neat. But you can't copy the Haskell standard library 1:1 without infringing.

If Copilot produces Haskell stdlib, the copyright holders can sue.

And of course that's again where politics comes in. Since this is a new phenomena that disadvantages many small copyright holders this is where society ought to step in and help them to enforce their copyrights.

(So it's important to note that copyright and the license terms don't kick in if it's not a derived work. And that's the discussion we should be having. Maybe as a quick hotfix society should tweak copyright law and make it more broad in terms of what it considers a derivative. We can also put in the specifier that "except for humans, if you swear you did not use an artificial neural net" or whatever. I don't think this makes much of a difference, unfortunately. AI - and other kinds of technological changes - makes a lot of existing inequality problems worse. The general solution is to fix the inequalities. Of course if politics were actually representative of people's needs the inequalities were also likely inconsequential.)


> I can't but that doesn't mean there are none.

I think the point here is that now you can. Your "AI" can draw it for you with extreme accuracy and then... is it yours? That's the real question here. Pretty simple.

What if we had a copyright law that did give like a percent of detail that has to be different, so you make a new drawing that has the required amount of difference, and it's yours right? Now your "AI" can be trained to make differences in drawings that specifically meet the requirement of however much difference between the original and yours will negate the ability of the original author to make a copyright claim. Pretty straightforward. It's not the case now, but when we have enough content that has to be judged by humans on whether or not it falls under copyright... Anyhoo... I'm not selling content, so I shouldn't care, right?</snark>




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: