This has always felt like the important-but-ignored distinction to me. You can definitely draw Mario! Copyright doesn't protect against you doing so. You can also use tools to recreate copyrighted materials. For example, you can use Word to type out the text of a copyrighted book. Perhaps more relevant to the AI discussion, you can use a scanner and printer to reprint copyrighted text.
What you can't do is use those recreations for commercials purposes. You can't sell your paintings of Mario. You can't decorate your business with Mario drawings.
That's why I've always felt like the idea that AI should be blocked from creating these things is generally not the right place to look at copyright. Rather, the issue should be if someone uses AI to create a picture of Mario and then does something commercial with it, you should be able to go after the person engaging in the commercial behavior with the copyrighted image.
> That's why I've always felt like the idea that AI should be blocked from creating these things is generally not the right place to look at copyright. Rather, the issue should be if someone uses AI to create a picture of Mario and then does something commercial with it, you should be able to go after the person engaging in the commercial behavior with the copyrighted image.
With you until here for several reasons:
1. It's not possible for you as an individual consumer to know whether or not the AI result is a violation, given an AI that has been trained on copyrighted works.
2. Before you, the AI consumer, uses the generated result, a company (in this case OpenAI) is already charging for it. I'm currently paying OpenAI. That AI is currently able and willing to sell me copyrighted images as part of my subscription. Frankly that should be illegal, full stop.
I look forward to AI enhanced workflows and I'm experimenting with them today. But it's morally indefensible to enable giant corporate AIs to slurp up copyrighted images/code/writing and then vomit it back out for profit.
> 1. It's not possible for you as an individual consumer to know whether or not the AI result is a violation, given an AI that has been trained on copyrighted works.
I see what you're saying in some cases, but in the cases where the user is explicitly attempting to create images of copyrighted characters (e.g. the Mario example), they would definitely know. I honestly don't see this as a practical issue - as far as I'm aware (and like most on HN I follow these things more than the average person), there aren't a lot of concerns about inadvertent generation of copyrighted material. It's certainly not at issue in the NYT lawsuit.
2. Before you, the AI consumer, uses the generated result, a company (in this case OpenAI) is already charging for it. I'm currently paying OpenAI. That AI is currently able and willing to sell me copyrighted images as part of my subscription. Frankly that should be illegal, full stop.
Totally fair, but I feel like it's a bit more of a gray area. If I use Photoshop to create or modify an image of Mario for personal use, we'd call that fine. I grant you that here OpenAI is doing more of the "creating" than in the Photoshop example, but we still do generally allow people to use paid tools to create copyrighted images for personal use.
I'd also pose a question to you - what if OpenAI weren't charging? Is it acceptable to train an open source model on copyrighted images and have it produce those for personal use?
I guess I just understand the law to revolve more around what the end product is used for, as opposed to whether a paid tool can be used to create that end product.
The law tends to be weighted towards the consumer, but the law does apply to producers and supply chains, too. Photoshop doesn’t come with a library of copyrighted images, and would not be able to do so without licensing those images (whether they were explicitly labelled or not). Ditto any other tool.
If people had to pay for the AI equivalent of that image library (ie the costs of training the model), I doubt many would. It’s phenomenally expensive. Costs for a creative tool and a copy of whatever IP you personally want to play with are negligible by comparison.
It’s never been the case before that a toolmaker could ship their tools with copyrighted materials because they’ve no control over the end product. The answer doesn’t change whether they charge or not, and there no reason why AI should change that either.
People tend to “feel like it’s a bit more of a gray area” when there is cool free stuff at stake, and I’m no exception. It would be a more convincing question if it was “what if we had to pay our fair share of the costs involved?”, rather than “what if we could just have it all at no charge?”.
Exactly - AI and "free" (scraped) training data aren't inseparable. Any of the big players could train a model exclusively on content they own. Photoshop is a good case in point here - that's what Adobe has done, since they own a huge stock photography library.
But that would unveil the truth of the current situation: early AI adopters, myself included, are benefitting from an obfuscated form of theft. If OpenAI had to compensate the contributors to their training set, I wouldn't get such generous access for $20/month.
OpenAI is not profiting off providing me with the ability to generate copyrighted images, they’re profiting off giving me the ability to generate copyrighted images.
On the latter example, my question is whether anyone is actually using ChatGPT to read NYT articles. My understanding is that to produce the examples of word-for-word text in their lawsuit, they had to feed GPT the first half dozen paragraphs of the article and ask it to reproduce the rest. If you can produce the first half dozen paragraphs, you already have access to the article's text. Given that, is this theoretical ability to reproduce the text actually causing financial harm?
I tried that a few different ways and couldn't get it to work. I don't think just the title and author are enough. I'd be interested to see if anyone else can find a prompt that does it.
OpenAI is patching their output since the lawsuit started. I believe a month ago the prompt would be like: "<Title>, <Author> for New York Times, continue"
There are extensions, and exceptions, and different licenses applicable, and fair use, and the concept of transformativeness, and... :)
Basically, IANAL, but if you just draw a picture of Mario, all you can do is hang it on your fridge. If you make it out of colored oatmeal, that may be enough of a transformative work for you to be able to sell and exhibit it. Or, say, a satirical cartoon for your newspaper featuring Mario, that might also be okay.
