[dlg]

I am not a lawyer, but I've had to argue about copyright with several.

In the United States, there are two bits of case law that are widely cited and relevant: In Kelly v. Arriba Soft Corp (9th), found that making thumbnails of images for use in a search engine was sufficiently "transformative" that it was ok. Another case, Perfect 10 (9th), found that thumbnails for image search and cached pages were also transformative.

OTOH, cases like Infinity Broad. Corp. v. Kirkwood found that that retransmission of radio broadcast over telephone lines is not transformative.

If I understand correctly, there are four parts to the US courts' test for transformativness within fair use (1) character of use (2) creative nature of the work (3) amount or substantiality of copying (4) market harm.

I'd think that training a neural network on artwork--including copyrighted stock photos--is almost certainly transformative. However, as you show, a neural network might be overtrained on a specific image and reproduce it too perfectly--that image probably wouldn't fall under fair use.

There are also questions of if they violated the CFAA or some agreement crawling the images (but Hiq v Linkedin makes it seem like it's very possible to do legally) and whether they reproduced Getty's logo in a way that violates trademarks (are they trying to use it in trade in a way there could be confusion though?)

[eslaught]

Search engines don't create market harm for a work because they don't compete with it. In fact, they do the opposite: they advertise the work, making it more accessible and increasing exposure.

These AI tools on the other hand seem to do the exact opposite. They can (or could, if they got good enough) absolutely compete with a work, and therefore seem like they create substantial market harm. The character of use also seems vastly different; AI tools are creating images explicitly to be consumed, vs a search engine is basically just an index, and only shows the image in so far as it needs to make it discoverable.

So three of the four tests for fair use seem clearly against AI image generation, at least to me. The only test that possibly goes in favor of AI is the amount or substantiality of copying, but AIs can easily reproduce images, or if not entire images, other substantial subsets of a composition.

I just don't get how these could possibly be fair use.

[gojomo]

As I see it, 3 of the 4 tests are strongly in OpenAI's favor; the 'market effect' is mixed.

(1) The use is highly transformative;

(2) the images used were offered to the anonymous browsing public (with watermarks);

(3) the end effect of training will only retain a tiny spectral distilled essence of any individual photo, or even a giant source corpus;

(4) there's a potential risk of market competition from the ultimate model output, for some uses – but that's also the most 'transformative' aspect.

Getty et al could potentially just ask creators of such models not to include their images – perhaps by blocking their crawling 'User-Agent' – and it might not make any real difference in the models.

[saurik]

I'm still not seeing the "transformative" argument: the point of transformation isn't "it is in a different format" but (to quote Wikipedia, which is, of course, dumb... I'm sorry ;P) where one "builds on a copyrighted work in a different manner or for a different purpose from the original". The reason a search engine thumbnail is transformative isn't because it has been transformed to make it smaller... it is because the purpose of the resulting use of the image is somewhat unrelated to the use the original author was going for when they made the original image. At issue here is then that, rather than using an original image from Getty Images, someone decided to take all of the images from Getty Images and churn them through some algorithm that generated an image that directly competed with the original images from Getty Images. So like, sure: if you really only narrowly want to talk about OpenAI, what they are themselves doing (training and distributing a model) might potentially be legal, but the people using the result would seem to be in serious hot water... oh, and actually, I think they run it all a service, don't they? So no: I don't even think that defense works, as OpenAI is in some sense not even selling a model, they are merely directly competing with Getty Images to provide sell photos to people.

[gojomo]

Autogenerated, often fantastical, never-seen-before AI images strike me as a paradigmatically 'transformative' use. It's novel. It's shocking to many practicioners how flexible & high-quality the images can be. It will unlock all sorts of new downstream creation.

The representation that feeds the generation is statistical, even to the point of being plausibly factual: these things/people/places/concepts can be abstractly represented as the balanced weights inside the model. And under US law, facts aren't copyrightable.

I could see a case being factored as: (1) the scraping/training/ephemeralization itself involves the usual copying of downloading/locally-processing images, like indexing, but all those 'copying' steps are fair-use protected, as science/transformative/de-minimus/whatever; (2) any subsequent new-image generation no longer involves any 'copying', only new creation from distilled patterns of the entire training corpus, in which Getty retains no 'trace tincture' of copyright-control. So there's no specific acts of illegal copying to penalize.

Also, a human artist would be allowed to review related Getty/etc preview images, free on the web, to familiarize themself with a person or setting, before drawing it themself, with their own flair – as long as they don't copy it substantially. Why wouldn't an AI artist?

[omnimus]

"AI artist" doesn't add any of its "own flair". It builds exclusively on past experience and work of humans. And it also directly completes with them without any thought of credit or compensation.

People are really underplaying how damaging this is going to be for the industry. It's going to completely decimate it. You can already see people using names of artists in the DALL-E prompt to get "their" work for few dollars avoiding any copyright or social issues.

Artists will suddenly be competing with AI on price and time - why we should pay you living wage when we instantly generate something close enough.

Why would anyone try to create some new aesthetic or push anything further if their effort will be replicated next week when the model gets updated with new source data. Everything is gonna get stuck to aesthetic of 2025 and before.

It's completely inhuman.

[gojomo]

The synergistic effect of all the AI's inputs absolutely results in a unique new 'flair', with extensions, reversals, and mash-ups of styles just as in human-made artistic styles.

And AI "builds exclusively on past experience and work of humans" just like any young new human artist equally does. In many cases, you can even tell the different models' outputs apart, not by raw quality or glitches, but by hard-to-describe aesthetic tendencies.

I share your concern on the effect on human artists – both the market for their work, and even their morale, when learning, knowing that decades of practice will still be outproduced by seconds of computation.

But I don't think the genie will be put back in the bottle, by either expansive interpretation of existing copyright law, or even new laws.

[throwaway675309]

One could make the same case about humans, nobody works in a vacuum. Even though he used it in a pejorative sense, Sir Isaac Newton, the famous English scientist, once said, “If I have seen further, it is by standing on the shoulders of giants.”

That humans are capable of developing their own style could still be argued that it's just a intermixing of previous work that they've seen, but they've combined it in a different way, which effectively is exactly what these generative systems do.

[leksak]

I'd argue that if an artefact such as a watermark is copying even more substantially than any other human would and that human would at best be labelled as unoriginal, or doing very derivative work or be in violation of copyright.

[throwaway82388]

Perhaps I’m misunderstanding your argument, but my counterexample would be: if a human digital artist transformed a Getty image, resulting a fantastical, never-before-seen result, using software like Photoshop, that use would be no more defensible. If anything, the vast scale at which this occurs in AI makes it worse.

[gojomo]

I think your hypothetical would depend on the character & extent of the transformation. Mere filters that leave the original recognizable? Probably an infringement. But creative application of transformations to express new ideas? Maybe not – especially if the derivative is a comment/parody on the original, that actually increases interest in it. Most art is a conversation with the past, reusing recognizable motifs & often even exact elements.

For example:

Andy Warhol died in 1987, 35 years ago. One of his 'Prince' collages dating to the early 80s used another photographer's photo, without permission. In 2019, one federal judge ruled that was not infringement. An appeals judge then said it was.

The Supreme Court has decided to take the case.

The US Copyright Office & Department of Justice agree with the photographer in briefs filed with the court... but the mere fact the Supreme Court took the case indicates they think there might be issues with the appeals court ruling. They might agree with the original judge!

Oral arguments come this October. See:

https://www.reuters.com/legal/litigation/us-backs-photograph...

So, when all the (possible) disputes over AI-training-on-copyrighted-images resolve – maybe in the 2030s or 2040s? – what will the laws say, & courts decide? It'll depend a lot on other specifics, & reasoning, that may not be evident now.

[gnopgnip]

These AI generated images are directly competing with stock images. AI tools are selling images to blogs and other customers that often would purchase stock images instead.

The "character of use" is not in favor of dall-e, it is a commercial use.

Copyright law does not require getty to block a user agents or ask them not to include their images.

Another issue here is that removing copyright management info like a watermark is a violation of the DMCA, separate from fair use or copyright infringement. These cases have statutory damages and attorneys fees awarded.

[gojomo]

Whether something is directly competing for the same business would have to be evidenced, and copyright doesn't mean protection from all possible competition - it's just one factor weighed. And fair use protects many commercial uses, too, depending on proportion/character-of-original/etc.

But also, none of these images are direct, or even necessarily subtantial, "copies" of other images. The generator learned from other images – the same as any human artist might.

No watermark has been removed; the bigger issue may be that the spectral watermark violates a trademark. (But, I doubt consumers are likely to be confused.)

[wettico]

"The generator learned from other images – the same as any human artist might."

A lot of people seem to make this comparison, but I don't think it's fair. It's wrong. A computer is capable of ingesting/processing and "learning" from images at a rate no human can possibly come close to matching. To elaborate, it is not actually learning in the way we normally think of it, as its "brain" is completely different from a human's brain. It is doing something entirely different that should have its own word. Human artists learn from other human artists' work. An AI does something else.

It's also worth noting that the art the AI was trained on was posted online when the technology didn't exist (or if it did in some form it was not in the state it is in now). So an artist having posted their art online for public consumption can't be equated with somehow consenting to its consumption by a web scraper / AI.

[gojomo]

It's great that human artists learn from, & introduce into their work, influences other than just patterns seen in other works.

But it's also great that AI artists can learn from more examples in a few minutes than a human artist might see in lifetime.

To say that's "not actually learning in the way we normally think of it" is superficially true, but it doesn't mean it's "not actually learning", or necessarily any worse than typical learning. It's so new, & we barely understand fully how it works or what its limits are. It might be better in many relevant & valuable aspects!

[prox]

It’s going to be interesting what the stock companies will do. Maybe they will make their own Image Generator. Perhaps we will see a case based on the new factor that is AI. An AI is not artist; they can’t be conflated. A decent artists can churn out maybe 5-10 works if he is productive. AI can churn out by the hundreds or thousands if needed. The process also isn’t the same.

Anyway it will be interesting to watch this space.

[rasz]

AI generated images cant be copyrighted.

[sokoloff]

I have a hard time agreeing with 3, given https://ibb.co/DzGR063

[bryanrasmussen]

aside from if it is not copyrighted the image, the Getty watermark usage probably might have a bunch of issues.

[hammock]

> Search engines don't create market harm for a work because they don't compete with it. In fact, they do the opposite: they advertise the work, making it more accessible and increasing exposure.

AMP, snippets, Knowledge Base and in-app browsers would like to have a word with you

[SquareWheel]

Knowledge Base I grant you, but snippets are a crucial feature to trust a result is correct before clicking through.

AMP is completely unrelated so I'm not sure why you mention it. Website owners have to create a specific version of their own site for AMP to even work.

[srg0]

It seems it is possible to generate images which are very similar to the existing stock photos if you feed getty images' description into DALL-E.

I tried it with a distinctive banana image:

https://imgur.com/a/0OrIr6e

[Dylan16807]

"very similar" insofar as it's following the narrow prompt, sure.

> Different runs can generate different size, orientation and placement of the bananas, as well as different shades of pink.

At that point it's definitely the curation causing any possible derivation. The image generator is innocently doing what you ask in an unbiased way.

[Thorrez]

Those bananas are completely different. There's no copyright infringement there. I could take a photo of a banana and photoshop it repeatedly onto a pink background. That would look just as similar, and there's no copyright problem there.

You can't copyright an idea.

[srg0]

Images are different, but it appears that DALL-E is inspired by the aesthetics and the layout of the copyrighted material.

Another example, picking a random image from the Getty Images site. "A young parkour flips through the city,guangzhou,china, - stock photo":

https://imgur.com/a/pPruwzA

The images are obviously different, but it appears that DALL-E maps the getty images description to similar tone, similar perspective, similar background, and similar weather conditions. I'm sure there are thousands of possible backdrops in Guangzhou, and many ways to show a parkour flip. Even in the Google image search results there's more variance than in the output of DALL-E.

So you can't copyright an idea, but you can certainly scrape a copyrighted DB with image metadata, and use it to create your own product. My point is that DALL-E itself might be a derivative work of Getty Images and thousands of other online catalogs.

[kriro]

Interesting. Adding "stock photo" to the string generated that getty tag? That is probably the most attackable (alas easy to fix) part of the issue. It will be an interesting question how close to the original a picture has to be to be considered the same (I'm sure there's some case law) and maybe there's some new research to be done regarding how to recreate the training data images with the correct search string (I suppose one could build an ML model for that).

Fun times ahead

[srg0]

No, I didn't get the tag. But I suppose that Getty metadata as well as the images were used for training.

[jcranmer]

From what I understand, the actual process of fair use boils down to "the judge decides in his/her gut if the use is fair, and then writes up the analysis to justify coming to that conclusion." If you look at the recent SCOTUS opinion in Google v Oracle, you can see how two judges can look at the same facts and come to almost diametrically opposed fair use analyses. My further understanding is that generally the #1 overriding concern in fair use analysis is money, which means you're more likely to see analysis along Thomas's dissent than Breyer's opinion.

In this case, let me give a fair use analysis that is going to suggest that this isn't fair. Factor 1 weighs against fair use: it's not transformative because, well, transformative is extremely narrowly interpreted against fair use. Factor 2 weighs against fair use because, well, it's factor 2 and it weighs against fair use unless the underlying copyright was paper-thin in the first place. In factor 3, it's weighing against fair use because it's not copying the minimal amount of the original work to get what it needs (it copied the watermark after all!). And factor 4 of course weighs against fair use because you're essentially creating stock images which is naturally in the exact same market that a stock image provider is in.

If you wanted to write a fair use analysis that finds fair use, you'd argue instead that the work was transformative, and the amount copied also weighs in favor of fair use (thus converting factors 1 and 3 to weigh in favor of fair use). You might try to argue that it's a completely different market, but I'm incredibly skeptical that such an argument could win over both a district court and an appeals court (although Breyer's opinion in Google v Oracle did basically follow this thread of analysis, its repetition is unlikely since everyone wants to pretend that Google v Oracle has 0 impact to anything outside of software). Such an analysis is possible, but unlikely, since the unspoken factor of "could you have paid for this" tends to be the factor that wins out over everything else.

Note that we are going to have a SCOTUS case in the fall that will specifically explore transformative uses in the context of fair use: Warhol v Goldsmith (https://www.scotusblog.com/case-files/cases/andy-warhol-foun...). I'm not going to hold my breath that the use will be found fair, though.

[csmpltn]

Putting aside the core question of the legality of training data on licensed material - what about the false advertising/copyright aspect that comes with slapping a "GettyImages" logo on some random nonsense generated by a "neural network"?

[visarga]

It's not worth discussing about Getty so much. AI labs will collect a dataset to predict if an image is watermarked. They will crawl to index the Getty images to make sure they are not in the training set. Then retrain and in 2 months the problem is solved. They can cut out a sizeable part of the training set without problem, the model will still be good.

They can also OCR the output to make sure there are no blacklisted words and use an index to skip all images that look too similar to the training data. Then the argument of copyright defenders is going to be weakened.

The fact that a prompt and curation are necessary also goes against the "AI works can't be copyrighted" narrative - it's generated by a human-AI team, so human work is part of the process.

The core of the issue I see is that human and AI both learn from the published media but an AI can both "see" and "draw" more than a human, so there is an important distinction there.

[csmpltn]

I understand that there are (both practical and theoretical) ways to reduce the chances of an AI generating an image that has copyrighted elements in it (such as the "GettyImages" logo).

I'm mostly curious about the legal aspects of having a black-box system that can - under some unknown circumstances - attach openly copyrighted or trademarked elements (such as a company logo) to a piece of work.

[sixothree]

> (2) creative nature of the work

Is AI even capable of having a creative nature. All that I see is re-use of source images.