meow_d

First of all, there are (at least) two components in the datasets that train models like Dall-E 2, Midjourney and Stable Diffusion. The image and the description, which are delivered as a pair. How well the generated image matches the prompt is only as good as the description. The AI doesn't understand anything about the picture, it (very, very reductively) just matches particular pixel distributions with particular word sequences.

The huge leap forward in generative AI models that we've seen this year comes more or less from the huge quantity of training data used to train the models. And from some smart engineering by creating systems of specialized neural networks. But the images aren't really the hard part to come by. We live in a culture saturated with images. For the longest part, the classifying part was the hardest, assembling the pair. That's because it was done by hand, by people paid or incentivized to do it (or who were somewhat aware of what they were doing, like when we solve a Google Captcha). What was innovative about Dall-E 2, and then opened the floodgates when the other labs realized they could do the same, is using publicly available data from the web. The image descriptions done for SEO and accessibility reasons stored in the alt fields are a treasure trove of readily classified data, for example. These descriptions are sometimes written by the artists, but for the amount of data needed, it's almost certain that they've used images described by critics, curators (from museum websites, for example) and especially by anonymous data entry and SEO people (on stock image sites) or by random internet users assembling image boards on Pinterest, blogs and the likes.

Secondly, this copious amount of classified image data is much more than what a single artist can produce. But the model is not fed only the one artist's works, but also the works of every other artist that tried to understand the style and replicate it (for better or worse).

If the model were only fed the one artist's images, with people paid to classify and describe them, it would be piss poor. As they've been for years. When they work, they're not "stealing" the labor and ideas of a couple of people, but of hundred of thousands, maybe millions. Sure, for the vast majority of the people involved it's an infinitesimally small amount of labor time and that's why they more or less can get away with it.

It's also very diffused, starting with the works of a popular artist, then all the other artists that get lumped with them, all the other artists explicitly influenced by them, and finally everyone who posts their attempts at replicating a particular style either through studies or through exercises made for fun. And not even this greatly expanded dataset of images would mean squat, if not for the countless internet users who basically create a raw version of the set of classified images that gets turned into training data. They, we, are the ones doing the labor of understanding the work, of placing it in artistic continuums, of relating pieces of art with styles with artists and so on. The AI just creates probabilistic models based on all this deeply layered human labor of creating, understanding, describing and classifying art.

Sure, the likeliest scenario is that those campaigns will amount to nothing, at best removing the ability to use a few names in a prompt (although this surely can be circumvented).

The training data doesn't have to be stored anywhere, the model doesn't have to look at the image to "sample" it in the way lots of people think it does. The legality of having assembled the data is at most grey, there are even fewer problems with using the model, and there's no actionable way to easily remove an artist's impact on the network. You'd have to retrain the network with a dataset that doesn't use the offending images. And even in this absurd scenario, the effect will be minimal because of the way the dataset is constructed (you'd remove all the images owned by Sam Yang or Karla Ortiz or whatever, but won't remove those by the people drawing like them, save a global class action lawsuit that would take forever anyway), but also because there are network effects at play. "I drink your milkshake" stuff since all the data is transposed in multidimensional mathematical spaces. At the scale that these networks operate, the models can be made fill a void with something they hallucinate. This is how something like the Loab was made.

But if people try and try, at some point a campaign will succeed, especially if it's coopted by a larger entity that better understands the tech and the nefarious possibilities opened by the "reforms" needed to address these artists' concerns. Think only about the way Hachette Book Group, HarperCollins Publishers, John Wiley & Sons, and Penguin Random House used Chuck Wendig to sue The Internet Archive, with effects that will surely create precedents affecting other non-profits and the library system.

You don't have to think very hard to image some sort of new DRM dictating how images are shared and described online. Let's even invoke a nightmare scenario and bring the blockchain into discussion. Both smart contracts and NFTs could "solve" the problem, in all sorts of ways that I don't want to think too much about because I'll get sick. But some quick examples, just to show I'm not full of shit and paranoid, would be minting a number of NFTs to an artist proportionally to how often their name is used in a prompt; or smart contracts transferring to the original artist a fraction of the money involved in each transaction where an image strongly influenced by them is involved.

I hope it's obvious how such an arrangement would be detrimental to most artists. First of all, as we've established, in order to get the model to do a compelling Kim Jung Gi, it'd need the input of a lot of other people who aren't Kim Jung Gi. His name would absorb the value created by all of them like a sponge. And there's nothing to say that such an arrangement would be relegated only to ML generated art. If all this architecture is put into place, it'd be trivial to implement systems to prohibit saying you're inspired by Ortiz, since her name attracts clients and gets better SEO, unless you've got the token received after following her workshop. This is a recipe for cartelization and monopolization.

But to get any of that, you need a legal framework to enforce it and at the moment you still can't patent or copyright a style. Yet!

People think about intellectual property rights as being something universal and eternal, but in fact the field is very much in flux. The doctrine stated above can be, if not abolished, then at least circumvented. The current worldwide intellectual property architecture is actually very modern, formed in the mid 90s with the TRIPS agreement. The term itself basically didn't exist before the mid '80s. There were other regimes governing what we call IP, because it was acknowledged that it wasn't ... property.

And the way it changed gravely impacted countries in the Global South. Let's not forget that art and media, even though it's what we most often talk about, is only the tip of the iceberg with IP. We're talking about technology, about patents, it affects industry, agriculture and pharmacology. This system enforces dependency relations that not only impoverish developing countries, but can also have deadly consequences as we've seen during the COVID epidemic, regarding the patents for vaccines.

Evolving what counts as IP on the internet, how it's managed and monetized, isn't a speculative issue, but very much and active concern. Although its goals are noble, GDPR has been an overall mess and actually helped consolidate big tech. The same goes for "link taxes" implemented or proposed in places such as Canada, Australia or the EU. And these are only "small" interventions, lacking a public facing campaign or lawsuit, like the one I mentioned above between Internet Archive and the large publishers or the one between Metallica and Napster, that ended up establishing the DMCA.

I mentioned all these issues to establish three things.

• The fact that what counts as property shifts all the time both in the intellectual and digital spaces.
• The fact that what counts as legitimate sharing, using, reusing changes
• The fact that these changes rarely, if ever, help the little guy or gal, instead entrenching existing hierarchies

The way the neural networks were trained is a problem of the commons, of social ecosystems, of incredibly diffused networks of inspiration, influence and incrementally small transformative work. To adequately address it we need to accept this and work from there.

Any campaign that takes as its starting point the figure of the sole heroic creator exploited by a thinking machine will, first of all, misdiagnose the problem and propose inefficient solutions. But its ideological core is actually dangerous, since it'll propose solutions in the realm of regulations and intellectual property, it'll think in terms of individuals (be them persons or firms) and ownership, not in terms of social bodies.

The endgame of such a project can't be anything other than a form of rights management over style, method and technique. The ML labs will take it on the chin, the star artists will secure the bag and the vast majority of artists will be hurt, all while opening a Pandora's box that, going by past experiences, ushers only terrifying developments for the world at large. Image generators and chatbots are probably the most visible products made with ML, but they're hardly the only ones.