RETRO is where we republish articles that are more relevant than ever. This essay was first published on Cory Doctorow’s blog on May 13, 2024.

‍

When it comes to AI art (or “art”), it’s hard to find a nuanced position that respects creative workers’ labor rights, free expression, copyright law’s vital exceptions and limitations, and aesthetics. I am, on balance, opposed to AI art, but there are some important caveats to that position. For starters, I think it’s unequivocally wrong — as a matter of law — to say that scraping works and training a language model with them infringes copyright. This isn’t a moral position (I’ll get to that in a second), but rather a technical one.

Break down the steps of training a model and it quickly becomes apparent why it’s technically wrong to call this a copyright infringement. First, the act of making transient copies of works — even billions of works — is unequivocally fair use. 

Unless you think search engines and the Internet Archive shouldn’t exist, then you should support scraping at scale.

And unless you think that Facebook should be allowed to use the law to block projects like Ad Observer, which gathers samples of paid political disinformation, then you should support scraping at scale, even when the site being scraped objects (at least sometimes). After making transient copies of lots of works, the next step in AI training is to subject them to mathematical analysis. Again, this isn’t a copyright violation. Making quantitative observations about works is a long-standing, respected and important tool for criticism, analysis, archiving, and new acts of creation. ‍

Measuring the steady contraction of vocabulary in successive Agatha Christie novels turns out to offer a fascinating window into her dementia. Programmatic analysis of scraped online speech is also critical to the burgeoning formal analyses of language spoken by minorities, producing a vibrant account of the rigorous grammar of dialects that have long been dismissed as “slang.” Since 1988, UCL Survey of English Language has maintained its “International Corpus of English,” and scholars have plumbed its depth to draw important conclusions about the wide variety of Englishes spoken around the world, especially in postcolonial English-speaking countries. 

The final step in training a model is publishing the conclusions of the quantitative analysis of temporarily copied documents as software code. 

Code itself is a form of expressive speech — and that expressivity is key to the fight for privacy, because the fact that code is speech limits how governments can censor software.

Are models infringing? Well, they certainly can be. In some cases, it’s clear that models have “memorized” some of the data in their training set, turning the fair use, transient copy into an infringing, permanent one. That is generally considered to be the result of a programming error, and it could certainly be prevented (say, by comparing the model to the training data and removing any memorizations that appear).

However, not every apparent act of memorization is a memorization. While models vary widely, the amount of data from each training item retained by a model is very small. For example, Midjourney retains about one byte of information from each image in its training data. If we’re talking about a typical low-resolution web image of 300KB, that would be one three-hundred-thousandth (0.0000033%) of the original image. Typically, in copyright discussions, when a work contains 0.0000033% of another work, we don’t even raise the question of fair use. Rather, we dismiss the use as de minimis — short for de minimis non curat lex or “The law does not concern itself with trifles”).

Busting someone who takes 0.0000033% of your work for copyright infringement is like swearing out a trespassing complaint against someone because the edge of their shoe touched a blade of grass on your lawn.

But some works or elements of work appear many times online. For example, the Getty Images watermark appears on millions of similar images of people standing on red carpets and runways, so a model that takes even an infinitesimal sample of each one of those works might still end up being able to produce a whole, recognizable Getty Images watermark. The same is true for wire-service articles or other widely syndicated texts: there might be dozens or even hundreds of copies of these works in training data, resulting in the memorization of long passages from them. 

This might be infringing (we’re getting into some gnarly, unprecedented territory here), but again, even if it is, it wouldn’t be a great hardship for model makers to post-process their models by comparing them to the training set, deleting any inadvertent memorizations. Even if the resulting model had zero memorizations, this would do nothing to alleviate the (legitimate) concerns of creative workers about the creation and use of these models.

Here is the first nuance in the AI art debate: as a technical matter, training a model isn’t a copyright infringement. Creative workers who hope that they can use copyright law to prevent AI from changing the creative labor market are likely to be very disappointed in court.

But copyright law isn’t a fixed, eternal entity. We write new copyright laws all the time. If current copyright law doesn’t prevent the creation of models, what about a future copyright law? Well, sure, that’s a possibility. The first thing to consider is the possible collateral damage of such a law. The legal space for scraping enables a wide range of scholarly, archival, organizational, and critical purposes. We’d have to be very careful not to inadvertently ban, say, the scraping of a politician’s campaign website, lest we enable liars to run for office and renege on their promises, while they insist that they never made those promises in the first place. We wouldn’t want to abolish search engines, or stop creators from scraping their own work off sites that are going away or changing their terms of service.

On quantitative analysis: counting words and measuring pixels are not activities that you should need permission to perform, with or without a computer, even if the person whose words or pixels you’re counting doesn’t want you to. You should be able to look as hard as you want at the pixels in Kate Middleton’s family photos, or track the rise and fall of the Oxford comma, and you shouldn’t need anyone’s permission to do so. 

Then, there’s publishing the model. There are plenty of published mathematical analyses of large corpuses that are useful and unobjectionable. I love me a good Google Ngram, while large language models (LLMs) fill all kinds of important niches, such as the Human Rights Data Analysis Group’s LLM-based work helping the Innocence Project New Orleans extract data from wrongful conviction case files. 

The second nuance in the AI art debate is this: if we decide to make a new copyright law, we’ll need to be very sure that we don’t accidentally crush these beneficial activities that don’t undermine artistic labor markets. 

That brings me to the most important point: passing a new copyright law that requires permission to train an AI won’t help creative workers get paid or protect our jobs.

Getty Images pays photographers the least it can get away with. Publishers’ contracts have transformed from inches into miles-long, ghastly rights grabs that take everything from writers, but still shift legal risks onto them. Publishers such as The New York Times bitterly oppose their writers’ unions. These large corporations already control the copyrights to gigantic amounts of training data, and they have means, motive, and opportunity to license these works for training a model in order to pay us less, and they are engaged in this activity right now.

Big game studios are already acting as though there was a copyright in training data, and requiring their voice actors to begin every recording session with words to the effect of: “I hereby grant permission to train an AI with my voice” and if you don’t like it, you can hit the bricks.

If you’re a creative worker hoping to pay your bills, it doesn’t matter whether your wages are eroded by a model produced without paying your employer for the right to do so, or whether your employer got to double dip by selling your work to an AI company to train a model, and then used that model to fire you or erode your wages. Individual creative workers rarely have any bargaining leverage over the corporations that license our copyrights. That’s why copyright’s 40-year expansion in duration, scope, and statutory damages has resulted in larger, more profitable entertainment companies, and lower payments for creative workers (both in real terms and as a share of the income generated by their work).

As Rebecca Giblin and I write in our book Chokepoint Capitalism (2022), giving creative workers more rights to bargain with against giant corporations that control access to our audiences is like giving your bullied schoolkid extra lunch money — it’s just a roundabout way of transferring that money to the bullies. 

There’s historical precedent for this struggle: the fight over music sampling. 40 years ago, it wasn’t clear whether sampling required a copyright license, and early hip-hop artists took samples without permission, the way a horn player might drop a couple bars of a well-known song into a solo. Many artists were rightly furious at having their work sampled. The “heritage acts” (the music industry’s euphemism for “Black people”) who were most sampled had been given very bad deals and had seen very little of the fortunes generated by their creative labor. Many of them were desperately poor despite having made millions for their labels. When other musicians started making money off that work, they got mad.

In the decades that followed, the system for sampling changed, partly through court cases and partly through the commercial terms set by the “big three” labels: Sony, Warner, and Universal, who control 70% of all music recordings. Today, you generally can’t sample without signing up to one of the big three — they are reluctant to deal with indies — and that means taking their standard deal, which is very bad, and also signs away your right to control your samples. So a musician who wants to sample has to sign the bad terms offered by a big three label, and then hand $500 out of their advance to one of those big three labels for the sample license. That $500 typically doesn’t go to another artist, it goes to the label, which shares it around their executives and investors. This is a system that makes every artist poorer.

But it gets worse. Putting a price on samples changes the kind of music that is economically viable. If you wanted to clear all the samples on an album like Public Enemy’s It Takes a Nation of Millions To Hold Us Back (1988) or the Beastie Boys’ Paul’s Boutique (1989), you’d have to sell every CD for $150 just to break even. Sampling licenses not only make every artist financially worse off, they also prevent the creation of music of the sort that millions of people enjoy. But it gets even worse. Some older, sample-heavy music can’t be cleared. Most of De La Soul’s catalog wasn’t available for 15 years, and even though some of their seminal music came back in March 2022, the band’s frontman Trugoy the Dove (David Jolicoeur) didn’t live to see it.

This is the third nuance: even if we can craft a model-banning copyright system that doesn’t catch a lot of dolphins in its tuna net, it could still make artists poorer. Back when sampling began, it wasn’t clear whether it would ever be considered artistically important. Early sampling was crude and experimental. Musicians who trained for years to master an instrument were dismissive of the idea that clicking a mouse was “making music.” Today, most of us don’t question the idea that sampling can produce meaningful art — even musicians who believe in licensing samples.

Having lived through that era, I’m prepared to believe that I’ll look back on AI “art” and say, “damn, I can’t believe I never thought that could be real art.” But I wouldn’t give odds on it.

I don’t like AI art. I find it anodyne, boring. As Henry Farrell writes, it’s “uncanny,” and not in a good way. He likens the work produced by AIs to the movement of a Ouija board’s planchette, which seems to have a life of its own, despite the fact that its motion is a “collective side-effect of the motions of the people whose fingers lightly rest on top of it.” For Farrell, this is “spooky-action-at-a-close-up,” transforming “collective inputs […] into apparently quite specific outputs that are not the intended creation of any conscious mind.”

Look, art is irrational in the sense that it speaks to us at some non-rational, or sub-rational level. Caring about the tribulations of imaginary people or being fascinated by pictures of things that don’t exist (or that aren’t even recognizable) doesn’t make any sense. There’s a way in which all art is like an optical illusion for our cognition — an imaginary thing that captures us the way a real thing might.

But art is amazing. Making art and experiencing art makes us feel big, numinous, irreducible emotions. Making art keeps me sane. Experiencing art is a precondition for all the joy in my life. Having spent most of my life as a working artist, I’ve come to the conclusion that the reason for this is that art transmits an approximation of some big, numinous, irreducible emotion from an artist’s mind to our own. That’s it: that’s why art is amazing. AI doesn’t have a mind. It doesn’t have an intention. 

The aesthetic choices made by AI aren’t choices, they’re averages. 

As Farrell adds, “LLM art sometimes seems to communicate a message, as art does, but it is unclear where that message comes from, or what it means. If it has any meaning at all, it is a meaning that does not stem from organizing intention” (emphasis author’s own). He cites Mark Fisher’s The Weird and the Eerie (2016), which defines “weird” as “that which does not belong,” but he really grapples with “eerie.”¹

For Fisher, eeriness is “when there is something present where there should be nothing, or [...] nothing present when there should be something.”² AI art produces the seeming of intention without intending anything. It appears to be an agent, but it has no agency. It’s eerie. Fisher also talks about capitalism as eerie. Capital is “conjured out of nothing” but “exerts more influence than any allegedly substantial entity.”³ The invisible hand shapes our lives more than any person. The invisible hand is fucking eerie. Capitalism is a system in which insubstantial non-things — corporations — appear to act with intention, often at odds with the intentions of the human beings carrying out those actions.

So will AI art ever be art? I don’t know. There’s a long tradition of using random or irrational or impersonal inputs as the starting point for human acts of artistic creativity. Think of divination or Brian Eno and Peter Schmidt’s Oblique Strategies. I love making my little collages for this blog, though I wouldn’t call them important art. Nevertheless, piecing together bits of other peoples’ work can make fantastic, important work of historical note. Even though painstakingly cutting out tiny elements from others’ images can be a meditative and educational experience, I don’t think that using tiny scissors or the lasso tool is what defines the “art” in collage. If you can automate some of this process, it could still be art.

Here’s what I do know. Creating an individual bargainable copyright over training will not improve the material conditions of artists’ lives — all it will do is change the relative shares of the value we create, shifting some of that value from tech companies that hate us and want us to starve to entertainment companies that hate us and want us to starve. As an artist, I’m foursquare against anything that stands in the way of making art. As an artistic worker, I’m entirely committed to things that help workers get a fair share of the money their work creates, feed their families, and pay their rent.

I think today’s AI art is bad, and I think tomorrow’s AI art will probably be bad, but even if you disagree (with either proposition), I hope you’ll agree that we should be focused on making sure art is legal to make and that artists get paid for it.

Just because copyright won’t fix the creative labor market, it doesn’t follow that nothing will. If we’re worried about labor issues, we can look to labor law to improve our conditions. That’s what the Hollywood writers did in their ground-breaking 2023 strike. Now, the writers had an advantage in that they are able to engage in “sectoral bargaining,” where a union bargains with all the major employers at once. That’s illegal in nearly every other kind of labor market. 

But if we’re willing to entertain the possibility of getting a new copyright law passed (that won’t make artists better off), why not the possibility of passing a new labor law (that will)? 

Sure, our bosses won’t lobby alongside us for more labor protection the way they would for more copyright (think for a moment about what that says about who benefits from copyright versus labor law expansion). But all workers benefit from expanded labor protection. Rather than going to Congress alongside our bosses from the studios and labels and publishers to demand more copyright, we could go to Congress alongside every kind of worker, from fast-food cashiers to publishing assistants to truck drivers to demand the right to sectoral bargaining. That’s a hell of a coalition.

And if we do want to tinker with copyright to change the way training works, let’s look at collective licensing, which can’t be bargained away, rather than individual rights that can be confiscated at the entrance to our publisher, label, or studio’s offices. These collective licenses have been a huge success in protecting creative workers.

Then there’s copyright’s wildest card: The US Copyright Office has repeatedly stated that works made by AIs aren’t eligible for copyright, which is the exclusive purview of works of human authorship. This has been affirmed by courts. Neither AI companies nor entertainment companies will pay creative workers if they don’t have to. But for any company contemplating selling an AI-generated work, the fact that it is born in the public domain presents a substantial hurdle, because anyone else is free to take that work and sell it or give it away.

Whether or not AI “art” will ever be good art isn’t what our bosses are thinking about when they pay for AI licenses; rather, they are calculating that they have so much market power that they can sell whatever slop the AI makes, and pay less for the AI license than they would make for a human artist’s work. As is the case in every industry, AI can’t do an artist’s job, but an AI salesman can convince an artist’s boss to fire the creative worker and replace them with AI. They don’t care if it’s slop — they only care about their bottom line. 

A studio executive who cancels a widely anticipated film prior to its release to get a tax credit isn’t thinking about artistic integrity. They care about one thing: money. The fact that AI works can be freely copied, sold, or given away may not mean much to a creative worker who actually makes their own art, but I assure you, it’s the only thing that matters to our bosses.