AI Training, the Licensing Mirage, and Effective Alternatives to Support Creative Workers

On May 1, a US court for the first time considered whether and under what circumstances developers’ use of in-copyright works to train AI models is consistent with “fair use” and copyright law. Kadrey v. Meta – in which a group of authors has joined together to sue Meta for, among other things, using their copyrighted works to train AI models without permission – is just one of many such hearings coming in the future. There are now over forty copyright cases against a variety of generative AI developers pending in US courts.

Job protection has become a core argument in the copyright debate. Major media companies and some artists worry that the output of AI models – including images, screenplays, books, and other works – will decimate the jobs of creators, particularly those whose works are used in the training data. They argue that requiring AI developers to get licenses from copyright holders will help protect creators’ livelihoods.

Automation can pose significant challenges for creators, but requiring licenses for training data won’t impede automation or provide meaningful support to creators. In fact, it can make challenges from automation worse. Rather than running headlong into a copyright dead-end, it is critical that we look towards alternatives that effectively reckon with the underlying fears about the impact of automation, powerful companies, and a changing economy.

The jobs debate in the courtroom and the culture

Fears about the impact on creators’ jobs featured prominently in the Kadrey hearing. Presiding Judge Vince Chhabria was skeptical about evidence of harm to the types of works in this suit, which include a number of fiction works and memoirs, and acknowledged that fears may seem speculative. (In fact, in a separate case brought by book authors against AI developer Anthropic, Judge William Alsup called the idea that AI models will supplant all human work “baloney,” adding “I don’t believe that will ever happen.”)

At the same time, Chhabria cautioned against a “formalistic” approach that simply looked to past precedent and technologies, and he pushed Meta to identify the limits of its argument:

You have companies using copyright-protected material to create a product that is capable of producing an infinite number of competing products. You are dramatically changing, you might even say obliterating, the market for that person's work… This seems like a highly unusual case in the sense that though the copying is for a highly transformative purpose, the copying has the high likelihood of leading to the flooding of the markets for the copyrighted works.

As Chhabria indicates, previous landmark rulings have broadly affirmed “highly transformative” uses, including the creation of other digital tools, are lawful. Chhabria’s willingness to explore how such use could still nevertheless be infringing echoes fears presented in the broader debate. For example, consider an album consisting of “recordings of empty studios and performance spaces” released by a group of over 1,000 artists in protest of proposed reforms to the UK copyright regime to allow AI training on in-copyright-works. The group stated that the silent album “represent[s] the impact we expect the government’s proposals would have on musicians’ livelihoods.” Listeners are not only asked to contemplate the lost livelihoods of the musicians, but also to suffer the loss of human-created music altogether.

Such fears are not altogether new. At the advent of recorded music in the early 20th century, prolific composer John Philip Sousa testified to Congress that “These talking machines are going to ruin the artistic development of music in this country…. We will not have a vocal cord left. The vocal cord will be eliminated by a process of evolution, as was the tail of man when he came from the ape.” Similarly, nearly 70 years later, Jack Valenti invoked the metaphor of a destructive flood in testimony to Congress on behalf of the American film and television industry about then-nascent video recording tools: “[W]e are facing a very new and a very troubling assault on our fiscal security, on our very economic life and we are facing it from a thing called the video cassette recorder and its necessary companion called the blank tape. And it is like a great tidal wave just off the shore.”

If Sousa’s and Valenti’s arguments had been accepted, perhaps it would have cut off the raft of new creators and creativity unleashed by recorded music and mass video recording tools. By the same token, such arguments about AI could cut off new creativity today. In fact, a “flood” of novel works (as opposed to mere regurgitation of training data) would seem consistent with copyright’s purpose – incentivizing creativity for the public’s benefit.

But: what if this time is different, or, as Chhabria puts it, what if this is a “highly unusual case”? Will requiring licensing actually help address the worries about automation, and help ensure creators can make a good living?

Licensing requirements for training won’t stop AI-powered automation

Let’s start with whether licensing requirements are even relevant to automation’s impact on jobs. Rightsholders and artists rightly tend to assume that high-performing AI models can also be built with licensed data. But if that’s the case, the entire licensing debate is effectively irrelevant to addressing the job threats of automation.

A profit-maximizing company that deploys a high-performing AI model to generate AI art will deploy that technology even if the training data was licensed. The illustrator, writer, painter, or other artist whose works are replaced by the output of the model will still potentially be out of a job – again, even if all of the in-copyright works used by the model are fully licensed.

The profit-maximizing company might even be in the creative industry. If record labels can make equally profitable music without paying for artists’ labor, that reduces a cost on their balance sheet. “Empty studios and performance spaces” are just more costs a record label can eliminate. A record label already owns licenses for myriad pieces of music; record labels already use AI in their work, and they can use their existing content library to train generative AI tools that further reduce costs of doing business.

Put simply, licensing is irrelevant to whether companies will deploy AI at the expense of human creators.

But at least human creators will get revenue from licensing deals, right? Not so fast. High-performing generative AI models are built by training on many billions of data points — individual artists will see minimal direct revenue from any licensing scheme.

For a sense of the magnitude we’re talking about here, consider Stable Diffusion, a model trained on over 2 billion images in 2022 that kicked off early controversy about text-to-image generators and copyright. Researchers led initial development of the model. StabilityAI, one of the initial companies to invest in and commercialize Stable Diffusion, was valued at around $1 billion at the time that the model rose to prominence. Even if all of that company’s value were liquidated and went directly to artists, without any middlemen, that is still just a one-time check of $0.50 per work. Back-of-the-envelope math for other companies or types of models are no more encouraging.

In reality, the math on a licensing market is even worse than that for artists. In any actual implementation of licensing requirements, there will necessarily be a middleman, and there are two types that stand to benefit the most.

First, large tech companies, which have licenses to the user-generated content from their services to use in training, or to sublicense to others. And second, large media companies, which control the copyrights in works produced by musicians, journalists, filmmakers, and authors.

In other words, even if all AI models were trained on fully licensed data, it will be Disney, Getty, Universal Music, Google, Meta, Bytedance, and other such large data aggregators that benefit the most – not the individual creators. Chhabria’s fear of obliterated markets and “flooding of content” could still happen, and the rewards will be reaped by the biggest incumbents in the market today.

In fact, licensing requirements can act as a barrier to competition not only for AI models but also for creative works. All creativity builds on the past; copyright protects specific expression, but not ideas, facts, and other building blocks of knowledge. In the hearing, Chhabria speculated that cookbook writers may face ruinous competition; however, recipes have never been copyrightable, and it is unclear what interests are served by giving existing cookbook writers not only a copyright on their work but also a monopoly on the market for cookbooks going forward.

Given that requiring licensing won’t stop the “flood” anyway, the protection for incumbents won’t meaningfully trickle down and protect creators. Artist Mat Dryhurst succinctly summarizes this future: “Everyone is just going to have to get used to coexisting in a world of infinite kinds of good media that used to be challenging to make. No denial. No time.”

There are better ways forward for the creative community

That public benefit of a “flood” of good media is cold comfort when it comes to worries about creators' jobs and livelihoods. If not licensing requirements for training to address those concerns, then what?

While it is unlikely that there will be one solution given the diversity of creative sectors and circumstances, the following broad topics are areas we see opportunity for further development of specific policy interventions and for organized advocacy to support creative workers.

• Regulating AI outputs: There has already been some momentum to craft rights of publicity laws or create special protections for “digital replicas.” These approaches limit uses of an artist’s name, image, likeness, and voice in AI outputs, such as commercial uses that could compete with that artist. Such frameworks don't restrict the broader training of AI models but rather regulate specific AI-generated outputs
• Ensuring workers enjoy the gains: AI is poised to generate returns for any company that deploys it effectively, but questions remain about how best to ensure workers share in those benefits. Collective bargaining as a tool has already been deployed to influence how AI models are adopted and deployed in creative work during the Hollywood strikes of 2023. In those negotiations, key concessions from movie and TV studios included ensuring that human writers would not receive lower pay just because an AI was used for initial drafting. These types of limitations protect individual creators’ pay and use in a given workplace, rather than limiting AI development generally.
• Limiting concentration of power: Most artists are not unionized, so it is especially critical to ensure that there is a competitive marketplace for artists to shop their work (or “sell” their labor, to put it in crudely economic terms). In her article on “Intellectual Property as Labor Policy,” scholar Xiyin Tang highlights the importance of addressing monopsony in media markets. For instance, the US government blocked the merger of publishers Penguin Random House and Simon & Schuster because it would have “harm[ed] competition to acquire the publishing rights to ‘anticipated top-selling books,’ resulting in lower advances for the authors of such books.” Similarly, activist and writer Cory Doctorow targets Google’s and Facebook’s hold over the market for advertising as a lever to benefit journalists and the news ecosystem.
• Creative mechanisms to redistribute the gains: Redistribution of the gains from AI – that is, taxes and subsidies – is also an important lever to make it easier for people to make a living as an artist. Direct funding for the arts, including to artists themselves, as well as to performance spaces and exhibit halls, is worth considering. Some thinkers are developing tax proposals that would be narrowly output-oriented and thus focused only on systems that produce outputs that substitute for certain works.

All of these ideas are potentially fruitful avenues for supporting creative workers. But none are receiving comparable engagement from the broader community of stakeholders who are or might be concerned with the impacts of automation on creative workers.

Of course, many people may still advocate for licensing requirements for a variety of reasons, irrespective of whether they are a meaningful support for artists’ livelihoods. At the same time, solutions in these other spaces could bring together stakeholders who are otherwise opposed to one another on the training and licensing issues. Such an agenda is badly in need of development and represents a productive path forward for collaboration.