Benjamin Sobel has written a provocative new essay about copyright law’s role in the development of generative artificial intelligence, and I think all IP scholars should read it. His essay, Copyright Accelerationism, is the kind of creative scholarship that our field needs more of. It is being published in the Chicago-Kent Law Review’s “AI Disrupting Law” symposium issue.
As all readers will know, copyright law may stand in the way of further development of generative AI, because AI models are trained on millions (or billions?) of unlicensed copyrighted works. Authors and artists have sued the major AI platforms, and if those lawsuits are successful, they could dramatically hinder AI creation.
The response from most copyright scholars so far has been to argue that AI training should be protected as fair use. According to Sobel, though, this is simply a reflexive response to the habits learned during the last copyright wars—those against Disney and the recording and motion picture industries over expansive assertions of copyright scope and duration. Sobel refers to this approach of defending narrow copyright scope against expansive assertions by industry as copyright decelerationism.
Sobel warns, however, against continuing down this path in the new copyright war with generative AI. According to Sobel, even if the decelerationists’ fair use arguments win out, their victory will inevitably be a narrow one. AI “learning” will receive fair use protection, but real human learning will not. Sobel sees no reason why a decision allowing OpenAI to copy for purposes of learning would overturn cases like Texaco. He notes, “Could you imagine a defendant having the temerity to argue that she only torrented a Beatles album because she wanted to learn to write songs in the style of Lennon and McCartney? She’d be laughed out of court.”
Sobel predicts that a win for AI fair use wouldn’t benefit all forms of learning: “a rising tide for AI would not lift all boats! … [A] win for AI would be a win for AI—nothing more and nothing less.” He continues: “A copyright regime that makes these works available gratis to train AI but inhibits their circulation among human readers is a regime that misconceives ‘progress’ by defining it not as human engagement with expressive works but as a narrow sort of technological development.”
What’s to be done? Sobel’s response is that scholars who care about real human progress and creativity should completely invert their strategy. Instead of embracing deceleration or incremental minimalism, we should do the opposite: copyright accelerationism. We should insist on the rigid application of copyright doctrine to AI to heighten its contradictions and, effectively, make AI unworkable.
Embracing copyright law’s most regressive features—its trivial originality threshold, its lack of formalities, its extreme duration, and its limited fair use protections—would cast onto AI’s powerful overlords the task of reforming copyright law for everyone. Marc Andreessen, whose venture capital firm backs several AI startups, has argued that copyright liability “would upset at least a decade’s worth of investment-backed expectations” causing billions of dollars of harm. Sobel’s response: “Sounds like you might’ve made a bad bet. What would you be willing to do to salvage it?” A copyright accelerationist “would refuse to cede an inch on AI-specific exemptions from copyright liability and demand instead that the AI industrialists yoke their interests to those of the broader community of human readers and authors.”
To be clear, Sobel doesn’t adopt this position to kill AI. He explains, “The copyright accelerationist is agnostic towards AI; the future she seeks to realize is one that is of human self-development equally as accommodating through old-fashioned reading as it is of human self-development through the creation and use of expressive AI.” The goal is equal copyright treatment for human and robot readers and creators.
Now, Sobel’s essay is fairly light on the sorts of changes to copyright law that might accomplish these goals and on the political economy of how copyright accelerationists will bring their arguments to bear on AI’s backers. And Sobel admits these limits and recognizes that accelerationism is far from a sure bet. Nonetheless, to my mind, this is exactly the sort of work that our field needs—provocative shots across the bow of old-school minimalist thinking.
