In the 74,833 words of a book I am writing, there are six words that, when strung together in a specific 12-word sequence, I cannot say. It’s a single line from the song Bloodbuzz Ohio by the National, which goes: “I still owe money to the money to the money I owe.”
My book is a memoir about the psychological toll that what I term “desperation capitalism” took on millennials in particular, and how it pushed tens of millions of people to try to find a way out of financial precarity by engaging in high-risk financial activity. It’s told through the lens of my own experience of falling deeper and deeper under the spell as I spent 11 months trading a few thousand dollars into more than $1.2m, and then 18 months chasing my losses all the way down to zero. Well, more than zero, in fact, since by the end I owed the US government nearly $100,000 in taxes on phantom gains that no longer existed.
Cue that line from the National as a perfect stage-setting epigraph – though only in theory. Song lyrics, my publisher informs me, are subject to notoriously strict copyright enforcement and the cost to buy the rights is often astronomical. Fat chance as well, then, of me quoting Eminem to talk about how Lose Yourself seeped into the psyche of a generation when he rapped: “You only get one shot, do not miss your chance to blow, this opportunity comes once in a lifetime.”
Oh would it be different if I were an AI company with a large language model (LLM), though. I could scrape from the complete discography of the National and Eminem, and the lyrics of every other song ever written. Then, when a user prompted something like, “write a rap in the style of Eminem about losing money, and draw inspiration from the National’s Bloodbuzz Ohio”, my word correlation program – with hundreds of millions of paying customers and a market capitalisation worth tens if not hundreds of billions of dollars – could answer:
“I still owe money to the money to the money I owe,
But I spit gold out my throat when I flow,
So go tell the bank they can take what they like
I already gave my soul to the mic.”
And that, according to rulings last month by the US courts, is somehow “fair use” and is perplexingly not copyright infringement at all, despite no royalties having been paid to anyone in the process.
I am neither a copyright lawyer nor a judge. I’m sure both have detailed and technical answers such as there are different legal frameworks around fair use and they apply differently to AI training versus direct reproduction. But suspend, if you will, all those technicalities for a moment. Is the spirit of copyright law – the same one that stops me from quoting a single 12-word line from the National in a book, but somehow permits ChatGPT to reproduce it word for word, as part of a new and instantaneously generated “song” – really being followed here?
Or is it simply power? Is it simply that Eight Mile Style, Eminem’s publishing company, could crush me for quoting Eminem in a book because I am small in comparison, or that Meta could drown Eight Mile Style, which is equally small in comparison, in teams of lawyers and years of delay?
(I should note that Eight Mile Style has filed its own copyright infringement lawsuit against Meta, in which it alleges “another case of a trillion (with a ‘T’) dollar company exploiting the creative efforts of musical artists for the obscene monetary benefit of its executives and shareholders without a licence and without regard to the rights of the owners of the intellectual property”. Hear, hear. I hope it wins.)
I imagine that a copyright expert would tell me that what ChatGPT did is akin to me scribbling lines from the National or Eminem in a personal notebook; that if I ever recorded and sold this AI-generated song quoting either one directly, then I would be guilty of infringing on protected material. But when it comes to LLMs, the output alone is the product.
LLMs are not “thinking” about that output. They are not learning and then transforming and being creative about anything whatsoever. They identify complex (and poorly understood) relationships between words and chunks of words within large amounts of text based on enormous amounts of training data (or what, to me, seems to be the wholesale heist of nearly the entire history of human literary and artistic output) to generate responses that are good enough to fool us into imagining that some type of consciousness was involved.
This is not a rant about me not being able to use a lyric in my future book. If that is the legal standard, then it is the standard (fair use doctrine for news media happens to be different). But let it at least be the same standard in essence all round.
There is what the law is, and then there is what the law ought to be. Are the rulings that have just been handed down to writers in the cases against Meta and Anthropic – another major player in the AI industry – ones that are beneficial to human creativity? Or are we stumbling into a world where not only does capital reign supreme, but where fake, for-profit “intelligences” face fewer restrictions on how they use human material than actual humans do?
We keep hearing from the AI founders and experts about how revolutionary their products are. About how AI and AI agents will disrupt absolutely everything in an irreversible way. Well then, perhaps AI should disrupt the law, too. If the law is producing situations that are technically correct but unfair, undesirable and threaten the existence of humanity’s writers, musicians and artists – life vocations that, for the most part, have never truly fallen within any sort of economic logic – then we must change the law.
Alexander Hurst is a Guardian Europe columnist
