A few weeks ago, Clarkesworld Magazine, a science fiction and fantasy literature journal that publishes short stories, posted an article about a flood of AI-generated content that skyrocketed in February following the release of ChatGPT: “Before the end In 2022 it was mostly plagiarism, now it’s machine-generated submissions,” the magazine’s Twitter account said.
Clarke wrote on his blog that he reached out to editors from other magazines who confirmed that this isn’t unique to Clarkesworld and appears to be a general trend.
Apparently, the launch of ChatGPT created a whole spam scene. Reuters reports hundreds of e-books on Amazon that list ChatGPT as author or co-author, including many books about ChatGPT written by ChatGPT. Reader submissions for the magazine have been closed for the time being.
What is happening today at Clarkesworld is happening tomorrow with all cultural distribution systems dealing with a new form of synthetic cultural spam. Even when we digitally watermark synthetic content, manual (and presumably machine) editing can easily destroy it. Open-source AI detector GPTzero recently published a case study that found a 50 percent false positive rate, and while these approaches are improving, I’m not sure watermarking and AI detection will ever be effective may be enough to stem a rising tide of synthetic spam in which individual works are impossible to distinguish from synthetic cultural products. The Voight-Kampff machine for literature has not yet been invented.
A few months earlier I had published an article in my newsletter in which I argued that humans would not be overwhelmed by a deluge of synthetic products, and that the machine-made works (so far) had had very little impact because they simply lacked the emotional punch is missing in order to be able to convince as art. A flood of AI spam does not contradict this: the human psyche as a system is far more complex than any company and its bureaucratic mechanisms. People aren’t easily fooled by ChatGPT’s smooth and middle-of-the-road stories, but I doubt whether the same is true for systems like GEMA or VGWort — and the figures from the Kindle store give me a clue Right: “Admittedly, consumer interest has been drowsy so far: (…) sales amounted to about a dozen copies.”
The basic model of institutions like GEMA or BMI is: You create music, publish it and report it to GEMA and if your music is played by a DJ, you collect. So what happens when that DJ is replaced with a generative AI playing synthetically generated music? A new song, but in the style of Nirvana, produced using a database full of statistical analyzes of the music of Kurt Cobain and a million other artists? This question is extremely tricky, if not impossible: If generative AI provides a calculated latent space in which users can use prompt engineering to interpolate between songs, who does GEMA want to pay out and how much? How does GEMA calculate how much actual Nirvana content is actually contained in a song in the style of “Nirvana, Bongo Boys and MC Hammer”?
Generative AI means the dissolution of the human archive of cultural expression into an explorable latent space. I can train a model on the structure of something, for example “The folding of protein molecules”, and the AI will calculate a multidimensional space from the previous research on protein folding with every possible solution that I can explore by choosing parameters. The same applies to music, art, pictures, film, literature, business coaching, 101 marketing tricks to increase your productivity, real estate offers, job applications, basically any form of communication that uses language in the form of text, audio and (moving) images.
How can BMI or GEMA or any other copyright exploitation organization secure income for creatives when a stochastic, interpolatable library exists that contains the entire history of human knowledge and provides an interpolative exploration of all combinatorial possibilities?
The dissolution of the human archive of cultural expression into explorable latent spaces also means the dissolution of the GEMA and VGWort licensing model, which relies on individually identifiable works, since this identification becomes blurred in an interpolatable latent space. The flood of AI spam from Clarkesworld SciFi Magazine is one of the first signs of this.