Magic and sacrifice in generative AI
I am thinking today about the unimaginable amount of labor and resources that goes into producing today’s generative AI products—the emotional, creative, and physical labor that goes into producing the art that “AI art” products are trained on, the underpaid and traumatizing labor that goes into labeling and supervising ChatGPT, the astronomical energy cost of running the GPUs. All of that is done to keep up the façade of instant, cheap, free, and “magical” automation.
Is that it, the key to magic? The magician Teller famously says, “Sometimes magic is just someone spending more time on something than anyone else might reasonably expect.” Time, money, resources—magic is unreasonable sacrifice. I am reminded of The Prestige, a film about two rival magicians and the lengths they go to in order to produce their illusions: bodily mutilation, murder, torture.
Alfred Borden: It takes nothing to steal another man’s work.
Robert Angier: It takes everything.
In the case of AI, there is an added layer of horror: the sacrifice is not being made by those who profit from the show, but rather from millions of people who are in many cases directly being harmed by that show.
Wherever the sacrifice comes from, to the audience it must be invisible for the trick to work. We press the button and we get an image or an essay without ever having to see the input data to the algorithm—and so, we think it is magic. But seeing magic where there is no magic can be dangerous. I wish we could offer the world a peek behind the curtain, a chance to comprehend the true scale of what goes into that illusion of intelligence. Think of Dorothy and her friends looking behind the curtain to find the Wizard of Oz, who ultimately turns out to be a puppeteer and ventriloquist with far too much time on his hands.
“I thought Oz was a great Head,” said Dorothy.
“And I thought Oz was a lovely Lady,” said the Scarecrow.
“And I thought Oz was a terrible Beast,” said the Tin Woodman.
“And I thought Oz was a Ball of Fire,” exclaimed the Lion.
“No, you are all wrong,” said the little man meekly. “I have been making believe.”
To “make believe”—to make us believe—that is the great trick of today’s AI products, and I worry that we will fall for it over and over again.