I frequently ask ChatGPT to write poems in the style of the American modernist poet Hart Crane. It does an admirable job of delivering. But the other day, when I instructed the software to give the Crane treatment to a plate of ice-cream sandwiches, I felt bored before I even saw the answer. “The oozing cream, like time, escapes our grasp, / Each moment slipping with a silent gasp.” This was fine. It was competent. I read the poem, Slacked part of it to a colleague, and closed the window. Whatever.
A year and a half has passed since generative AI captured the public imagination and my own. For many months, the fees I paid to ChatGPT and Midjourney felt like money better spent than the cost of my Netflix subscription, even just for entertainment. I’d sit on the couch and generate cheeseburger kaiju while Bridgerton played, unwatched, before me. But now that time is over. The torpor that I felt in asking for Hart Crane’s ode to an ice-cream sandwich seemed to mark the end point of a brief, glorious phase in the history of technology. Generative AI appeared as if from nowhere, bringing magic, both light and dark. If the curtain on that show has now been drawn, it’s not because AI turned out to be a flop. Just the opposite: The tools that it enables have only slipped into the background, from where they will exert their greatest influence.
Looking back at my ChatGPT history, I used to ask for Hart Crane–ice-cream stuff all the time. An Emily Dickinson poem about Sizzler (“In Sizzler’s embrace, we find our space / Where simple joys and flavors interlace”). Edna St. Vincent Millay on Beverly Hills, 90210 (“In sun-kissed land where palm trees sway / Jeans of stone-wash in a bygone day”). Biz Markie and then Eazy-E verses about the (real!) Snoop Dogg cereal Frosted Drizzlerz. A blurb about Rainbow Brite in the style of the philosopher Jacques Derrida. I asked for these things, at first, just to see what each model was capable of doing, to explore how it worked. I found that AI had the uncanny ability to blend concepts both precisely and creatively.
Last autumn, I wrote in The Atlantic that, at its best, generative AI could be used as . I’d been using DALL-E to give a real-ish form to almost any notion that popped into my head. One weekend, I spent most of a family outing stealing moments to build out the fictional, 120-year history of a pear-flavored French soft drink called P’Poire. Then there was Trotter, a cigarette made by and for pigs. I’ve spent so many hours on these sideline pranks that the products now feel real to me. They are real, at least in the way that any fiction—Popeye, Harry Potter—can be real.
But slowly, invisibly, the work of really using AI took over. While researching , I asked ChatGPT to give me an overview of the U.S. market for beverages with this ingredient, but had to do my own research to confirm the facts. In the course of working out new programs of study for my university department, I had the software assess and devise possible names. Neither task produced a fraction of the delight that I’d once derived from just a single AI-generated phrase, “jeans of stone-wash.” But at least the latter gave me what I needed at the time: a workable mediocrity.
I still found some opportunities to supercharge my imagination, but those became less frequent over time. In their place, I assigned AI the mule-worthy burden of mere tasks. Faced with the question of which wait-listed students to admit into an , I used ChatGPT to apply the relevant and complicated criteria. (If a parent or my provost is reading this, I did not send any student’s actual name or personal data to OpenAI.) In need of a project website on short order, I had the service create one far more quickly than I could have by hand. When I wanted to analyze the full corpus of Wordle solutions for a on the New York Times games library, I asked for help from OpenAI’s Data Analyst. Nobody had promised me any of this, so having something that kind of worked felt like a gift.
The more imaginative uses of AI were always bound to buckle under this actual utility. A year ago, university professors like me were already fretting over the technology’s practical consequences, and we spent many weeks debating whether and how universities could control the use of large language models in assignments. Indeed, for students, generative AI seemed obviously and immediately productive: Right away, it could help them write and do . (Teachers to use it, too.) The applications seemed to grow and grow. In November, OpenAI CEO Sam Altman the ChatGPT service had 100 million weekly users. In January, the job-ratings website Glassdoor put out a survey finding that of professionals, including 77 percent of those in marketing, were using ChatGPT at work. And last month, Pew Research Center that almost half of American adults believe they interact with AI, in one form or another, several times a week at least.
The rapid adoption was in part a function of AI’s novelty—without initial interest, nothing can catch on. But that user growth could be sustained only by the technology’s transition into something unexciting. Inventions become important not when they offer a glimpse of some possible future—as, say, the right now—but when they’re able to recede into the background, to become mundane. Of course you have a smartphone. Of course you have a refrigerator, a television, a microwave, an automobile. These technologies are not—which is to say, they are no longer—delightful.
Not all inventions lose their shimmer right away, but the ones that change the world won’t take long to seem humdrum. I already miss the feeling of enchantment that came from making new Hart Crane poems or pear-soft-drink ad campaigns. I miss the joy of seeing any imaginable idea brought instantly to life. But whatever nostalgia one might have for the early days of ChatGPT and DALL-E will be no less fleeting in the end. First the magic fades, then the nostalgia. This is what happens to a technology that’s taking over. This is a measure of its power.