TECHNOLOGY

The New Generation of A.I. Apps Could Make Writers and Artists Obsolete

Sudowrite and Dall-E use machine learning to produce astonishingly serviceable text and images in a flash. What does that mean for stodgy old human beings?
Image may contain Human and Person
Illustration by Quinton McMillan. Images by Getty and Shutterstock. 

For decades we’ve been warned that artificial intelligence is coming for our jobs. Sci-fi books and movies going all the way back to Kurt Vonnegut’s Player Piano portray a world where workers have been replaced by machines (or in some instances, just one machine). More recently, these ideas have moved from the annals of novels into the predictive economic papers of governments and consulting firms. In 2016, the Obama administration authored a report warning that the robots were coming, and that millions of Americans could soon be out of a job. In 2021, McKinsey predicted that algorithms and androids would vaporize 45 million jobs by 2030. And the Brookings Institute prophesied in a 2019 study that 52 million U.S. jobs would be affected by algorithms by 2030. 

While no one can agree on exactly when the robots will take over, or how many jobs they will swallow up, the assumption has generally been that garbage collectors, bus drivers, and interstate truckers will be among the first to lose their livelihoods to A.I. Lately, however, it’s starting to look as if people like me—creatives—are even more imminently in danger. Over the past few months, new advancements in A.I. have made it clear that writers, illustrators, photographers, journalists, and novelists could soon be driven from the workforce and replaced by high-tech player pianos.

Don’t believe me? Just read the opening of this short story:

It was late when I first heard the loud clicking noise coming from outside. As I looked out of my bedroom window, the tall grass swayed in an unseen breeze. And then, a shadow passed over it—which is when I saw it. A creature standing just outside the window, staring right at me. Its eyes were deep red with a venomous glow; they seemed to burn with a fire that made me shudder. The creature’s body was black as coal, with a large, thin tail near its rear. It hissed at me and tapped its long, sharp claws impatiently on a tree trunk. I grabbed the nearest weapon I could find—a baseball bat—and headed outside. The creature’s breath reeked of sulfur, and its hushed breathing carried an echo of a crackling campfire. When it saw me, its feet hit the soil with thundering booms as if a skyscraper had crashed to the earth with each footfall. It lunged forward at me; its claws slashing wildly in the air.

That paragraph wasn’t written by an MFA, or a sci-fi author, but rather by a new online machine-learning platform called Sudowrite, which is billed as a tool to help with the creative-writing process. For the above paragraph, I wrote the first sentence—about the loud clicking noise—and the A.I. wrote the rest. The technology works using a platform from OpenAI, an artificial-intelligence research laboratory with a billion dollars in funding from Microsoft, and investments from Elon Musk. More specifically, it’s built on GPT-3, a component of the company that focuses on text. GPT-3 scrapes billions of words and learns from them using natural-language processing. Then, it gets to work. After reading Sudowrite’s first draft, I said I wanted more description. So the A.I. suggested we add some “smells,” and revised the text accordingly in a few seconds. While the story isn’t Pulitzer Prize–worthy (yet), I was startled by the algorithm’s ability to turn phrases like “the tall grass swayed in an unseen breeze” and “its hushed breathing carried an echo of a crackling campfire.”

Advancements in writing are just the beginning. Another OpenAI tool currently being likened to magic by people in Silicon Valley is a visual platform called Dall-E. Using a version of GPT-3, Dall-E can create truly astounding renditions of artworks and illustrations. Like GPT-3, Dall-E has learned how to draw and paint by combing through billions of images. It’s now conversant in styles, objects, shapes—you name it. Just type a set of commands into Dall-E, and it will nearly instantaneously generate an image to illustrate them. For example, if you ask it to draw “an astronaut riding a horse in a photorealistic style,” it will create several options to choose from. If you tell it to instead make a “pencil drawing,” it will render new images in that style. You can order up stained glass, spray paint, Play-Doh, cave drawings, or paintings in the style of Monet. You can replace the astronaut with a teddy bear. A dog. Elon Musk. Or have the horse riding a horse. The possibilities are endless, and the end results are terrifyingly impressive—so much so that one of the top questions associated with a Google search of the platform is “Is Dall-E fake?”

These new technologies have scrambled our assumptions about creativity and computers. Philosophers have long believed that computers would never be able to create “art” because machines don’t have emotions. They don’t feel pain or joy; therefore, they can’t express those feelings in a creative way. As it turns out, however, computers don’t need feelings to make art. They can simply copy what humans have already made. “It’s not so much that the computer is ‘thinking’ like a human artist,” said Hod Lipson, a scientist who specializes in artificial intelligence and robotics at Columbia University. “It’s more that they are producing output based on what they have seen.”

Now, I have to be honest. I didn’t interview Hod Lipson. I didn’t even search the web for his quote. Instead, I asked another GPT-3-powered program called ShortlyAI to write a post for me about algorithms replacing journalists and artists. The algorithm found Lipson’s work, read his research papers, videos, talks, and interviews, and selected this quote as the most compelling to add to this piece. When I reached out to Lipson to see if this was something he said, and that the A.I. hadn’t just made it up, Lipson said it was accurate, but noted that a more accurate point would be to say that “it’s more that, like a human, the computer too is producing output based on what it has experienced.” 

Here’s what else the A.I. wrote:

Robots, machine learning, and other A.G.I.s (artificial general intelligence) will soon be writing daily news stories about stocks, sports scores, weather-related events and a variety of breaking news alerts. In the next decade or less, artificial intelligence will make huge strides in their capability to produce high-quality written news. This will bring an end to the era of the journalist as we know it. The difference between A.G.I.s and journalists is the ability to comprehend and absorb information, not just to report it. An A.G.I. will be able to understand the “what” of a story, as well as its “why,” without breaking a sweat over details, because it simply doesn’t care about them in terms of story-writing.

See, now you’re wondering which parts of this column I wrote, and which ones were written by a computer. Amit Gupta, the cofounder of Sudowrite, believes that these writing platforms won’t completely replace humans. Instead, this new tech will supplement our all-too-human efforts, not unlike the way the predictive text on an iPhone does. “Financial reporting and sports reporting will be the first to be sent to automation because they are a little bit more rote,” Gupta said, but he thinks that most journalism will be produced in conjunction with, rather than completely by, A.I.—at least for the foreseeable future. “Auto-complete is already kind of working its way into Google Docs and Gmail; that type of A.I. writing is going to be ubiquitous, so as I’m writing a story, it will say, ‘Oh, here’s a potential quote.’” (Yes, I did actually interview Gupta for this piece.)

GPT-3 is far from perfect. It can be wonky, repetitive, even racist. As TechCrunch noted in 2020, GPT-3 language models have associated the word Islam with terrorism, and female pronouns with the word naughty. That’s because it’s crawling text that has been written by humans in the past. These platforms are learning not from other A.I. entities but from us—which has its upsides from a creative standpoint, and its downsides from a humans-are-jerks perspective. Dall-E’s own documentation warns that words like personal assistant and flight attendant will generate images of women, and words like CEO will likely give you white men.

Those issues may give you and me pause, but they aren’t doing much to slow the technology’s evolution. Last week, Google announced its own text-to-image platform, called Imagen, which is just as impressive (and scary) as Dall-E. If anything, Imagen’s photorealistic images are even more indistinguishable from the real thing.

For now, the companies behind these image platforms aren’t making them available to the public—mostly while they work out the kinks, but also, as Google has noted, while they work on ways to prevent the software from being used for nefarious purposes. “Datasets of this nature often reflect social stereotypes, oppressive viewpoints, and derogatory, or otherwise harmful, associations to marginalized identity groups,” Google said in its latest announcement about its software. “[This] guides our decision to not release Imagen for public use without further safeguards in place.”

Once these services are publicly available, all bets will be off. Think about the speed with which fake news obliterated trust in the journalism landscape online during the 2016 election cycle. An A.I. that can create fake images indistinguishable from real ones could have disastrous effects on the public’s ability to separate fact from fiction. I don’t even want to imagine what kind of horrifying imagery the QAnon folks could dream up, let alone the denizens of 4Chan and other bowels of the internet. I’m sure Russia would have a field day with this tech too. As with all new technologies, there will be good and bad results. The question is, will the creators of these platforms be able to discover the negatives before this tech is released into the public? (Which, let’s be honest, has never happened before in technological history.) Perhaps it will take years to foresee all the potential outcomes of these services. There’s a fun irony in the thought that the only thing saving our jobs from being replaced by algorithms is the fear that humans will do terrible things with the technology.

Some early experiments point to positive and creative applications. GPT-3 was recently used to create a new song with lyrics written and rapped by an A.I. Eminem—it isn’t half bad. There’s also an A.I.-written play now being performed in London that supposedly asks what technology can teach us about ourselves. (A notice to ticket buyers warns ominously that the play “may contain strong language, homophobia, racism, sexism, ableism, and references to sex and violence.”) One group, called Calamity A.I., is using GPT-3 to write screenplays that are then filmed with actors—as you can see from “Lady Bird: A Lost Scene Written by Artificial Intelligence,” the A.I. isn’t quite at the level of Greta Gerwig. But the same software did come up with a relatively delicious chocolate-chip-cookie recipe. There are other iterations of OpenAI technology used to generate music from scratch, like Jukebox, which can write its own music, including singing and instruments in numerous genres and styles.

As someone who has been covering new technologies for more than 20 years, I’ve seen a lot of contraptions, apps, sites, platforms, and newfangled inventions, and few have ever felt like they would change the way we live. Twitter, the iPhone, and driverless cars have been among the handful of standouts. The more I see of these new machine-learning algorithms, though, the more I realize that the future is coming quickly. And in that future, a good number of people could lose their jobs—not to mention their grasp of what originated in the mind of a human or a machine. Soon, you’ll be asking yourself every time you read an article, Did a human write this, or did an algorithm? The answer is: You’ll never know.