“Shut the doors, please.” The director of a video production studio in London is wearing headphones and holding a clipboard as he prepares for our first take. He looks me up and down.
“The avatar is going to be wearing whatever you’re wearing,” he says. “You’re okay to have it like this?”
“Yep,” I nod, eager and ready.
He takes another look. “If you can take off the watch that’d be nice.”
Oh. I set it on the table.
“Thank you.”
As if we’re in Hollywood, an engineer in jeans snaps a clapper board. Behind me is a green screen. Above me are large lights radiating heat. In front is a camera with a teleprompter, the sort used by TV news anchors, and it’s showing some of the strangest dialogue I’ve ever seen.
I read the single sentence in front of me: “All the boys ate a fish.”
At the director’s instructions, I roll my eyes from top to bottom and left to right, before reading a one-minute script with cringey sentences like, “You can see the positivity shine through as I’m friendly and warm as I speak.” I’m asked to read it several times.
The goal is to train an AI model to create a digital clone of myself. The “fish” statement covers all possible mouth movements, and the rest is just enough to train the model on a cheerful version of my face, body and gestures. The company creating my avatar, Synthesia, has made similar versions for more than 15,000 companies including McDonald’s, Accenture and Amazon. Want to make a marketing video? There’s a buffet of more than 150 avatars to choose from that speak in more than 120 languages, all based on real humans. There’s no need to rent studio space, cameras or lighting — just type a script and your avatar will say it. One manufacturer says they’ve saved 70% in video production costs with the method.
Deepfakes used to be a scourge of the internet. Now they’re a legitimate tool for getting a human on video more cheaply and at scale. Instead of hiring an actor to present a corporate training video and paying for their travel and time, a company can use an avatar for a fraction of the price. You can expect to see more digital clones like these in the coming years, of celebrities in TV ads or of ageing bands like ABBA in whizzy new concerts. The proprietor of one virtual concert company tells me he’s been solicited by the families of several dead singers, eager to cash in on their abiding fame with clones that could revive them on the stage. Speculating on his avatar’s ability to act in the afterlife, movie megastar Tom Hanks recently half-joked, “If I wanted to, I could get together and pitch a series of seven movies that would star me in them in which I would be 32 years old from now until kingdom come.”
Synthesia says its videos are technically not deepfakes since they are generated from scratch, and deepfakes manipulate a pre-existing video of someone. But the spread of these avatars is surely one of the most head-spinning impacts of the rapid advances in so-called generative AI, which can now conjure artwork, mimic voices in music, clone entire faces and bodies, generate pop songs, screenplays, short stories and news articles. It hammers a virtual nail in the coffin of the creative process we know today, accelerating a transformation that is seeing humans outsource the work of their imagination and even their own likenesses.
Like those before it, this technological revolution will come at a price. The introduction of the printing press in the 16th and 17th centuries allowed us to spread ideas and literacy, but it also lost us the oral tradition of storytelling. In our own century, social media connected us but also inadvertently made us more disconnected. In the case of generative AI, there’s a chance it will gradually erode our creative skills and lead to soulless, increasingly derivative content on our computer screens, phones and TVs, a cribbed amalgamation of the content on which AI models have been trained. Art and writing will increasingly become “content.” We’ll see a lot more of what the recently deceased novelist Martin Amis decried as “herd writing,” clichéd phrases like “the heat was stifling.” What’s even more likely is that we’ll lose the experience of connecting with a human artist through their work.
Notwithstanding those potential consequences, a new generation of creators is understandably keen to exploit generative AI’s commercial potential and jumping into the fray. Consider Sydney Faith, an author who self-publishes young adult fiction on Amazon. In January she used ChatGPT to write Legends of the Shadow Woods, a collection of short stories based on Greek mythology.
She left most of the writing in the hands of the AI, asking it at the start, “What kind of fantasy world should I write about?” When the software gave several suggestions, she picked one, then got it to generate chapter outlines, and then to generate three paragraphs at a time, which she then copied and pasted into a document. It was a repetitive process, she says, but eventually she had a rough draft of each short story. The book, which discloses that it was written by ChatGPT, now has a four-star review on Amazon. “Passable,” says one reviewer, who says they were expecting worse. Faith says she’s sold a few dozen copies so far.
Those pitching generative AI to the creative classes argue that it makes getting content down on the page so much easier. Natalie Monbiot, who runs Hour One, a company that makes avatars like mine, says these new AI tools can solve “blank page syndrome.” She refers me to one of her clients, an entrepreneur named Ian Beacraft, who has been producing videos where an avatar of himself reads out technology news. He’s been using ChatGPT to write his scripts, making it quicker and easier to produce his videos. The blank page used to be a sticking point, Monbiot says, before invoking Silicon Valley lingo: These new tools are “removing friction” from the whole process.
Devin Finley is another artist who’s been using generative AI to replace himself on the screen. A New York-based actor, Finley has put his baritone voice to characters in video games and audio plays for about 10 years, but in the fall of 2022, an industry friend told him that Hour One was looking for actors who would “go virtual” by creating avatars of themselves, digital clones that could be automated to work on their behalf. The gig could be lucrative, the friend said.
Finley welcomed the chance to make some passive income and created a clone of himself, becoming one of the first of more than 150 other actors to have a deepfake made by Hour One. Many of them weren’t professional actors, but students, waiters, paralegals and a range of gig workers, according to Monbiot. Finley, who now has “multilingual synthetic actor” on his resume, earned an upfront fee for his time, and a few months after creating his avatar, it starred in the marketing videos of a media company, Monbiot said, declining to reveal the client’s name.
If these few examples point to a broader adoption of artificial “artistic” content, a mimicry of the real thing, then we can’t blame machines for everything. Humans have been laying the groundwork for this shift over the past decade or so, with Hollywood churning out sequels and re-imagined classics, formulaic TV shows like NCIS, and a Netflix diet of addictive television designed to keep you watching the next episode. News sites have become aggregators of other articles. Much of the content we see on screens today, in other words, is already a rehash of someone else’s work.
For now, most AI-generated video content is being pioneered on YouTube, but down the line, a television producer could use generative AI tools like ChatGPT to create a rough draft of a script that then gets polished by humans. They could, for instance, take a human-written rough draft and use ChatGPT to repurpose it in the style of Nora Ephron, Aaron Sorkin or other storied screenwriters. That prospect has given Hollywood executives unexpected leverage against screenwriters who are currently striking over financial security. One of the demands of the Writers Guild of America has been for studios to ban the use of AI to write scripts, shifting a job with a steady salary to a form of gig work. So far, the studios have rejected those requests.
What that likely means is that our TV consumption will have fewer human stories, and more AI-sourced derivative content that helps studios and streaming companies protect their bottom lines. Some of the most notable recent films have come from the personal experiences of their creators, from Steven Spielberg’s childhood in The Fabelmans, to Daniel Kwan’s Asian-American upbringing in Everything Everywhere All at Once. Perhaps movies like this, borne out of real-world, personal experience will narrow in reach to niche audiences, while the general public will lap up artificially designed concoctions of previous films that sold lots of tickets.
In that sense, we’ll lose more of the ingredients that make artwork great. The story behind how a movie or piece of art is created is critical to how we end up judging it as good or bad, says Agustín Fuentes, a Princeton University anthropologist and author of The Creative Spark: How Imagination Made Humans Exceptional. A work of art has value because of the process of creating it — the care and thought and even the errors contribute to its beauty. “Think of the Mona Lisa,” says Fuentes. “I’m not impressed by it, but it is hugely important in the history of art, and seeing it matters because of the story behind its making and its historical context. None of that is replicable by AI. An AI model can make a perfect image copy of the Mona Lisa, but it cannot produce the Mona Lisa.”
In March, the New York-based poet and novelist Joseph Fasano tweeted a letter from a schoolteacher, asking if he’d take part in an experiment: “Would you be willing to come into our classroom and go head to head with ChatGPT: human poet versus AI poet?” Fasano and the chatbot would each have five minutes to come up with a poem for three different topics. The problem with this is that ChatGPT can generate poems in seconds because it’s been trained on poems that humans have spent countless hours on. What it spits out isn’t poetry; it’s content.
Rick Rubin, the famed record producer who worked with music artists from Johnny Cash to Run-D.M.C., was recently asked in a podcast hosted by Bloomberg Opinion columnist Tyler Cowen about generative AI’s impact on art. His response was that for now, art from generative AI was mostly “decorative,” because it lacked humanity. “It’s the soul in it that makes it good,” he said.
What do soul, humanity and historical context mean anyway? It is hard even for the experts to define creativity, just as it is hard to put into words exactly what we will lose when AI-made content takes up more of what we read and see. But that loss will probably have something to do with the invisible force of human connection to which people like Rubin allude, and the act of feeling “moved” by created work. Twenty years ago, when I got my first reporting job at a local radio station, I learned an open secret among news readers: The best way to get people to perk up and listen to a news bulletin was not for the news reader to deepen their voice or to copy the mannerisms of other news readers, but simply to pay careful attention to the meaning of each word as they said it. The difference between reading with meaning and reading mindlessly was technically indistinguishable and difficult to articulate. But it worked. To this day, I only have to pay close attention to the words in a book I’m reading to my kids, and they’re enraptured.
Does that allude to the soul that Rubin described? It’s hard to know. But machines are not sentient, and they have no relatable struggle or backstory to move us when we encounter their created work. As AI researchers make generative AI models more sophisticated, with billions more parameters and datasets to draw from, art created by AI will probably also appear more inventive. That will only reinforce how little we know about what creativity is, and even erode our sense of exceptionalism among animals and machines. One of the reasons Sam Altman came to believe in the possibility of artificial super intelligence before co-founding OpenAI was his realization that if human intelligence could be simulated, humans weren’t all that unique to begin with.
Little wonder that AI scientists wrangle over whether AI is, or ever will be, creative. Melanie Mitchell, a computer science professor at Portland State University, says in her 2020 book, Artificial Intelligence: A Guide for Thinking Humans, that computers can be creative but they’re not quite there yet. When I emailed her in May 2023 to ask if she still believed that, she said “yes.”
Demis Hassabis disagrees. The British scientist who leads Google’s AI efforts was claiming eight years ago that AlphaGo, the Go-playing AI program developed by his unit DeepMind, was displaying remarkable signs of creativity. In its 2016 match against Lee Sedol, the program made a highly unconventional play known as “move 37,” surprising both Lee and the game’s commentators. The move, in which AlphaGo sacrificed a group of stones in the corner of the board to gain a positional advantage elsewhere, was so unexpected that experts thought it was a mistake, and Lee took an unprecedented 15 minutes to consider his response.
AlphaGo went on to win the game decisively. Hassabis and many other AI scientists put that down to the system’s creative prowess, posing a tantalizing possibility, that software had managed to create something totally unique from seemingly nothing.
It’s also possible that believing that makes us suckers. For Hassabis, Altman and other entrepreneurs who want us to buy into a vision of unfathomable smart AI systems, selling a story of software whose abilities are as unpredictable and mysterious as humans makes good business sense. In April 2023 for instance, Google CEO Sundar Pichai appeared on an episode of 60 Minutes to talk about Google’s ChatGPT competitor, known as Bard, where he mentioned a phenomenon he called “emergent properties.”
“Some AI systems are teaching themselves skills that they weren’t expected to have,” Pichai said on the program, explaining that one of the company’s AI models was able to translate Bengali even though it had only ever seen a few words of Bengali. Yet the company’s own research paper showed its leading AI model had been trained on Bengali. The system wasn’t being creative or intuitive. Its creators were exaggerating its capabilities.
I am not downplaying the awe-inspiring potential of these machines. I’m also keen to avoid grasping for vague justifications for our exceptionalism as humans. With the exception of those who can’t wait for the Singularity to happen, we all want to distinguish ourselves from AI systems that seem to be rapidly on course to surpass us. But as more people use generative AI to write books and screenplays or conjure videos that could go viral on TikTok — a phenomenon that is undoubtedly coming thanks simply to the economics — it’s hard to deny that our creative skillsets will grow flabbier, and another avenue for human connection will narrow. The mysterious AI models with their “emergent properties” will seep further into our creative fields, working their magic while reinforcing the power and influence of the technology companies that created them.
For now, my digital doppelganger is still sitting on some Synthesia server somewhere, waiting to be dusted off and used for a presentation, a TikTok video or a message to someone. As strange as it may sound, some staff members at the consultancy EY, previously known as Ernst & Young, have started sending talking avatars of themselves to clients instead of emails. I have yet to find a good reason to use it other than to promote this story, though.
The actor Devin Finley, meanwhile, seems to have settled into a new career pattern, where his digital clone can earn a little extra money on the side. He recalls that when he was first asked about creating an avatar of himself, he hesitated. “Originally I thought this might be something that could take away from who I am,” he tells me. “Then I realized I am a unique, living being.”
In a future world of artificial content, that might well become a novelty.
Parmy Olson is a Bloomberg Opinion columnist covering technology. A former reporter for the Wall Street Journal and Forbes, she is author of “We Are Anonymous.”
Send questions/comments to the editors.
