Sign up for the daily CJR newsletter.
On major news sites, you may have started noticing a button—maybe a triangular “play” symbol, or a little headphones icon—allowing you to hear an article read aloud. Tap the button, and you may be told that the piece is being “read by an automated voice.” News outlets including the New York Times, the Washington Post, and the Wall Street Journal and magazines such as The New Yorker, The Atlantic, and Vanity Fair have all embraced automated text-to-speech tools to oralize their journalism in recent years. ElevenLabs, an AI voice company that partnered with The New Yorker in 2023 to provide this function, said this year that it had surpassed five hundred million dollars in annual recurring revenue. ElevenLabs offers more than ten thousand artificial voices, from James (“deep, raspy, and grim”) to Ngan (“cute, bubbly, and authentic”).
“It didn’t sound like a robot; it didn’t sound like something from Star Wars or Lost in Space or something like that. It sounded pretty damn good, and it’s nearly instantaneous in its production,” David Remnick, the editor of The New Yorker, told Semafor late last year, about an ElevenLabs voice. Monica Racic, the deputy editorial director at The New Yorker—where about 20 percent of subscribers now listen to narrated stories—told me that it’s about “giving all readers greater flexibility in how they want to consume our journalism—whether they’re reading or listening while commuting, walking the dog, or doing the dishes.” ElevenLabs has held early talks with investors about a secondary offering that would value it at twenty-two billion dollars.
But the quiet proliferation of AI-read news, thanks to tools created by ElevenLabs, tech giants like Microsoft, and other specialist companies such as News over Audio, or NOA—which promises “news and insight for busy people”—has made me curious: Does hearing journalism narrated by a synthetic voice change a reader’s relationship to the story?
As I’ve listened to a bunch of synthetic narrators in recent weeks, I have been impressed by the tools’ human mimicry and time-saving potential. I absorbed a profile of Gianni Infantino, the FIFA president, while loading the dishwasher. I listened to a story on inequality as I took a walk in the sun. I heard a bunch of daily news articles while sitting on the sofa. But all the while, I never quite forgot I was listening to a robot.
The speech patterns of the AI narrators tend to contain markers of artificiality. The voices can be breathless, not sensing when to slow down, to linger over a line or a word for impact. They emphasize strange things (one voice kept stressing the word and). Quirks and tics—such as the simulated drawing of breath or rising intonation, added to sound more personable—arrive at strange times, giving them an uncanny quality. Despite all the advances, there is still an unsettling feeling that comes with hearing journalism read out by AI.
Advocates for the technology would probably tell me that voice narration is only going to improve, that these lumps will get smoothed out. I’m sure that’s right. To listen to synthetic speech today is nothing like encountering the robotic voices of the early internet—such as the tinny vocals of Fred, an Apple text-to-speech tool that Radiohead used on their 1997 track “Fitter Happier.” Nearly thirty years later, AI-generated voices—emerging from neural networks trained on thousands of hours of recorded human speech—are capable of all the rasps, inhalations, and emphases that you might hear on an NPR broadcast, an audiobook narration, or an unscripted podcast. ElevenLabs did not respond to an interview request for this article. But Mati Staniszewski, the chief executive, told Al Jazeera in February that the company was able to create “the first model that could sound human,” able to speak “based on the contextual understanding of what was written.”
The more I listened to AI voices from ElevenLabs and other providers, however, the more I came to realize what was bothering me: the absence of human discernment. I listened to one article—a classic 1997 New Yorker profile of Donald Trump, then a struggling real estate mogul, by Mark Singer, who died in June—first read by the author, then read by a synthetic voice. (This was possible because Singer reads out the article in the audiobook version of his 2016 book Trump and Me.) I found the real narrator preferable. It wasn’t that the AI version was unlistenable. My complaint was that it could not capture Singer’s clever irony, or the pathetic grandiosity in his tone when he quotes Trump, or the lavish emptiness when he gets to those Dickensian lists of Trump Tower ornaments. As the article crescendoed—and reached the biting passage that Trump “had aspired to and achieved the ultimate luxury, an existence unmolested by the rumbling of a soul”—the AI voice continued with consistent pacing and assertiveness, missing the pathos, flattening the tinge of sadness audible in Singer as he contemplates a man who sees an interior life as “an intolerable inconvenience.” To listen to this human-sculpted work read by a nonhuman mediator was to lose something essential.
To be sure, there are clear benefits to voice-narrated news, especially for people with disabilities—blindness or low vision. “We are always looking for ways to make our journalism more accessible and have been experimenting with automated voice for a number of years,” Graham James, a spokesperson for the Times, told me. (James declined to tell me which firm the Times was working with.) ElevenLabs, through its impact program, says it provides free licenses to a range of nonprofits and to individuals with accessibility needs. Racic, of The New Yorker, said that “while the technology certainly isn’t perfect, we’re continuing to make improvements with the goal of making our journalism easier to access and engage with. If anything, the biggest change is that readers are engaging with us more.”
A spokesperson for the Journal, which partnered with Microsoft for its text-to-speech function beginning in 2020, told me that about 92 percent of users responded positively to the feature during a test. In the past year, the Journal’s “Read to Me” option was used about five million times, with a 65 percent completion rate. “I really see this as additive. We haven’t taken anything away,” Taneth Evans, the head of digital, told me in an interview. “If we were saying you have to listen to this now, and you can’t read it, I’d understand objections.” The plan is to add more choice for how people engage with the newspaper’s work, Evans said. “People told us that they wanted to consume our journalism while also commuting, or cooking, or exercising.”
Some formats work better than others; I preferred listening to the Journal’s hard-news stories read by an AI narrator, for instance, to a nonfiction magazine piece from The New Yorker. (It’s notable that The New Yorker does not use AI to narrate its fiction, which is read by the author whenever possible, because this is more like a “performance,” Racic said.) But if they’re properly labeled as AI, “I don’t think the voices themselves fundamentally change readers’ relationship to our journalism,” Racic told me. “The reporting and the story remain the same.”
The urge to listen, I think, stems from the era we inhabit: a world without enough time. AI narration allows us to buy back time from our day. But I think we would do well to stop and reflect on whether the world being built around us—spinning at 1.5x speed or faster—is one we actually want to live in. Singer noticed a similar desire to collapse time in his Trump profile. I was sitting in my living room when I heard Singer read that, during a flight to Mar-a-Lago, Trump put on Bloodsport, a 1988 martial arts movie, and assigned his son Eric, then thirteen, “the task of fast-forwarding through all the plot exposition.” Trump’s goal, he said, was “to get this two-hour movie down to forty-five minutes.”
Has America ever needed a media defender more than now? Help us by joining CJR today.
