Sign up for The Media Today, CJR’s daily newsletter.
Every year around this time, Spotify launches “Wrapped,” a feature that presents users with data about their listening habits for the year. The feature was introduced in 2016 and quickly became a popular way for listeners to share their unique music tastes with friends. (Forbes described it as a BuzzFeed quiz for music lovers.) Each year, Spotify has kept the releases fresh by adding fun and personalized elements, such as last year’s “Sound Town,” which identified a part of the world that is aligned with a listener’s most-streamed artists. Wrapped coming out is now a highly anticipated event. Some diehard fans jokingly compare it to the Super Bowl or the Met Gala.
This year, however, cute new elements were largely missing; instead, Spotify users got the usual statistics about their habits and a podcast that was “delivered using generative AI by two dynamic hosts,” in collaboration with Google. The two “hosts” banter about the user’s most listened-to artists, songs, and musical genres in an eerie, humanlike back-and-forth. While technologically impressive, the result fell flat for many users. Some noted on social media that the hosts mispronounced artists’ names, got basic facts wrong, and spoke with a slightly sarcastic tone. “It’s nothing more than two uncanny voices repeating data points and descriptors their models generated from tracking my data ceaselessly,” one user complained. “It feels both like listening to a doctor go through your bloodwork results and a psychic vaguely supposing facts about your life,” Vox’s Kyndall Cunningham wrote.
The new AI-generated podcast was built using NotebookLM, a research tool launched by Google last year to help users make sense of complex information. Powered by the company’s AI chatbot Gemini, NotebookLM prompts users to upload source material (say, an academic paper or YouTube video) and then becomes an “expert” on it. Previously, users could only interact with NotebookLM via text, but in September, Google added Audio Overview—the technology that this year’s Spotify Wrapped podcast relies on—to help users who learn best by listening. The results stand out from other AI-generated audio, partly because they’re peppered with intonation and linguistic disfluencies like “um” and “like.” As Stuart Anderson-Davis, my colleague at the Tow Center for Digital Journalism, pointed out in a recent piece for CJR, “the casual listener would be unlikely to notice much difference between this AI-generated output and the NPR broadcasts on which the tool seems to have been trained.”
While Audio Overview may be an ingenious way to make a dry academic paper come to life, it has not been a universally well-received replacement for the creative process imbued in Spotify Wrapped. Some commentators blamed layoffs at Spotify for prompting the company to rely too heavily on AI-generated content. (A year ago, the company laid off 17 percent of its workforce, citing “dramatically” slowed growth; shortly thereafter, it raised the cost of many of its subscription plans.) In short, for some, Wrapped no longer feels like Spotify’s number one concern. “This year’s recap,” Brady Brickner-Wood wrote in The New Yorker, “revealed a company that seems chiefly concerned with profit margins and squashing its competition.”
While NotebookLM’s audio feature didn’t thrill users on Spotify, it has been largely successful outside of the music-streaming sphere as a writing and research assistant, including in the world of journalism. Jensen Huang, the CEO of Nvidia, an AI chipmaker, told the tech magazine CNET that he “used the living daylights” out of the AI tool to scan through and listen to documents. Chris Moran, head of editorial innovation at The Guardian, wrote in a post on LinkedIn that NotebookLM is a great tool for journalists to keep reports in one place and share them with collaborators. “If you’ve written a draft of an article based on a report and want to double check your citations, you could upload your draft and ask it to find supporting evidence in the original documents,” Moran wrote. Audio Overview, Moran argued, is the least impressive part of the NotebookLM package.
And, if placed in the wrong hands, AI-generated podcasts in general could be misused as propaganda weapons—in Anderson-Davis’s CJR piece, he explored some of the risks associated with providing free access to a tool that produces “ultra-realistic podcasts in minutes.” In a series of experiments, he demonstrated how a politically motivated actor could use Audio Overview to produce misleading content; after uploading conservative articles to NotebookLM and tweaking the prompt he entered, he managed to produce a ten-minute podcast about Matt Gaetz, Donald Trump’s then-pick for attorney general, that was highly favorable and skipped right over the sexual misconduct allegations against him. (Gaetz denies wrongdoing; he recently withdrew from contention to be attorney general.) Since Spotify doesn’t prohibit AI-generated content—unless it violates copyright laws—the Tow Center was able to upload the episode directly to Spotify within fifteen minutes. “It seems inevitable that AI-generated podcasts will become an increasingly prominent feature of digital communications in the years to come,” Anderson-Davis concluded—“flooding the public sphere with yet more pseudo-journalistic pink slime, slop, and disinformation.”
As with most AI tools, NotebookLM’s Audio Overview is a double-edged sword. It can be a helpful tool for those genuinely seeking to learn dry information, while also easily being misused by those aiming to deceive and mislead. The social media backlash against Spotify appears to show a clear boundary for the extent to which technology companies can integrate AI-powered features into the creative process while maintaining the trust of their users. Putting Spotify Wrapped into the mouths of two soulless podcast hosts was that line. Users’ skepticism toward the use of AI in this low-stakes setting may offer a glimmer of hope for our susceptibility to nastier stuff. And perhaps it will push companies to think twice before further incorporating AI features into their products.
Other notable stories:
- This morning, the US Press Freedom Tracker is out with a new report finding that at least forty-eight journalists have been arrested or detained by police on US soil this year—a higher figure than the previous two years combined, and the third-highest annual total since the Tracker started counting, in 2017. Nearly 90 percent of the arrests came at protests related to the war in Gaza, and nearly half came in New York. “The Tracker has consistently found that protests are the most dangerous location for journalists in the United States: Four out of every five arrests (as well as assaults) of journalists since 2017 took place at demonstrations,” Stephanie Sugars writes. In October, CJR’s Feven Merid wrote about the crackdown particularly on visual journalists at protests, citing a previous round of data from the Press Freedom Tracker.
- Yesterday, Donald Trump said that he would appoint Kari Lake—the former local news anchor turned 2020 election denier and failed Republican candidate for governor and US Senate in Arizona—to run Voice of America, the US-backed international broadcaster. VOA is supposed to be editorially independent, but Trump was widely accused of attempting to politicize it in his first term, and his appointment of Lake would seem to reflect the same impulse. As CNN’s Brian Stelter notes, however, it’s “not clear at all” that Trump will be able to impose Lake: the current director, Mike Abramowitz, has only been in post for six months, and, since Trump was last in office, Congress created a Senate-approved bipartisan board to oversee appointments at VOA and sister networks.
- Nexstar—the media company that owns the NewsNation cable network and The Hill, among other properties—said that it would make cuts to its broadcast and sales divisions; less than 2 percent of staffers will be affected, but that could still equate to hundreds of people. In other media-jobs news, Stat, the health news site that grew significantly during the pandemic, laid off eleven staffers—a surprise given that its owner, which also owns the Boston Globe, is generally profitable. (The Globe’s Aidan Ryan has more.) And the union representing tech workers at the New York Times—who recently went on strike amid stalled contract talks—reached a tentative deal with management.
- Donald Tusk, the prime minister of Poland, said that his government would list TVN and Polsat, the country’s two biggest private TV stations, as “strategic companies”—a move that would put them on a footing with, for example, state-owned energy firms, and would block them from being sold to what the government deems to be hostile actors. Tusk’s announcement came amid rumors that TVN, which is currently owned by the US media company Warner Bros. Discovery, could be sold to a businessman close to Viktor Orbán, Hungary’s far-right leader. Notes from Poland has more (and we profiled TVN last year).
- And for CJR, Aluf Benn, the editor in chief of the left-leaning Israeli newspaper Haaretz, writes that his journalists will not be cowed by the Israeli government’s recent move to boycott the paper, including by pulling paid ads—ostensibly in response to controversial remarks that its publisher made recently about “Palestinian freedom fighters.” Haaretz “will not be silenced,” Benn writes. “The boycott led to a surge of new subscriptions, and we launched a legal battle against the advertising ban. In a democracy, the government should not be allowed to treat its budget as a punishment tool against its critics.”
Has America ever needed a media defender more than now? Help us by joining CJR today.
