ChatGPT Needs More Cowbell

AI struggles to write a good jingle.

You’d be forgiven if you can’t hum the 18th-century Cumbrian folk song “Do Ye Ken John Peel.” But in 1942, a version of that tune, reworked with lyrics about Pepsi-Cola, was the most recognized melody in America.

Three years earlier, two men walked into the office of Pepsi-Cola’s president, carrying a phonograph. They played a demo of what would become one of America’s earliest advertising jingles. To the tune of “Do Ye Ken John Peel,” it went: Pepsi-Cola hits the spot / Twelve full ounces, that’s a lot / Twice as much for a nickel, too. Pepsi-Cola is the drink for you. The jingle became a hit. People played it on jukeboxes around the country; it was translated into 55 languages. Electronic chimes atop a Pepsi-Cola plant in New York rang the first seven notes on the half hour. Other companies recognized the power of hummable commercials, and there were soon so many on the radio that listeners complained. Eventually, The New York Times’ radio station banned these “singing commercials,” but companies could get around the ban by launching instrumental versions instead.

America has had a love-hate relationship with jingles since, but they’ve continued to provide paychecks for musicians. Most of the music industry is made up of people in the gray area between “rock star” and “hobbyist,” like the session musicians and composers who make not just albums, but commercial soundtracks and jingles. When a report surfaced in October that OpenAI was developing a music-generation tool similar to products like Suno and Google’s Lyria, I wasn’t worried about the rock stars. These artists at least have their celebrity to trade on. But advertising musicians’ work is usually anonymous, and you don’t need to be Stravinsky to compose the 800-588-2300-EMPIRE tune. Are the jinglers going to be okay, or will advertising melodies be yet another livelihood cut down by new technology?

---

Jingle writing has become less of an art and more of a science over the past several decades. There are conventions now that did not exist in the “Do Ye Ken John Peel” era, and conventions could make it easier for the work to be automated.

For starters, jingles have gotten shorter and shorter over the years, morphing from 30-second songs to what are sometimes called “mnemonics,” or short sonic tags like the Netflix tudum and Liberty Mutual’s “Liberty, Liberty, Liiiberty, Liiiberty.” (Liberty Mutual’s CMO, Jenna Lebel, told me that they originally tested a version of the song that repeated liberty six times, but they felt that was a bit much and cut it down to four, which went over well.) Sonically, jingles—I just can’t get myself to call them mnemonics—are simple. Timothy D. Taylor, an ethnomusicologist at UCLA who wrote a history of advertising music, used to keep a database of every jingle he came across, labeled by genre. “Many of them were written like they were a song for children,” he said. “The very simplicity of them was part of the reason they stick in your head.”

With all of these limitations, jingle writers focus on measurable, proven qualities. “In many ways products like Suno think about music the same way that advertising agencies do—through words about mood and genre, and through references to existing artists,” Ravi Krishnaswami, a musicologist who studies AI music at Brown and a commercial-music composer himself, wrote to me in an email. Agencies might give composers a brief asking for a song that feels “curious,” “playful,” “witty,” and “quirky,” for instance—the types of descriptors you might see on a Spotify playlist—and then gauge which jingles best evoke these emotions in listeners. In theory, AI could be programmed to create a ditty with these qualities—perhaps producing something as catchy as “Like a good neighbor, State Farm is there” or “I am stuck on Band-Aid” (both of which were composed by Barry Manilow, by the way).

But good commercial music needs to target people’s emotions, and AI music is bad at doing that, at least for now. In one analysis by Stephen Arnold Music and SoundOut, a music-testing company, AI generally was not able to deliver tracks that matched the emotions requested by the prompters. Krishnaswami told me that, although AI can produce good jingles right now, “I’m not sure about ‘great’ yet.”

I could see what he meant. When I played around with one generator, it gave me nice-enough music for B-roll in a cat-food commercial. But could AI give me something as delightfully weird as the Meow Mix song? No. I asked for a jingle I could use for a coffee company, hoping it could give me something like the Folgers theme. Instead, it gave me “Brewed Bliss,” in which a store-brand Jason Mraz sang of “coffee magic in your soul.”

Maybe lyrics aren’t its thing, I thought. (We can’t all be Barry Manilow.) So I asked the generator for a series of five notes a coffee company could use as its sonic tag. It then gave me a 2-minute-27-second song called “Coffee Chime.” A soprano with Broadway-caliber diction narrated the song as it played, actually singing aloud the words “A five-note sequence that feels warm and inviting,” and “a rising pitch on a xylophone, ending with a soft chime that lingers, evoking the sound of a spoon stirring in a cup.” I tried to make it clear that I wanted something under five seconds that did not include self-narrating lyrics. But I just got another long song, more “sound of a spoon.”

Many of these jingles were ridiculous; many were deeply average. But the real problem was that I couldn’t remember any of them. A good jingle needs to stick with you—Liberty Mutual chose its melody because people could sing it back after hearing it only once, Lebel told me. Repetition is the main way a song can become an earworm, but it also needs to be the right amount of familiar and unusual, Bradley Vines, the director of neuroscience for the consumer-intelligence company NielsenIQ, told me. Memorable songs have a common melodic shape—think of how many pop songs’ melodies rise and fall in pitch, forming an arch (so does the tune of McDonald’s “I’m Lovin’ It”). But earworms usually have a surprising change in pitch, like that satisfying jump in the “Like a good neighbor” melody. “Given that a driver for earworms is the unexpected, an AI system trained on existing earworms may not be able to create a new one” by following what has worked before, Vines said.

The music I generated may have been forgettable because it was either too weird (no common melodic shape) or too bland (no surprising interval or rhythm changes). But research also indicates that people just don’t remember AI-generated content as well as the kind made by humans. Vines told me that when NielsenIQ showed people AI- and human-created advertisements, they couldn’t reliably distinguish between the two. However, when they looked at AI commercials, the memory systems in their brains were less active. “There was something that wasn’t as familiar and relatable in those ads,” he said. “There is something at the level of the nonconscious that still distinguishes” between human-created and AI-generated media.

Which isn’t to say that advertisers aren’t using the technology—they do, but mostly for demos, not the finished product. A composer who is not much of a singer might send a sample track with AI vocals on it. Or someone testing out whether to go with a band or a string quartet for a commercial track could easily try out both, before hiring people for the real advertisement. Part of this is pragmatic: Companies may be afraid of a lawsuit if, say, their prompt for “an upbeat 1960s British rock song with a 12-string Rickenbacker” produces a jingle that sounds too much like the Beatles. “I don’t see anybody going, Hey, AI, write me the next Coca-Cola jingle,” Jonathan Wolfert, the president of JAM Creative Productions, told me. (Ironically, Coca-Cola’s current holiday campaign features AI-generated visuals—but even so, the company has emphasized that the music in the ads was performed by humans.)

Whenever music generation gets good enough, companies will be ready. Liberty Mutual uses AI for voice, music, and sound effects in the early-stage commercials it shows focus groups. Then, it uses humans for the final versions. For now, people generally feel like AI-generated ads are inauthentic, but Lebel told me that Liberty Mutual has been testing them “with the hopes to someday incorporate that in market.”

---

Many people I spoke with said that a great jingle, or any great music, needs a “human touch.” When I’d ask them to define it, though, the consensus seemed to be, “I know it when I hear it.” But maybe it’s simple: AI is not a human. It does not exist in the sensory world, and it’s not exposed to the inspiration that daily life provides. Joni Mitchell wrote “Nathan La Franeer” about a real New York City cab driver who took her to the airport. Paul McCartney got the image for Eleanor Rigby’s “face that she keeps in a jar” from his mother’s cold cream. And the name Rigby, he saw on a sign while visiting Bristol.

Jingles are not Song to a Seagull or Revolver, but the best ones come from those same incalculable leaps of the creative mind. I thought of David Lucas. Now 88, Lucas was a prolific jingle writer from the ’60s through the ’90s, but by night, he was a rock musician. He was a producer on Blue Öyster Cult’s hit “(Don’t Fear) the Reaper,” and was the one who suggested the song needed a cowbell. (He thought nothing of it until decades later, when SNL did an entire sketch about it, and needs more cowbell entered the cultural lexicon.)

Lucas was once in a meeting about Fanta. He told me about passing bottles of it around with the other advertising guys, laughing, having a good time. As he got into a cab to head back to his studio, he thought, Well, that was fun. And then he got the lyric “It’s fun to be thirsty.” It’s a weird line—who likes to be thirsty? Not me. But after weeks of listening to jingles generated by AI, none of them stuck in my brain like Lucas’s Fanta jingle, which I had heard only a single time. And I doubt that AI—or most humans, for that matter—would have thought that “(Don’t Fear) the Reaper” needed more cowbell.