Can AI make a hit song?

Episode Summary

Executive Summary: The episode examines how AI is already affecting music creation, from songwriting and stem separation to voice modeling, mixing, and mastering. Through a hands-on experiment making a song entirely with AI tools, the hosts and musicians Charlie Harding and Ian Kimmel find that current AI is most useful for small workflow tasks, not for fully replacing human creativity. They argue the tech is promising but still janky, and its biggest impact may be in the music industry’s business, credit, and rights systems rather than in making better songs.

Main Topics: Hands-on AI music experiment (Priority: 5/5): The hosts create a song using multiple AI tools—ChatGPT, Suno, AudioShake, Splice, Arcade, VoiceSwap, Magenta, and iZotope—to test what AI can actually do in a real music workflow. AI as workflow automation, not creativity replacement (Priority: 5/5): Charlie and Ian emphasize that AI is most useful for tedious prep work, email writing, label/stem organization, and basic ideation, but not for replacing skilled music-making or good taste. Limits of current AI music quality (Priority: 5/5): The generated outputs are often generic, cliché, lo-fi, or unusable, and the process frequently slows the musicians down more than it helps. Industry data, metadata, and scale problems (Priority: 4/5): Music is a niche industry with messy metadata and small companies, making it harder for AI systems to train well or solve real production problems compared with larger tech domains. Copyright, licensing, and automated enforcement (Priority: 5/5): The discussion turns to how AI may intensify already complicated music-rights norms, with concerns about automated claims, licensing deals, and black-box content ID systems. Historical parallels to auto-tune and drum machines (Priority: 4/5): The guests compare AI to earlier technologies like auto-tune and drum machines: initially stigmatized, later normalized, and often repurposed creatively by outsiders. Voice identity and creative comfort (Priority: 3/5): The episode touches on the weirdness of hearing AI versions of one’s singing voice and how voice manipulation can be useful for artists who dislike their own vocals.

Key Arguments: Current AI music tools are best at narrow, repetitive tasks—like stem labeling, beat generation, and email drafting—rather than full artistic production. The song-making demo showed that AI can generate ideas, but humans still needed to fix structure, lyrics, melody, arrangement, and taste. Most AI music output today is closer to generic background music or “C+ content” than compelling, release-ready songs. Music industry metadata is already messy, which limits training data quality and makes sophisticated AI solutions harder to build. The biggest future risk is not just AI replacing musicians, but AI-driven rights enforcement and automated payments/claims becoming overly aggressive and error-prone. Creative norms in music are becoming more corporate and credit-heavy, which may reduce experimentation and complicate provenance. AI may eventually matter more as a behind-the-scenes layer inside existing tools than as a standalone replacement for musicians.

Data Points: Podcast series structure: Second episode in a three-part series - The episode is framed as part of The Vergecast’s broader AI series. Time spent recording voice samples: About an hour - David Pierce recorded his singing voice for Voiceify.ai testing. Song-building workflow time: About 2.5 hours - Ian estimates the team spent roughly two and a half hours collecting AI-generated material before arranging the track. Stem-labeling use case: 400–500 tracks - Charlie cites Jacob Collier sessions as an example of huge unlabeled projects that take a long time to prepare. Writing session split: Four equal splits - Ian describes modern songwriting sessions where four people may split credit equally even if contribution levels differ. Platform adoption: 3.4 million companies - Zapier ad copy cites the number of companies already automating with its platform. LinkedIn network size: Over 1 billion professionals and 130 million decision makers - LinkedIn Ads ad copy positions the platform’s B2B targeting scale.

Pivotal Quotes: "the only tool I've really used to assist in any part of the creative process has been ChatGPT, but even that is quite limited." — Charlie Harding: Charlie explains that his AI use is minimal and mostly utilitarian. "AI is a cannon of C plus content." — David Pierce: David summarizes his view that AI can produce lots of mediocre material quickly, which still has business consequences. "I think the biggest problem was that we used that piano, which came from what sounded like a pharma ad." — Ian Kimmel: Ian reflects on how poor source material pushed the demo song toward a generic, commercial sound.

Implications: AI will likely seep into music as a background utility for prep, editing, and rights management rather than as a replacement for artists. But if automated licensing and content ID expand without care, they could make music creation and distribution more restrictive, error-prone, and corporate.