How AI-powered tools like OpenAI's Jukebox are being used to create deepfake music, which could lead to a host of ethical and intellectual property issues
'It's Christmas time! It's hot tub time!" sings Frank Sinatra. At least, it sounds like him. With an easy swing, cheery bonhomie …
Context & Ripple Effects
When OpenAI unveiled Jukebox in May 2020, the neural net could generate raw audio singing in an artist's style — enough to put a plausible 'new' Sinatra recording on anyone's hard drive. The Guardian's piece flags what that demo implied: voice cloning without consent or compensation, a question copyright analysts had already raised in 2019 over whether AI output has an author at all.
What looked like a research curiosity in 2020 became a product category within four years: Suno composes full songs with human vocals from text prompts, Udio launched a competing generator built by ex-DeepMind staff, and ElevenLabs extended synthetic voice tooling toward lyrics and samples. By early 2026, some labels had stopped fighting and started signing, taking licensing deals that mirror their YouTube revenue splits.
First-order effects
- Artists whose styles are cloned — the Guardian uses Sinatra as the test case — have no contractual claim on recordings they never made, leaving estates and rightsholders to police synthetic performances after publication rather than before.
- OpenAI's decision to publish a model that reproduces named artist styles shifts enforcement costs onto rights-holders, who must detect and challenge imitations across platforms rather than license upfront.
Second-order effects
- Suno, Udio, and ElevenLabs turned Jukebox's demo-grade cloning into consumer products with text-prompt interfaces, forcing labels to respond to mass-scale imitation rather than isolated experiments.
- Labels now face a fork between litigation over training data and voice likeness versus monetizing it through licensing deals — a split visible in the industry by 2026, when some signed revenue-share agreements modeled on user-generated content.
Third-order effects
- If the YouTube-style revenue-split pattern holds, voice and style become licensable assets with standardized payouts, converting what was framed in 2019-2020 as an authorship crisis into a negotiated market — while artists outside major-label deals risk having no mechanism to opt in or out.
- The unresolved question of whether AI output itself carries authorship remains the structural gap: settlement deals can price likeness, but they do not settle who owns the generated recording.
The trend: Generative music is moving from research demos like Jukebox to a licensed commercial layer, with labels deciding whether to sue the cloners or split revenue with them.