Taylor Swift's TAS Rights Management filed trademark applications last week for two spoken phrases โ "Hey, it's Taylor Swift" and "Hey, it's Taylor" โ plus a stage photo showing Swift in a pink-guitar-and-iridescent-bodysuit setup, The Verge reported on April 28. Swift's team has not publicly said the filings are aimed at AI imitators, but the timing and pattern fit: Swift has been at the center of years of AI music imitation and a wave of sexualized AI deepfakes. Matthew McConaughey did the same thing earlier this year, trademarking video clips of his "alright, alright, alright" catchphrase explicitly to protect against AI misuse. The legal vehicle in both cases is trademark, not copyright.
The reason trademarks are showing up in this fight is what copyright doesn't cover. Copyright protects a specific song or recording โ a fixed artistic expression. It does not protect an artist's voice as such, the timbre or cadence with which they say a word, or their general likeness. AI voice cloning produces output that sounds like the artist without copying any specific copyrighted work, which puts most of it outside copyright's reach. Trademark protects a brand identifier โ a phrase, a logo, a recognizable signature. Once a phrase is trademarked in commerce, the owner can challenge not just identical reproductions but, as IP attorney Josh Gerben told The Verge, also "confusingly similar" imitations. That is the legal hook: a "Hey, it's Taylor Swift"-flavored AI promo for a competing product becomes trademark exposure, even if no Swift recording was sampled. Universal Music Group used a related angle to take down an AI-generated Drake song by citing the Metro Boomin producer tag at the start.
Two patterns matter. First, this is the IP system patching itself in real time. The legal infrastructure that protects creative work was built around fixed copies of fixed works. AI generates non-fixed approximations at scale, and the people who can afford lawyers are filing trademark workarounds. Expect more catchphrase trademarks from major artists and athletes through 2027, and expect a parallel tier where less-resourced creators get nothing. Second, the trademark hook is going to constrain AI product design. If your text-to-speech product can produce a clip that says "Hey, it's Taylor Swift" in something close to Swift's voice, that is now a trademark exposure before any specific song is involved. Voice-cloning startups, AI music tools, and content-generation platforms all need to start treating celebrity catchphrase-plus-voice combinations as a separate compliance category from copyrighted music.
For builders, three concrete things. First, if you are building voice synthesis, music generation, or any AI tool that outputs imitations of named individuals, the threat surface just expanded โ copyright filtering is no longer sufficient, and trademark databases (USPTO TESS in the US) need to be part of your safety pipeline. A "verboten phrases" filter against trademarked celebrity catchphrases is straightforward to build; not building it is going to get expensive. Second, the legal strategy is being explicitly tested โ watch the McConaughey trademarks granted earlier this year for what gets enforced and what doesn't. The first few infringement actions will set the precedent. Third, the IP system is moving to celebrity trademarks rather than fixing copyright for AI generation, which is the telling signal: there is no political appetite in 2026 to update copyright law for AI, so the system is patching at the edges. Builders who want certainty will not get it from Congress โ they will get it from trademark cases, takedown precedents, and platform policies, in that order.