- 28% of the new music uploaded to Deezer every day is 100% AI. 70% of the plays on those tracks are fraudulent. ~$2–3B in royalties diverted a year. And on Aug 2, 2026, ignoring AI labeling becomes illegal in the EU. The clock is under two months. 🧵
- A year ago the question was "is generative audio safe to use." That collapsed in 6 weeks. Oct 30: UMG settles with Udio. Nov 25: WMG partners with Suno. Both pivoting to licensed, opt-in models. The legal panic is fading. Three harder problems replaced it.
- Portability. The licensed Suno/Udio outputs live in a walled garden — most tiers can't download or export off-platform. A media company needs an asset that ships across broadcast, streaming, social, cinema, and in-game. Walled-garden audio is broken before it starts.
- Registrability. The US Copyright Office's position: prompt-only AI outputs aren't copyrightable. Your AI jingle has zero downstream IP protection — a competitor can lift it free. Meanwhile the AI-training liability floor just got set: Anthropic settled at ~$3,000 per work.
- Detectability. EU AI Act Article 50 hits Aug 2, 2026: synthetic audio must carry machine-readable marking detectable as AI. Penalties run up to €15M or 3% of global turnover. Metadata alone won't cut it — the draft Code wants watermarking too.
- Here's the trap: there's no single watermark to adopt. SynthID-Audio is Google-only, closed detection. AudioSeal is open (MIT) but speech-first, weaker on music. XAttnMark holds 68% detection under adversarial attacks where AudioSeal drops to 15%. You can't bet on one.
- And watermarks die in the real content journey. Most social platforms strip C2PA metadata on upload. A mark that survives MP3 can die in Opus transcoding. Radio captured by a phone mic in a car loses anything that can't survive the analog gap. Survival has to be tested.
- April 2023: a fake Drake/Weeknd track hit 20M views before UMG pulled it. In 2026 it takes 90 minutes. The real question isn't "is it fake" — it's how fast you take it down and notify artist, label, DSPs and platforms in parallel. Detection is one input; the rest is workflow.
- The quiet liability: ad agencies. Suno's Pro/Premier plans don't include indemnification. 4A's tells agencies to negotiate AI indemnity clauses with clients — most haven't. One AI jingle in a national spot triggering a rights claim, and the agency eats the cost.
- We don't build the watermark — Google and Meta did. The gap is integration. A label runs 6–12 disconnected systems, and Spotify + 15 labels just backed a DDEX AI disclosure standard the ERN 4.3 spec has no fields for. The plumbing doesn't exist. We build the auditable chain.
- Real question for rights-tech and trust & safety leads: when a watermark dies in Opus transcoding or the analog gap, who in your org gets the alert — and what's your takedown SLA once it does? Most teams don't have a clean answer. What's yours? #MusicTech
- We wrote up the full Article 50 readiness picture — the stack, the deadlines, and where a consultancy actually fits: https://veriprajna.com/solutions/ai-audio-licensing-provenance