A landmark week for AI music policy: the NO FAKES Act cleared the Senate committee, The Atlantic exposed 21 million songs in AI training datasets, and the Suno copyright case barrels toward a July ruling after forensic discovery revealed millions of copyrighted recordings in its training data.
NO FAKES Act clears Senate committee with $750K platform penalty
The Senate Judiciary Committee unanimously advanced the NO FAKES Act on Thursday, sending the most significant US AI voice-protection bill toward a full Senate vote. The legislation gives individuals the legal right to control how their voice and likeness are used in AI-generated content, with penalties up to $750,000 per violation for platforms that fail to remove unauthorized deepfakes. The right transfers to heirs for up to 70 years after death. First Amendment carve-outs protect news, parody, education, and research.
If you're a voice actor, vocalist, or session musician, this is the bill to track. Three Republican senators flagged First Amendment concerns but still voted yes. It needs to pass the full Senate and House before becoming law, but the unanimous committee vote signals strong bipartisan support. The music industry — labels, publishers, and the AFM alike — backed it heavily.
The Atlantic publishes searchable database of 21 million songs used for AI training
Atlantic reporter Alex Reisner published four searchable databases revealing over 21 million copyrighted tracks used to train AI music models. The largest, LAION-DISCO-12M, contains over 12 million tracks. A second holds 9 million, with two smaller collections of roughly 100,000 each. Artists ranging from Taylor Swift and Bad Bunny to Aphex Twin, Radiohead, and Hainbach appear in the results. Google and Stability AI have both confirmed using these datasets in published research papers.
This is the first practical tool that lets individual artists and small labels check whether their recordings were ingested by AI systems — no lawyers required. Built as part of The Atlantic's AI Watchdog project, it converts abstract copyright claims into searchable, concrete evidence. If you release music and haven't checked yet, you should.
Suno discovery reveals millions of copyrighted recordings; pivotal July ruling nears
Audio fingerprinting by Audible Magic — conducted during two weeks of forensic inspection at Suno's counsel's office — identified millions of recordings owned by Universal and Sony in Suno's training data. The labels have filed to expand their original 560-work complaint to over 61,000 specific recordings, calling even that "only a small fraction" of what was found. A summary-judgment hearing is scheduled for July before Chief Judge Saylor in Massachusetts.
This is the case that will define whether training AI music models on copyrighted recordings constitutes fair use. Sony is the last major label still litigating against both Suno and Udio — Universal settled with Udio in October 2025, Warner settled with Suno in November 2025. Whatever the court decides will set the legal framework for every AI music company's training practices.
Musicians union sues UMG and Warner over AI settlement payouts
The American Federation of Musicians filed a breach-of-contract suit against Universal and Warner in New York federal court, alleging the labels' settlements with Suno and Udio triggered "new uses" provisions in their collective bargaining agreement. Under these provisions, labels must compensate session musicians when their recorded performances are put to new commercial applications. The AFM says the labels created significant new revenue from AI licensing deals but have refused to pay the musicians whose performances were used to train the models.
This matters beyond the specific lawsuit. If the AFM wins, every future AI training deal between labels and AI companies would require revenue-sharing with session players. That would meaningfully change the economics of licensed AI music generation and could set a precedent for other unions in adjacent creative industries.
Lionel Richie files to trademark his voice against AI cloning
Lionel Richie filed four trademark applications at the USPTO for signature sung phrases including "Hello, is it me you're looking for?" and "Easy Like Sunday Morning." The filings follow Taylor Swift's earlier attempts and signal a growing celebrity strategy: using trademark law — not just copyright — to fight AI voice cloning. Trademark attorney Josh Gerben notes the applications face hurdles because the USPTO requires sound marks to function as "source identifiers" tied to a product or service, not just famous lyrics.
Suno raises $400M at $5.4B valuation amid mounting legal pressure
Suno closed a $400M Series D led by Bond Capital, more than doubling its November 2025 valuation of $2.45B. The company reports over 2 million paying subscribers and $300M in annual recurring revenue. The raise landed alongside the discovery revelations, the "Say No To Suno" aerial protest over a Santa Monica investor summit, and the AFM lawsuit — a collision of soaring commercial metrics and intensifying legal and reputational risk.
The investors are betting that AI music generation survives its legal reckoning largely intact. With Warner already settled and licensing, that bet looks plausible. But Sony is still litigating, the GEMA case in Germany was just pushed to July 31, and session musicians are now demanding their cut. The $5.4B number prices in a lot of optimism.
LALAL.AI updated its DAW plugin to run six-stem separation entirely locally — vocals, drums, bass, acoustic guitar, piano, and electric guitar — without sending audio to the cloud. Insert it on any track, pick the stem you want to extract or remove, and it processes in-session. No artificial limits on track count or file length for Pro subscribers, and it works in any VST3-compatible DAW including Ableton Live, Reaper, and Audacity.
The offline processing is the real differentiator. Most competing stem separators either require cloud round-trips or run as standalone apps outside your DAW. Having it as a VST3 insert that processes locally means you can use it on client sessions without worrying about uploading proprietary material to someone else's servers. AU support is still in beta.
Lionel Richie's trademark filings, following Taylor Swift's earlier attempts, point to an emerging legal strategy that bypasses copyright entirely. The theory: if a sung phrase is recognizable enough to function as a "source identifier," then AI-generated imitations could constitute trademark infringement — regardless of whether the AI output technically copies any copyrighted recording. Copyright asks whether training data was used fairly; trademark asks whether the output misleads consumers about who's behind it.
For voice actors and vocalists, this could become a more direct path to protection than the complex copyright battles currently winding through federal courts. The NO FAKES Act, if signed into law, would create an even more explicit statutory right. But trademark registration, unlike legislation, doesn't require Congressional action — it just requires the USPTO to agree that a voice functions as a brand identifier. Watch whether Richie's applications survive examination. If they do, expect a wave of similar filings from artists with distinctive vocal signatures.