First criminal sentence in AI streaming fraud, Italy opens an investigation into Suno's terms of service, and Google makes its AI watermark detector available to the public — a week that tested AI music's legal boundaries from three directions.
First AI streaming fraud sentence: 18 months for an $8M bot scheme
A federal judge in Manhattan sentenced Michael Smith, 54, of North Carolina to 18 months in prison on October 6 for using AI-generated songs and thousands of bot accounts to collect over $8 million in fraudulent streaming royalties. Between 2017 and 2024, Smith uploaded hundreds of thousands of AI-generated tracks to Spotify, Apple Music, Amazon Music, and YouTube Music, then used up to 10,000 bot accounts with VPN masking to inflate play counts past anti-fraud systems.
Prosecutors wanted 46 months. The judge called the guidelines "somewhat inflated" for a non-violent offense. Smith must also forfeit $8.1 million and serve two years of supervised release. If you run a label or distributor, this is the Department of Justice signaling that AI-powered fraud is now a criminal matter, not just a platform policy issue. Expect streaming services to tighten bot detection further.
Italy opens investigation into Suno over moral rights and unfair terms
Italy's competition authority (AGCM) opened a formal investigation into Suno on October 6, citing potentially unfair terms of service under Article 33 of Italy's Consumer Code. The core issue: Suno's license requires users to waive moral rights — the right to attribution and to object to alteration of your work — which Italian copyright law does not permit. The AGCM also flagged unilateral price changes, broad account suspension powers, liability limits, and a mandatory US arbitration clause.
This is not a ruling yet, and no penalty has been announced. The authority will launch a public consultation in the coming weeks. But it matters because moral rights are protected across most of the EU, and a finding here could force Suno to rewrite its terms for all European users. If you create music with Suno and care about being credited, watch this space.
Google opens SynthID Detector to the public
Google launched a free, public web portal at synthid.com on October 7 that lets anyone check images, video, and audio for SynthID watermarks. The company says it has now watermarked over 240,000 years' worth of AI-generated audio — up from 60,000 years in May. Detection currently covers content from Google, OpenAI, NVIDIA, and Kakao, with Apple listed as coming soon.
For audio specifically, SynthID marks output from Google's Lyria music model and NotebookLM's podcast feature. The watermark is inaudible and survives MP3 compression and speed changes. The obvious limitation: it only detects SynthID marks. A file from Suno, Udio, or any other non-participating platform returns no signal, and a negative result proves nothing. Still, this is the first free, universal tool for checking AI audio provenance, and it sets a baseline for what detection infrastructure looks like.
Sonilo raises $11M to turn video into soundtrack
Sonilo, a San Francisco startup founded by former TikTok AI and music staff, closed an $11 million seed round led by B Capital with Redpoint participating. The company builds a "Sound World Model" that analyzes video content and generates matching music and sound effects. Shutterstock is an early licensing partner, and Sonilo says it's finalizing publishing and master recording deals with major rights organizations.
A consumer app is planned for Q4. The pitch is clear: automated scoring for short-form video, aimed at creators who need music that fits their content but can't afford or find the right sync license. If you license production music, this is a competitor. If you make it, this is potentially a new distribution channel — assuming the licensing deals hold up.
Fortnite pulls Ironmouse album cover after fans spot Gemini watermark
A small star-shaped Google Gemini watermark was discovered hiding among the decorative stars on the "Club Ironmouse" album cover, used for a Fortnite Jam Track since September 19. Fans spotted it on October 2 and tagged Ironmouse — a VTuber known for opposing AI-generated content — who demanded the art be replaced. By October 5, the cover was swapped out on Fortnite, Spotify, and Apple Music. Trap Nation, the label behind the release, admitted AI had been used and apologized.
The lesson here is prosaic but important: AI-generated assets are leaking into commercial releases through outsourced artwork, and visible watermarks are the canary. Google recently added an option to turn off visible Gemini watermarks, but invisible metadata markers remain. If you commission cover art or visual assets, check for watermarks before release.
Third Circuit: using copyrighted material for AI training is not fair use
The Third Circuit Court of Appeals ruled on September 29 — reported widely in early October — that ROSS Intelligence infringed Thomson Reuters' copyrights by using Westlaw headnotes to train its legal search AI. The court found the use was "highly commercial and only minimally transformative" and identified harm to both existing markets and an emerging licensing market for training data. It is the first federal appellate decision on fair use and AI training.
This is not a music case — it's about legal research summaries, and the opinion explicitly distinguishes itself from generative AI. But it matters because it establishes that copying copyrighted works for AI training can fail the fair use test at the appellate level, which gives plaintiffs in the UMG/Sony v. Suno and similar cases a persuasive precedent. The music-specific rulings are still coming, but the legal terrain just shifted.
Suno launched Speech in public beta on October 1 — its first product beyond music generation. Speech creates spoken voice and background music together as a single track. Simple mode works from a short prompt; Advanced mode accepts a custom script with controls for voice gender, speech style, and variation. Clips can run up to about eight minutes. A toggle turns off the music for voice-only output.
The practical applications are obvious: podcast intros, narrated content with underscore, audio ads, voiceover demos. Quality is uneven in the beta — expect inconsistent accents and dramatic pauses that nobody asked for. But the core idea of generating voice and music as one cohesive piece, rather than layering them separately, is a genuinely new workflow. Worth testing if you produce any kind of spoken-word audio with music beds.
Three stories this week — SynthID going public, the Gemini watermark catching an outsourced album cover, and Google quietly marking 240,000 years of audio — point to a single trend: AI audio watermarking is becoming real infrastructure before regulators have decided what to require. Google, OpenAI, and NVIDIA are embedding marks. Detection tools are now free. But participation is voluntary, coverage is partial, and a clean scan proves nothing.
The gap between what's watermarked and what isn't is where the interesting problems live. Suno and Udio don't use SynthID. Most open-source models don't watermark at all. A world where some AI audio is marked and some isn't may be worse than no watermarking, because it creates a false sense of certainty. For audio professionals, the practical takeaway: learn what SynthID Detector can and can't tell you now, before clients start asking you to run checks.