Google faces lawsuit over training Lyria on YouTube data

🔑 Enhanced Key Takeaways

•The lawsuit specifically alleges that Google's Lyria 3 was trained on a massive corpus of "at least 44 million clips and 280,000 hours of music" from YouTube.
•Google's motion to dismiss asserts that YouTube's Terms of Service grant a "worldwide, non-exclusive, royalty-free, sublicensable and transferable license" to use uploaded content, including for derivative works, which extends to AI training by Google and its affiliates.
•Beyond copyright infringement, the plaintiffs allege that Google's AI tools, including Lyria 3 and ProducerAI, remove copyright management information (CMI) and falsely attribute ownership to end-users.
•The legal action seeks class certification and damages for 16 alleged violations under federal and state law, encompassing claims like fraud, illegal distribution, false endorsement, and false advertising.
•Google counters that the musicians cannot plausibly claim their voices are distinctive enough to function as trademarks, nor that Lyria outputs are mistaken for their work, and dismisses Biometric Information Privacy Act (BIPA) claims as speculative due to alleged metadata stripping during training.

📊 Competitor Analysis▸ Show

Feature/Aspect	Google Lyria 3/Pro	Suno AI	Udio	AIVA
Max Song Length	Up to 3 minutes (Pro), 30 seconds (Clip)	Up to 8 minutes (v4.5)	Variable, often shorter clips, focus on remixing	Longer, full compositions
Structural Control	Explicit control over intros, verses, choruses, bridges, outros (Pro)	Song Editor for sections, less explicit structural prompting	Granular control, instrumental separation	Deep structural understanding (orchestral/classical)
Vocals/Lyrics	Generates vocals and timed lyrics, supports user-provided lyrics	Realistic and multilingual vocals, auto-generated lyrics	Realistic vocals, strong for vocal song generation	Primarily instrumental, some vocal capabilities
Multimodal Input	Text, audio, and image inputs	Text prompts, audio upload as starting point	Text prompts	Text prompts, musical parameters
Watermarking	SynthID watermarking, C2PA support	Not explicitly mentioned in detail, focus on licensing	Not explicitly mentioned in detail, focus on licensing	Not explicitly mentioned in detail
Licensing/Copyright	Google cites YouTube ToS for training; outputs embedded with SynthID	Paid plans grant commercial rights, ongoing lawsuits	Cleanest licensing (deals with UMG, WMG), ongoing Sony lawsuit	Full copyright ownership on Pro plans
General Pricing	Free in Gemini (with limits), higher limits for paid subscribers/Pro	Free tier (limited), Pro plan ~$10/month	Free tier (limited), paid plans available	Free basic, Standard $15/month, Professional $49/month
Overall Quality	High-fidelity, structural coherence, strong overall	Best overall for complete songs with vocals	Strong for vocal song generation, granular control	Dominates orchestral/classical music

🛠️ Technical Deep Dive

Architecture: Lyria 3 utilizes latent diffusion, applied to temporal audio latents.
Training Data: Trained on audio data, which is annotated with text captions at different levels of detail. The training data undergoes processing steps including deduplication, safety filtering, and quality filtering to enhance reliability and compliance.
Hardware: Lyria 3 models are trained using Google's Tensor Processing Units (TPUs), which are specialized hardware designed for large-scale computation in training and serving generative AI models.
Model Variants:
- Lyria 3 Pro: Google's premier model for full-length song generation, capable of creating tracks up to approximately three minutes long with professional-grade structural awareness, including intros, verses, choruses, bridges, and outros.
- Lyria 3 Clip: Optimized for speed and high-volume requests, this variant generates high-quality 30-second audio clips.
- Lyria RealTime: A distinct model designed for interactive, real-time music generation, allowing users to continuously shape and morph music by blending styles, instruments, and musical attributes with advanced controls over key, tempo, density, and brightness.
Input/Output: The models support multimodal inputs, including text prompts, audio, and images, to generate high-fidelity 48kHz stereo audio. Outputs can include realistic vocals and time-aligned lyrics.
Safety & Transparency: All Lyria 3 and Lyria 3 Pro outputs are embedded with imperceptible SynthID watermarking technology, allowing for the detection of AI-generated or edited audio. The models also support C2PA standards for content origin and authenticity.

🔮 Future ImplicationsAI analysis grounded in cited sources

The lawsuit's outcome could redefine the scope of "fair use" in AI training.

Courts are still establishing a clear framework for whether AI training constitutes copyright infringement, and this case, along with others, will contribute to that legal precedent.

Content platforms like YouTube may face increased pressure to implement more explicit and granular creator consent mechanisms for AI training.

The lawsuit highlights creator concerns about unauthorized use, and YouTube has already introduced a "third-party training" toggle, though its effectiveness for past training is unclear.

AI music generators will increasingly pursue direct licensing deals with rights holders to mitigate legal risks.

Competitors like Udio have already settled with major labels to build licensed AI music platforms, indicating a trend towards legal compliance to avoid similar lawsuits.

⏳ Timeline

2023-11

Google DeepMind introduces Lyria, collaborating with YouTube Shorts for the Dream Track feature.

2024-12

YouTube implements a "third-party training" toggle, allowing creators to opt-in their content for AI model training, with default opt-out.

2025-06

Google DeepMind launches the Lyria RealTime API for interactive music generation.

2026-02-18

Google releases Lyria 3, making it available in the Gemini app and YouTube's Dream Track.

2026-03-06

A group of independent musicians files a class-action lawsuit against Google, alleging Lyria 3 was trained on their copyrighted YouTube content without permission.

2026-06-08

Google files a motion to dismiss the lawsuit, citing broad licensing terms in YouTube's user agreement.