Anthropic Allegedly Copied Copyrighted Song Lyrics Without Permission to Train Claude

Music publishers alleged that Anthropic copied copyrighted song lyrics without permission to train Claude and that Claude reproduced protected lyrics in its outputs. Anthropic denied infringement and has argued that its training use is fair use. Publishers alleged that the conduct deprived rightsholders and songwriters of licensing control and revenue and undermined markets for licensed lyrics.

Risk classification

• Primary risk domain: 6 Socioeconomic & Environmental
• Primary risk subdomain: 6.3 Economic and cultural devaluation of human effort

The training of generative AI models on unlicensed creative works devalues human artistry and songwriting by enabling automated systems to reproduce or imitate copyrighted music without compensation.

Additional risk subdomains

• 7.3 Lack of capability or robustness: The AI systems exhibited porous guardrails that failed to prevent the reproduction of copyrighted lyrics when prompted with slight variations or misspellings.

Causal factors

• Entity: Human
• Intent: Intentional
• Timing: Pre-deployment

The decision to scrape and train AI models on unlicensed copyrighted music and lyrics was an intentional action taken by humans during the pre-deployment development phase.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

Limited Risk: The systems involved are Claude (a chatbot) and Suno (a generative music creator), which fall under the category of chatbots and AI-generated content creators requiring transparency.

AI system and alleged parties

• AI system: Claude (Anthropic)
• AI purpose: Music Generation; Writing Assistant
• Behaviour type: Assistant
• Alleged developer: Large language model developers, Chatbot developers, Anthropic
• Alleged deployer: Anthropic, AI service providers
• Alleged harmed parties: Universal Music Group, Songwriters, Round Hill Music, publishers, Musicians, Music publishers, Intellectual Property rights holders, Copyright holders, Concord Music Group, BMG Rights Management, ABKCO Music

Harm severity

Highest direct severity in any category: Severe. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Severe, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Negligible, indirect Negligible
• Civil rights: direct Negligible, indirect Negligible
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Negligible, indirect Negligible
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Financial

Reported: The reports describe lawsuits seeking statutory damages of up to $150,000 per work, potentially exceeding $1 billion in each case, and a prior settlement of $1.5 billion by Anthropic.

Directly caused: No final damages have been awarded yet in the active lawsuits, but the average potential financial claim per occurrence is up to $1 billion.

Indirectly caused: N/A

Inferred additional harm: Songwriters and publishers are losing licensing revenue due to unauthorized training, which likely totals millions of dollars.

People affected

• Occurrences reported: 2
• People reportedly harmed: 500
• People reportedly exposed: 500

Potential causes

Management

• Prioritizing Speed Over Compliance: Rushed model training using scraped data to compete in the AI market.
• Bypassing Licensing Markets: Management chose to bypass functioning licensing markets to avoid paying.
• Inadequate Risk Assessment: Failed to assess the legal risks of training on massive unlicensed datasets.

Technology

• Porous Guardrails: Suno's guardrails failed to block misspellings of protected artist names.
• Model Memorization: Claude reproduced copyrighted lyrics on demand and closely copied structures.
• Stripped CMI Tools: Selected Newspaper and Scraping Browser tools to strip copyright info.

Data Inputs

• Unauthorized Torrent Datasets: Anthropic used torrents from Library Genesis and Pirate Library Mirror.
• Scraped Platform Audio: Suno scraped millions of clips from YouTube Music, Deezer, and Genius.
• Lack of Licensed Data: AI companies bypassed established licensing markets to acquire training data.

Human Factors

• Co-founder Eagerness: Co-founder celebrated torrent availability, ignoring copyright implications.
• User Prompts Exploiting Guardrails: Users prompted models to bypass filters using slight misspellings.

Process and Methods

• Active CMI Removal Process: Deliberately chose tools that strip headers, footers, and copyright notices.
• Inadequate Dataset Auditing: Failed to audit training datasets for unlicensed copyrighted works.
• Concealing Torrenting: Anthropic allegedly concealed torrenting activities during legal discovery.

Regulatory Environment

• Unsettled Fair Use Doctrine: Lack of clear legal precedents on whether AI training constitutes fair use.
• Pleading Stage Challenges: Courts struggled to define contours of AI licensing markets during litigation.

Information quality

• Classification confidence: High
• Reason for confidence: The reports provide detailed, consistent information regarding the legal complaints, the specific allegations of copyright infringement, the datasets used, and the defenses raised by the AI companies.
• Ambiguities identified: The exact amount of financial harm actually suffered by the publishers is legally disputed and yet to be determined by the courts.
• Alternative interpretations: The incident could be interpreted purely as a commercial/legal dispute over fair use rather than an AI safety failure, though it directly concerns AI training practices and guardrail limitations.

Major music publishers filed copyright infringement lawsuits against AI companies Anthropic and Suno for unlicensed scraping of copyrighted songs to train generative AI models. While representing a significant legal and economic dispute over intellectual property within the tech and creative sectors, the incident poses negligible direct threats to national security.

• Overall national security impact: Minor
• Response level: Moderate
• Scope: Single nation
• Primary target: United States
• Alleged perpetrator: Anthropic and Suno

Threat characteristics

• Imminence: Long-term. Represents an ongoing legal and regulatory debate over intellectual property rights rather than an active, immediate national security crisis.
• Autonomy: Human-controlled. The AI systems function as generative tools or assistants that respond to specific human prompts and require human initiation.
• Novelty: Evolved capability. Represents a significant advancement of existing intellectual property disputes adapted to large-scale generative AI training practices.

Impact by dimension

• Physical security: Negligible. The incident is a commercial copyright dispute with no impact on physical systems, critical infrastructure, or human safety.
• Information security: Negligible. No intelligence compromise, classified data theft, or information warfare operations are indicated in this civil legal dispute.
• Sovereignty: Negligible. This is a civil legal dispute between private commercial entities and does not threaten state authority, elections, or government operations.
• Economic security: Minor. While involving billions in potential damages and affecting the domestic creative economy and AI industry, it remains a civil IP dispute rather than a strategic threat to national economic security.
• Societal stability: Negligible. No mass surveillance, systematic discrimination, or threat to civil liberties or societal stability is present in this copyright case.