Stable Diffusion's Training Data Contained Copyrighted Images

Stability AI reportedly scraped copyrighted images by Getty Images to be used as training data for Stable Diffusion model.

Risk classification

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

The training of Stable Diffusion on billions of copyrighted images without consent devalues the creative efforts of artists and threatens to disrupt the creative industry by flooding the market with competing synthetic images.

Causal factors

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

The incident stems from intentional decisions by humans at Stability AI, Midjourney, and DeviantArt to scrape copyrighted images and deploy generative models without licensing.

EU AI Act risk tier

• Risk tier: 3 Limited Risk

Limited Risk: The systems involved are generative AI models that produce 'AI-generated content' from text prompts, which carries transparency obligations under the Act to ensure users are aware of the AI's nature.

AI system and alleged parties

• AI system: DreamUp, Midjourney, Stable Diffusion (DeviantArt, Midjourney, Stability AI)
• AI purpose: Visual Art Generation; Image Generation
• Behaviour type: Tool
• Alleged developer: Stability AI, Runway, LAION, EleutherAI, CompVis LMU
• Alleged deployer: Stability AI
• Alleged harmed parties: Getty Images contributors, Getty Images

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Negligible. 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 Negligible, indirect Severe
• 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 report explicitly describes potential financial losses, noting that the value of the misappropriated images is roughly $5 billion assuming nominal damages of $1 per image.

Directly caused: The report does not specify a concrete, realized direct financial loss figure for the plaintiffs, but notes the potential statutory damages of $5 billion for the class.

Indirectly caused: The report notes that Stable Diffusion's ability to flood the market with infringing images will inflict permanent damage on the market for art and artists.

Inferred additional harm: Widespread financial losses are highly likely for millions of artists whose market value and licensing opportunities are devalued by competing AI-generated works.

People affected

• Occurrences reported: 1
• People reportedly harmed: 3
• People reportedly exposed: 3

Potential causes

Management

• Prioritizing speed over ethics: Released Stable Diffusion rapidly to secure market share ahead of licensing.
• Ignoring viable licensing options: Chose to bypass existing licensing frameworks to reduce development costs.
• Betrayal of community trust: DeviantArt integrated AI tools, violating their own community's trust.

Technology

• Lossy compression of training data: Diffusion and autoencoders store and reconstruct compressed copies of images.
• Watermark replication: Models recreate Getty watermarks, proving direct copying of source images.
• Text-prompt conditioning: Prompts mask direct copying while generating derivative works from data.

Data Inputs

• Unlicensed web scraping: Scraped billions of copyrighted images from websites without authorization.
• Use of unvetted datasets: Utilized LAION dataset containing millions of copyrighted works from artists.
• Lack of metadata consent: Processed image metadata alongside images without creator permission.

Human Factors

• Disregard for artist consent: Developers assumed scraping public web images was acceptable without consent.
• User demand for artistic styles: Users prompt models to generate art in specific living artists' styles.
• Lack of user awareness: Users unaware that generated images derive from copyrighted training data.

Process and Methods

• Lack of pre-training licensing: Failed to establish licensing agreements with rights holders before training.
• No initial opt-out mechanism: No process for artists to opt-out of training datasets before model release.
• Absence of data vetting process: No internal reviews to filter out copyrighted content from training sets.

Regulatory Environment

• Unclear generative AI laws: Lack of specific regulations regarding copyright fair use in AI training.
• Legal gray area on scraping: Ambiguous legal boundaries regarding web scraping for commercial AI models.
• Lack of regulatory foresight: Regulators failed to establish guidelines before commercial AI deployment.

Information quality

• Classification confidence: High
• Reason for confidence: The reports provide clear, detailed, and consistent accounts of the lawsuits filed by artists and Getty Images against the AI companies. The legal arguments, technical descriptions of the diffusion process, and the core dispute over copyright and fair use are well-documented across multiple sources.
• Ambiguities identified: The exact legal boundaries of 'fair use' versus copyright infringement for AI training remain unresolved in the courts at the time of the reports.
• Alternative interpretations: The defendants argue that the training process constitutes 'fair use' rather than copyright infringement, presenting an alternative legal interpretation of the same facts.

A major legal dispute emerged as artists and Getty Images sued Stability AI, Midjourney, and DeviantArt over the unauthorized scraping of copyrighted images to train generative AI models. While representing a significant milestone in intellectual property law and the economics of the creative sector, the national security implications remain negligible to minor.

• Overall national security impact: Minor
• Scope: Multiple nations
• Primary target: No clear primary
• Other affected: United States, United Kingdom
• Alleged perpetrator: Stability AI, Midjourney, and DeviantArt

Threat characteristics

• Imminence: Long-term. The dispute represents an ongoing legal and strategic concern regarding intellectual property rights in AI training, rather than an active operational crisis.
• Autonomy: Human-controlled. The AI systems involved are generative tools that act based on direct human prompts and were developed through intentional human decisions to scrape data.
• Novelty: Evolved capability. While copyright disputes are established, the scale of data scraping for training generative AI models represents a highly evolved capability and a new legal frontier.

Impact by dimension

• Physical security: Negligible. No physical security, kinetic threats, or critical infrastructure impacts are reported or indicated in this incident.
• Information security: Negligible. No indicators of disinformation campaigns, intelligence compromise, or classified data theft. The incident is strictly a commercial copyright dispute.
• Sovereignty: Negligible. No disruption to state authority, government decision-making, or electoral systems. The dispute is being handled through established judicial processes.
• Economic security: Minor. While the economic impact on the creative industry and individual artists is potentially significant, it does not currently threaten national strategic industries, critical supply chains, or overall financial stability.
• Societal stability: Negligible. No threats to civil liberties, mass surveillance, or population-scale social manipulation. The impact is limited to economic and copyright concerns for creators.