AI Image Generation and Copyright Issues - Legal and Ethical Challenges

AI Image Generation Overview - What's at Stake

Image generation AI systems like Stable Diffusion, Midjourney, and DALL-E create high-quality images from text prompts. These models are trained on billions of image-text pairs, and their data collection methods and output rights attribution have sparked worldwide debate across legal, creative, and technology communities.

Technical mechanism:

Diffusion models generate images from random noise by reversing the process of gradually adding noise to images. During training, patterns are extracted from massive image-text pairs, acquiring the ability to generate images matching text conditions. Crucially, generated images are described as newly synthesized combinations of learned patterns rather than copies of training data.

Problem structure:

AI image generation issues divide into three stages: First, copyright issues in training data collection (whether using copyrighted works without permission for training is permissible). Second, copyright attribution of outputs (whether AI-generated images receive copyright and who owns it). Third, usage-stage issues (legal risks of generating and using images similar to existing copyrighted works).

International responses:

Japan, the US, and EU each take different approaches to organizing AI-copyright relationships. Japan's Copyright Act Article 30-4 relatively broadly permits copyrighted work use for information analysis purposes. In the US, fair use applicability has been contested in multiple lawsuits. The EU AI Act (2024) imposes transparency requirements, including that providers of general-purpose AI models make publicly available a sufficiently detailed summary of the content used for training.

Training Data and Copyright - Is Unauthorized Training Permissible

Most image data used for AI model training is collected from the internet. Large-scale datasets like LAION-5B contain over 5 billion image-text pairs crawled from the web. Many of these images are copyright-protected, used for training without rights holder permission.

Japan's legal framework (Copyright Act Article 30-4):

Japan's 2018 copyright law revision made "use not aimed at enjoying thoughts or sentiments expressed in works" subject to rights limitations. AI training qualifies as "information analysis," principally permissible without copyright holder permission. However, cases "unduly prejudicing copyright holder interests" are excluded. Interpreting "unduly prejudicing" is considered one of the points of debate.

US legal framework (Fair Use):

US copyright fair use (Section 107) comprehensively judges four factors: purpose/character of use, nature of copyrighted work, amount used, and market impact. Whether AI training constitutes fair use has been contested in multiple lawsuits. Getty Images v. Stability AI and Anderson v. Stability AI outcomes are likely to influence future direction.

Creator backlash:

Many artists and illustrators oppose unauthorized use of their works for AI training. Tools like Glaze and Nightshade add noise that is subtle to human eyes but disrupts AI training. ArtStation and DeviantArt have introduced AI training opt-out features for creators.

Licensed datasets:

To avoid rights issues, Adobe Firefly is reported to be trained on sources such as Adobe Stock licensed images and public domain works. Shutterstock provides AI models trained on their licensed images, building systems that compensate contributing creators fairly.

Copyright of AI-Generated Works - Who Holds the Rights

Whether AI-generated images receive copyright protection, and if so who owns it, remains unsettled. Different countries have reached different conclusions, with no unified international consensus yet established.

Japan's position:

According to Japan's Agency for Cultural Affairs guidance on AI and copyright (2024), whether AI outputs receive copyright depends on "creative contribution" presence. Simply inputting prompts likely doesn't constitute creative contribution, but prompt crafting, parameter adjustment, and post-generation selection/editing demonstrating human creative involvement may qualify for copyright protection.

US position:

The US Copyright Office maintains that autonomously AI-generated portions don't receive copyright. In the Zarya of the Dawn case (2023), copyright was denied for Midjourney-generated images themselves but granted for human-performed layout and text arrangement. AI-generated images used as materials with human creative editing/composition receive copyright for those human-contributed portions.

EU position:

The EU protects only "human intellectual creations" as copyrightable works. Autonomously AI-generated images aren't recognized as copyrightable, but human use of AI as a tool with creative judgment may qualify for protection.

Practical implications:

If AI-generated images lack copyright, third parties can freely copy, modify, and commercially use them. For businesses using AI images in product design or marketing, this means inability to claim exclusive rights. Conversely, for those using AI-generated images made by others, the risk of infringing copyright in the image itself is considered small.

Commercial Use Considerations - Business Risks of AI Image Utilization

Using AI-generated images in business involves multiple legal and ethical risks beyond copyright. Understanding and properly managing these is important corporate risk management.

Similarity to existing works risk:

If AI-generated images resemble existing copyrighted works in training data, copyright infringement claims are possible. Prompts mimicking specific artist styles (e.g., "in the style of [artist name]") increase risk of infringing original artists' copyright or trademark rights.

Portrait rights and publicity rights:

Generating images resembling real people may violate portrait or publicity rights. Commercial use of generated celebrity likenesses is illegal in some jurisdictions. Related to deepfake concerns, regulations have been developed in various countries.

Trademark infringement:

AI-generated images containing existing brand logos or trademarks risk trademark infringement. Logos and brand elements in training data may unintentionally appear in generated images. Verify generated images don't contain existing trademarks before commercial use.

Service terms of use:

• Midjourney: Commercial use permitted on paid plans. Companies above a certain revenue level require a higher-tier plan.
• DALL-E (OpenAI): Permits commercial use of generated images. Must comply with OpenAI content policy.
• Stable Diffusion: Open-source model follows CreativeML Open RAIL-M license. Commercial use conditions also follow that license, but illegal/harmful content generation is prohibited.
• Adobe Firefly: Designed for commercial use. Reportedly trained on licensed data, which may help keep copyright risk low. Provides IP Indemnification coverage.

Ethical Challenges and Social Impact - Technology Progress and Responsibility

AI image generation raises broader ethical and social challenges beyond legal issues. Society's response appears to lag behind technology's rapid advancement in some respects.

Economic impact on creators:

AI image generation proliferation raises concerns about reduced work for illustrators, photographers, and designers. Stock photography, concept art, and advertising visuals reportedly see some AI substitution of human creators. Conversely, some creators leverage AI as a productivity tool.

Misinformation and deepfakes:

High-quality image generation technology risks misuse for fake news images, political propaganda, and fraud. Non-existent person photos, fabricated incident images, and altered evidence photos threaten societal trust. C2PA (Coalition for Content Provenance and Authenticity) provenance technology is being developed as countermeasure.

Bias and representation:

AI models reflect training data biases. Tendencies toward specific race, gender, and cultural representations can reinforce stereotypes. Issues reported include "CEO" prompts generating predominantly white males and "beautiful person" prioritizing specific races in outputs.

Environmental impact:

Large-scale AI model training requires enormous computational resources with significant CO2 emissions. The Stable Diffusion v1-4 model card lists 150,000 hours on A100 PCIe 40GB as the compute time used to estimate the CO2 emissions of Stable Diffusion v1. Inference (image generation) power consumption is relatively small, but the total grows as user numbers grow.

Practical Countermeasures - Managing Risk While Leveraging AI

Concrete countermeasures for utilizing AI image generation in practice, organized from a risk management perspective. Complete risk elimination is difficult, but appropriate management can help reduce risk to practical levels for responsible adoption.

Usage policy development:

Establish organizational AI image generation usage policies. Document which services to use, permitted purposes, approval processes, and output management methods. Specifically clarify client work usage permissions, internal material usage scope, and public content usage conditions.

Attribution disclosure:

When using AI-generated images, disclosure is ethically recommended. Add credits like "AI generated" or "Created with [tool name]." Some industries (journalism, academia) prohibit or strictly limit AI-generated image use - verify industry guidelines before deployment.

Similarity checking:

Before commercial use, verify generated images don't resemble existing copyrighted works. Search for similar images using Google reverse image search or TinEye to assess infringement risk. Avoid prompts intentionally mimicking specific artist styles to reduce legal exposure.

License verification:

Thoroughly read AI service terms of use, understanding commercial use conditions, output rights attribution, and prohibitions. Conditions vary significantly between services - select appropriate services per project. Prioritize services offering IP indemnification (Adobe Firefly, Getty Images AI) for commercial applications.

Recording human creative involvement:

When potentially claiming copyright on AI outputs, record and document human creative involvement. Document prompt crafting, parameter adjustments, and post-generation selection/editing processes as evidence of creative contribution. This can aid future rights claims and dispute resolution if ownership questions arise.