The Digital Dilemma of Copyright: Legal and Ethical Boundaries of Generative AI

In today's rapidly advancing world of artificial intelligence, generative AI has taken the globe by storm with its amazing creative abilities, spanning from text and images to music and video. AI seems capable of doing anything. However, while marveling at these technological breakthroughs, a fundamental question is becoming increasingly prominent: the copyright issues involved in the creation process of these AI systems. As the boundaries of machine "creation" become blurred, traditional copyright legal frameworks face unprecedented challenges. This article will delve into the copyright dilemma posed by generative AI, analyze existing cases and legal developments, and contemplate possible solutions.

Training Data: The Starting Point of Copyright Disputes

The capabilities of generative AI stem from its training data. Whether it's GPT, DALL-E, Midjourney, or Stable Diffusion, these models master creative techniques by learning from a vast amount of human-created works. However, this basic training process itself raises the first copyright challenge.

Data Acquisition and the "Fair Use" Debate

Large language models from companies like OpenAI and Anthropic include a large amount of online text in their training data, often containing copyrighted content. According to a New York Times investigation, OpenAI's training data includes over 11,000 books, including unauthorized bestsellers. This has led to several important lawsuits:

• In December 2023, The New York Times sued OpenAI and Microsoft, accusing them of using millions of copyrighted news articles to train ChatGPT without permission.
• Authors George R.R. Martin (author of Game of Thrones) and 17 other prominent authors filed a class-action lawsuit against OpenAI for copyright infringement.
• Getty Images sued Stability AI (developer of Stable Diffusion), accusing it of unauthorized scraping of millions of Getty photos for training.

The core point of contention in these cases is whether AI companies can argue that their data usage falls under the "Fair Use" doctrine under US copyright law, because:

1. They do not directly copy or display the original content.
2. The models extract patterns from the data rather than specific content.
3. This usage is "transformative" and serves a different purpose.

However, content creators argue that:

1. Commercial companies are profiting from unauthorized large-scale use of copyrighted content.
2. AI products directly compete with original authors in the market.
3. No compensation mechanisms are provided.

The preliminary ruling in the case of "Andersen v. Stability AI" by the United States District Court for the Southern District of New York in 2023 may be instructive. The judge ruled that claiming fair use simply because it was used to train AI is insufficient and requires case-by-case analysis.

EU Data Mining Exception and Global Differences

Unlike the United States, the European Union explicitly provides a "text and data mining exception" in its copyright law. Articles 3 and 4 of the Digital Single Market Directive allow research organizations and other entities to conduct text and data mining, but Article 4 also grants rights holders the right to "opt out."

Japanese copyright law is even more open, explicitly allowing the use of copyrighted content for data analysis purposes, making Japan a friendly environment for AI research.

This inconsistency in the global legal framework has led to the phenomenon of "legal arbitrage" in AI development, where companies may choose to train models in regions with more lenient legal environments.

AI-Generated Content: Who is the Creator?

Another core question is: Who should own AI-generated content? Can it be protected by copyright?

Different Stances of Global Copyright Agencies

The U.S. Copyright Office denied copyright protection to the Midjourney-generated image "Teatro D'opera Spatial" in 2023, and clarified in its policy statement Copyright and AI: "The principle of human authorship is the cornerstone of copyright protection, and U.S. copyright law only protects human intellectual creations." However, the Copyright Office also stated that works created in collaboration between humans and AI, the human creative contribution can be protected.

The UK Intellectual Property Office has taken a more flexible stance, according to its Copyright and AI-Generated Works guidance, AI-generated works may be protected by copyright, but the "author" will be considered "the person who made the arrangements necessary for the creation of the work."

The revision of China's Copyright Law in 2020 did not explicitly exclude AI creation, which to some extent provides space for copyright protection of AI-generated content.

Lessons from Landmark Cases

In 2022, American author Charles Bazan attempted to register a copyright for the novel Zarya of the Dawn, which included illustrations generated by Midjourney. The Copyright Office ultimately granted copyright only for the text portion and denied it for the image portion.

More controversially, Microsoft and OpenAI's GitHub Copilot is facing a class-action lawsuit for potentially including original code snippets from the training data in the generated code. This "memorization" phenomenon raises concerns about whether AI systems directly copy original content.

Work Similarity: The Blurred Line from Inspiration to Plagiarism

Another issue raised by generative AI is: When AI-generated content is similar to existing works, how do you determine if it constitutes infringement?

"Style Imitation" and Copyright Boundaries

The most typical controversy comes from the field of image generation. Users can ask AI to "create in the style of Van Gogh" or "like a Disney animation," raising concerns about style plagiarism. In 2023, several artists including Kelly McKernan and Greg Rutkowski filed lawsuits accusing Stability AI and Midjourney of infringing their artistic styles.

However, traditional copyright law does not protect style, technique, or ideas, only specific expressions. This principle is challenged in the AI era, as AI can systematically learn and imitate the stylistic characteristics of artists.

Removing Original Elements from Training Data

Some AI companies are trying to mitigate copyright risks through technical means:

• OpenAI added a filter to DALL-E 3 that refuses to generate requests that imitate the style of a specific artist.
• Anthropic's Claude model refuses to repeat complete copyrighted content.
• Midjourney explicitly prohibits users from entering certain well-known artist names as prompts.

However, studies show that the effectiveness of these measures is limited. Researchers at Stanford University found that even without directly using the artist's name, using related descriptive words can allow AI to generate works with similar styles.

Legislation and Market Adaptation: The Road Ahead

Faced with these challenges, legislators, businesses, and creators around the world are exploring different paths:

Emerging Legislative Attempts

The EU Artificial Intelligence Act requires providers of generative AI to disclose the copyrighted content used in their training and to provide rights holders with an opt-out mechanism.

The Clarifying Lawful Use of Our Valuable Intellectual Property Act (CLUE Act) proposed in the US Senate attempts to clarify whether AI training constitutes fair use, but has not yet been passed.

The Guiding Opinions on Strengthening Copyright Protection of Artificial Intelligence-Generated Content issued by the China National Copyright Administration in 2023 proposes that artificial intelligence-generated content can be protected by copyright if it has originality and is creatively expressed by natural persons.

Licensing and Compensation Models

Some companies have begun to explore licensing models:

• The Associated Press reached an agreement with OpenAI to authorize the latter to use its news archives.
• Shutterstock has established partnerships with OpenAI and Stability AI to allow the use of its image library to train AI and establish a compensation fund.
• Adobe Stock has created a "Generative AI Friendly" licensing program that explicitly allows some content to be used for AI training.

These arrangements represent a possible market solution: ensuring that original creators are compensated through direct licensing agreements.

Technical Solutions

Technologies such as blockchain and digital watermarks have also been proposed as tools to solve copyright issues:

• The C2PA (Coalition for Content Provenance and Authenticity) developed content authentication standards to help identify AI-generated content.
• Midjourney and DALL-E embed metadata in the generated images, indicating their AI origin.
• Some startups have developed tools to distinguish between AI-generated content and human creation.

Multi-Party Balance: Rethinking the Interests of All Parties

Solving the copyright issues of generative AI requires finding a balance between multiple stakeholders:

The Legitimate Rights and Interests of Content Creators

Original authors deserve protection and compensation. The Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA) made AI usage restrictions one of its core demands in the 2023 strike. The final agreement requires production companies to obtain consent and provide compensation before using AI to copy actors' images.

Similarly, the music industry is also actively exploring protection mechanisms. Universal Music Group CEO Lucian Grainge said in 2023: "AI can be a tool for artists, but it cannot be a substitute."

Technological Development and Public Interests

At the same time, overly strict restrictions may hinder technological progress and social benefits. AI researchers point out that if each training data requires individual authorization, it will make model development too expensive and limit innovation.

A potential balanced solution is to establish a statutory licensing system, similar to the mechanical licensing system in the music industry, allowing the use of protected content but requiring the payment of reasonable fees.

Conclusion: Towards a Symbiotic Future

The copyright challenges posed by generative AI have no simple solutions. The speed of technological development far exceeds the speed of legal adjustment, and this gap requires joint efforts from all parties to bridge.

The most likely development direction in the future is: the gradual improvement of the legal framework, the innovation of commercial licensing models, the support of technical tools, and more in-depth dialogue and cooperation among all stakeholders. In this process, we need to both protect the legitimate rights and interests of creators and provide space for technological innovation, and ultimately establish an ecosystem in which AI and human creativity coexist.

Just as the fundamental purpose of copyright law is to promote the development of knowledge and culture, facing the challenges of the AI era, we need to return to this core concept and find a balance that can simultaneously incentivize human creativity and technological innovation. This is not only a legal issue, but also a deeper reflection on how we define creativity itself and how humans and machines should coexist in this new era.