The Generative AI Revolution: Authorship, Ownership, Training Data, and Global IP Risks

Artificial Intelligence (AI) has transitioned from a theoretical concept to a practical tool that creators, studios, and platforms utilize daily. This shift has forced the legal community to confront immediate questions regarding authorship, ownership, licensing, and enforcement. As lawyers are increasingly asked to provide guidance on these complex issues, understanding the current global regulatory and judicial landscape is paramount.

Quick Summary

• Copyright ownership depends on the role of the human in the creative process.

• Training data and the ability of models to reproduce content has created major infringement risk.

• Enforcement faces new challenges like synthetic-to-synthetic training and “copyright laundering.”

• Ethical boundaries include likeness protection and trade secret leakage through prompting.

• Global approaches differ across the U.S., EU, and Asia, creating complexity for cross-border creators.

1) Who Owns AI-Generated Work? The Authorship Dilemma

The fundamental question of ownership hinges on the role of the human in the creative process. Under current U.S. copyright law, human authorship is a bedrock requirement; no court has recognized copyright in material created entirely by non-humans.

Purely AI-Generated Content

The U.S. Copyright Office (USCO) maintains that content generated entirely by an AI system, without a guiding human hand, cannot be protected by copyright. This position was recently affirmed by the U.S. District Court for the District of Columbia in Thaler v. Perlmutter, which ruled that copyright law protects only works of human creation.

The Problem with Prompts

Many users believe that “prompt engineering” constitutes authorship. However, the USCO concludes that, based on current technology, prompts generally function as unprotectable ideas or instructions. Because a user lacks sufficient control over how an AI interprets a prompt to generate an expressive output, the user is not considered the “author” of that output.

AI as an Assistive Tool

Copyright protection remains available when AI is used merely as a tool to assist a human creator. If a human makes significant creative choices, such as selecting, arranging, or modifying AI-generated elements in a way that reflects their own intellectual creation, those human-authored aspects are copyrightable.

The Trade Secret Loophole

While copyright requires a human author, trade secret law has no such requirement. This means that purely AI-generated outputs, such as a specialized pharmaceutical formula or a proprietary algorithm, could potentially be protected as trade secrets as long as the owner takes reasonable measures to keep the information secret.

2) Training Data Risks: The Proof of Infringement

The massive datasets required to train large language models (LLMs) have created significant legal risks for AI developers.

Access and Substantial Similarity

High-profile lawsuits, such as The New York Times vs. OpenAI, allege that AI models were trained on millions of copyrighted works without permission. The core of these claims is that models can sometimes reproduce content verbatim or produce “hallucinations” that attribute made-up facts to reputable sources, demonstrating that the training data was “memorized” rather than truly transformed.

Filtering Failures

While developers may implement filtering systems to exclude copyrighted material, these systems face technical gaps. Copyright databases are often fragmented or incomplete, and automated systems struggle to distinguish between protected creative expression and unprotectable facts.

Licensing at Scale

The USCO has noted that it is not currently “practically possible” to obtain individual licenses for the sheer volume of content necessary to power cutting-edge AI systems. This has led to a two-tiered system where some developers strike deals with major publishers (like Axel Springer or Politico), while others rely on the fair use doctrine to justify mass scraping.

3) Enforcement Challenges: The AI Ouroboros and “Copyright Laundering”

A major emerging threat to IP enforcement is what experts call the “AI Ouroboros”, the practice of training successor AI models on the synthetic outputs of their predecessors.

Copyright Laundering

By using multi-generational pipelines, developers can effectively “launder” copyrighted material. As data passes through successive models, it is diffused into statistical abstractions, making it nearly impossible for a human author to prove that a late-generation model was trained on their specific work.

Fruit of the Poisonous Tree (AI-FOPT)

To combat this, legal scholars have proposed a new evidentiary rule: if a foundational model’s training is adjudged infringing, all subsequent models derived from it carry a rebuttable presumption of taint. This would shift the burden of proof to the developer to demonstrate a “clean lineage” or a “curative rebuild”.

Model Weights as Infringing Articles

In the UK, the case of Getty Images v. Stability AI tested whether model weights themselves could be considered “infringing copies”. The High Court ruled that for an article to be an infringing copy, it must have at some point contained or stored a copy of a copyrighted work; because weights do not store works in this manner, the claim of secondary infringement failed.

4) Ethical Boundaries: Protecting Likeness and Confidentiality

The rise of generative AI also creates new ethical risks regarding individual identity and corporate privacy.

The No Fakes Act

Legislation like the proposed bipartisan No Fakes Act aims to protect individuals from unauthorized AI-generated deepfakes of their voice or likeness. This is especially critical in the music industry, where AI can recreate an artist’s voice to produce unauthorized performances, damaging reputations and interfering with an artist’s control over their identity.

Trade Secret Leakage

A significant risk for companies is that employees may inadvertently share trade secrets while prompting public AI models. If a model trains on these inputs, the secret information may be disclosed to third parties, potentially destroying its legal status as a trade secret.

Reasonable Measures

To maintain trade secret protection, companies must now adapt their risk management strategies. This includes using “Enterprise Licenses” with confidentiality clauses or developing in-house AI models that do not train on user data for public distribution.

5) Global Perspectives: US, EU, and Asia

Approaches to AI and IP vary significantly across jurisdictions, creating a complex web for international creators.

European Union

The EU has taken a proactive regulatory approach with the EU AI Act, which introduces transparency obligations for providers of general-purpose AI models. The EU generally roots authorship in human creativity, requiring a “personal touch” for copyright protection.

Asia-Pacific

Japan and Singapore have adopted broad exceptions for text and data mining (TDM), even allowing commercial AI training on copyrighted materials. In contrast, Chinese courts have shown more flexibility, sometimes granting copyright to images generated through highly complex, multi-stage prompting processes.

Ukraine

Uniquely, Ukraine has adopted a sui generis right to provide alternative protection for AI-generated images, creating a “middle way” between traditional copyright and no protection at all.

Conclusion: The Future of Creative Compliance

The legal framework for AI and Intellectual Property is far from settled. While the U.S. Copyright Office continues to monitor technological developments, creators and lawyers must navigate a landscape of reactive litigation and evolving state-level transparency mandates. As AI models become more sophisticated, the focus of enforcement will likely shift from analyzing final outputs to auditing the lineage and ethics of the training data. For creators, the message is clear: the more human involvement you can prove, the stronger your legal claim.

FAQ

Can AI-generated content be copyrighted?

Under current U.S. copyright law, human authorship is required, and purely AI-generated content without a guiding human hand is not protected by copyright.

Are prompts enough to claim authorship?

The USCO’s position is that prompts generally function as unprotectable ideas or instructions, and a user typically lacks sufficient control over how an AI interprets the prompt.

What is “copyright laundering” in AI?

It refers to training successor AI models on synthetic outputs from earlier models, diffusing potential copyrighted material into abstractions that are difficult to trace.

What is the risk of trade secret leakage with AI tools?

Employees may inadvertently share trade secrets while prompting public AI models, and if those inputs are disclosed later, trade secret protection can be destroyed.

Why do global rules matter for creators?

Different jurisdictions take different approaches (U.S., EU, Asia), so cross-border licensing, compliance, and enforceability can change depending on where content is distributed or used.