From Unlawful Inputs to Unprotected Outputs: A Balanced Framework for Generative AI
Abstract
This article examines the growing legal conflict between generative AI systems and intellectual property rights. It argues that the legality of AI training data and the protection of AI-assisted outputs cannot be separated. The article reviews two global approaches; one that protects output based on human contribution, and another that rejects protection when the training process relies on unauthorized copying materials. It then proposes a balanced legislative solution built on three pillars: lawful training through prior licensing, a clear human-creativity threshold, and a strong enforcement remedy through algorithmic disgorgement, requiring the destruction of infringing models. The goal is to provide a practical, rights-respecting framework that supports innovation while safeguarding creators.
SUMMARY
• First: the legality of the training data used to build AI models, and
• Second: the copyright status of the outputs generated by those models.
This article argues that these two issues cannot be separated. If the Tree (training process) is unlawful, then the Fruit (output) cannot be treated as clean or eligible for protection. The central question becomes: Can an AI model trained on unlicensed copyrighted works produce an output that deserves legal protection?
The Flexible Approach: Protecting Outputs Based on Human Contribution
This approach focuses on the final stage of creation. It argues that even if the training process was questionable, the human user’s creative effort in guiding the model should be enough to justify protection.
Supporters of such an approach compare AI to tools like a camera or a paintbrush. If the user invests time, skill, and judgment, the output should be protected.
Supporting Examples
- · Li v. Liu (Beijing Internet Court, 2023 ) : The court held that an AI-generated image was copyrightable because the user exercised meaningful control through prompt engineering and several adjustments.
- · Scholarly Support : Scholars such as Mark Lemley warn that over-regulation may slow innovation. While Lemley does not explicitly support copyright for AI outputs, he emphasizes the importance of maintaining incentives for technological development.
The Judicial Retreat: How U.S. and U.K. Courts Have Moved Away from the Strict Approach
Recent judicial developments in both the United States and the United Kingdom demonstrate a clear retreat from the idea that unlicensed training-phase copying constitutes direct copyright infringement.
- · United Kingdom (High Court, November 2025): The Court rejected Getty Images’ primary copyright claim, holding that the act of training a generative AI model on copyrighted images did not, on its own, amount to direct infringement. Only limited trademark-related liability was upheld.
- · United States (N.D. California, April 2026): The Court declined to treat unlicensed training as actionable infringement and instead upheld only trademark and unfair-competition claims relating to the generation of synthetic images containing altered Getty watermarks.
Together, these rulings confirm that no major jurisdiction currently recognizes a per se rule prohibiting unlicensed AI training. The strict “fruit of the poisonous tree†theory has therefore lost its anticipated judicial foothold.
The Strict Approach: The (Fruit of the Poisonous Tree) Doctrine
Although courts have moved away from a licensing-first model, a number of scholars continue to argue that AI outputs are inherently tainted when the underlying training data is unlicensed.
This approach relies on the legal principle that what is built on an unlawful foundation remains unlawful.
This view is advanced by authors such as Benjamin Sobel who argues that generative AI systems may produce “probabilistic reproductions†of their training data is reinforced by a broader body of scholarship on non-literal copying and stylistic similarity.
Prominent scholars such as Pamela Samuelson , Jane Ginsburg , and the Nimmer treatise have long emphasized that copyright protection extends beyond verbatim duplication to include structure, style, creative choices, and expressive patterns. Their work supports the view that AI outputs may incorporate protected expressive elements derived from training datasets, even when the copying is indirect or statistical in nature.
Collectively, these authorities provide the intellectual foundation for the strict approach, which argues that unlawful training contaminates the resulting model and its outputs. However, despite its conceptual strength, this approach remains purely academic, with no jurisdiction having adopted it as a binding legal rule.
A Balanced Proposal: The (Prior Cleansing) Standard
To resolve this conflict, this article proposes a practical legislative framework built on three pillars.
1. Lawful Training Through Prior Licensing
No AI output should receive protection unless the developer or user can prove that the model was trained on lawfully obtained, licensed, or authorized data. This removes the “contamination †at the source and ensures that outputs can be evaluated fairly under existing copyright rules.
2. A Clear Human-Creativity Threshold
Once lawful training is confirmed, the human user must show real creative involvement. This includes meaningful prompt design, selection and arrangement of elements, editing, and creative judgment.
Simple or generic prompts should not qualify for protection. The goal is to protect genuine human creativity, not automated production.
3. Algorithmic Disgorgement: A Strong Enforcement Remedy
If a model is trained unlawfully, and the infringing data constitutes a core or substantial component of the model’s specialized capabilities, the appropriate remedy should be algorithmic disgorgement. This includes deletion of the model, deletion of any outputs derived from it, and destruction of the infringing training data. Such a remedy is reserved for severe violations where the unlawful material forms an essential part of the model’s functionality, ensuring proportionality while maintaining strong deterrence. This remedy already exists in the FTC’s enforcement actions in data-privacy cases such as:
These cases require companies to delete models built on unlawfully and unauthorized collected data. Applying this remedy to IP law would create a strong deterrent and encourage legal compliance from the start.
CONCLUSION
The debate over AI and copyright cannot be resolved by focusing only on outputs or only on training data. The two are inseparable. A lawful and transparent training process is the foundation for any legitimate claim to protection. At the same time, human creativity must remain at the center of copyright law.
By adopting a clear licensing requirement, a measurable creativity threshold, and a strong enforcement remedy through algorithmic disgorgement, lawmakers can create a balanced system that protects creators, guides developers, and supports responsible innovation.
This framework offers a practical path forward for policymakers, courts, and industry leaders seeking clarity in a rapidly evolving digital landscape.