Tackling copyright issues in AI image generation through originality estimation and genericization

Area of Science:

• Computer Science
• Artificial Intelligence
• Intellectual Property Law

Background:

• Generative AI's rapid advancement raises significant copyright concerns, evidenced by numerous lawsuits.
• AI models can generate images of copyrighted characters, posing legal and ethical challenges.
• Existing copyright mitigation techniques for AI have limitations, leaving substantial risks.

Purpose of the Study:

• To propose a novel genericization method for generative AI outputs to mitigate copyright infringement.
• To develop a metric for quantifying data originality within generative models.
• To introduce PREGen (Prompt Rewriting-Enhanced Genericization) as a practical implementation.

Main Methods:

• Developed a genericization technique to modify AI model outputs, reducing imitation of copyrighted features.
• Introduced a metric to quantify data originality by sampling from generative models.
• Implemented PREGen by combining the genericization method with an existing technique.

Main Results:

• PREGen reduced the generation of copyrighted characters by over 50% when character names were used in prompts.
• PREGen nearly eliminated the generation of copyrighted characters when names were not explicitly mentioned.
• The developed originality metric was applied effectively within the genericization process.

Conclusions:

• The proposed genericization method and PREGen offer practical solutions for strengthening copyright protection in generative AI.
• This research advances computational approaches for responsible AI development and copyright compliance.
• The findings provide methodologies to reduce the likelihood of AI generating infringing content.