Who Owns the Output? AI Copyright & IP Explained w/ Chris Paniewski | Startup Legal Basics

Key Concepts

• Large Language Models (LLMs): AI models trained on vast amounts of text data, capable of generating human-like text.
• Copyright Law: Legal framework protecting original works of authorship, including literary, dramatic, musical, and certain other intellectual works.
• Fair Use: A legal doctrine that permits limited use of copyrighted material without requiring permission from the rights holders, for purposes such as criticism, comment, news reporting, teaching, scholarship, and research.
• Human Authorship: The requirement that a human being must be the originator of creative expression for copyright protection to apply.
• Derivative Work: A work based upon one or more pre-existing works, such as a translation, musical arrangement, dramatization, or any other form in which a work may be recast, transformed, or adapted.
• Intellectual Property (IP): Creations of the mind, such as inventions; literary and artistic works; designs; and symbols, names, and images used in commerce.
• Statutory Damages: Damages awarded by a court when a copyright has been infringed, even if the copyright holder cannot prove actual damages.
• Creative Commons: A type of copyright license that allows creators to grant certain permissions for others to use their work.
• Enterprise Versions: Specialized versions of software or platforms designed for business use, often with enhanced security, privacy, and licensing terms.

AI, Copyright, and Intellectual Property: A Deep Dive

Protecting Copyrighted Content in LLMs

• Guardrails: LLMs are implementing measures to prevent the generation of infringing content, particularly concerning well-known IP like Star Wars.
• Fair Use Argument: LLMs aim to bolster their fair use defense by demonstrating efforts to protect copyrighted content.
• New Technologies and IP Rights: The emergence of AI, like previous technological advancements (DVDs, VOD), doesn't automatically grant the right to exploit existing IP without permission.
• Example: LLMs may refuse prompts that directly request the creation of content based on Disney's Star Wars IP.

AI-Generated Content and Copyrightability

• Human Authorship Requirement: Copyright law mandates that creative works originate from human creativity.
• LLMs and Creative Control: If an LLM significantly contributes to the creative elements of a work, it may lack sufficient human authorship for copyright protection.
• Copyright Office Rejection: Attempts to register AI-generated images without substantial human input have been rejected by the copyright office.
• Example: If a user prompts an LLM to design a weapon and the LLM generates a laser sword, the user may struggle to IP protect that design due to the AI's creative contribution.

The Level of Human Input and Copyright

• Detailed Prompts: Providing LLMs with highly detailed prompts increases the likelihood of copyright protection.
• Photography Analogy: The Supreme Court has ruled that photographs are copyrightable because humans control the scene, posing, camera settings, and the moment of capture.
• Camera as a Tool: The camera is considered a tool, and the human is the creator in photography.
• Monkey Selfie Case: A photograph taken by a monkey was deemed uncopyrightable because no human was operating the camera.
• AI as Inventor: Attempts to register patents with AI as the named inventor have been unsuccessful.
• Example: A security camera taking a picture without human involvement cannot be copyrighted.

Opportunities and Strategies for Startups

• Partnerships with IP Owners: Startups can collaborate with IP owners to develop AI-powered applications based on their IP.
• Proof of Concept: Building a proof of concept and presenting it to the IP owner can lead to licensing or revenue-sharing agreements.
• Balancing AI Use and IP Protection: Startups must balance the benefits of using AI for development with the need to protect their own IP.
• Documenting AI Involvement: Startups should meticulously document which parts of their work are AI-generated and which are human-created.
• Human Integration: Integrating AI-generated content with human input, such as text, organization, and integration with other applications, can create copyrightable subject matter.
• Example: A startup could develop a "Jedi Me" feature for Disney Plus, allowing users to transform themselves into Jedi Knights.

Derivative Works and Fair Use

• Building on Existing Work: When building on top of existing work, it's crucial to ensure that all necessary rights are obtained.
• Wirecutter Example: Using an LLM to summarize reviews from Wirecutter and Consumer Reports could deprive those services of traffic and revenue, potentially violating fair use.
• Damages: Copyright infringement cases require demonstrating damages to the rights holder.
• Transformative Work: A human-created meta-analysis of reviews from multiple sources could be considered a transformative work.
• Creative Commons Licenses: Using content licensed under Creative Commons may allow for broader use, depending on the specific license terms.
• Example: Andy Warhol's use of Campbell's soup cans raised questions about derivative works and the need for rights.

Training LLMs and IP Considerations

• Human Trainers: Hiring humans to train LLMs by summarizing and distilling information from various sources raises IP concerns.
• Derivative Work Analysis: The human trainer's work must be analyzed to determine if it's a derivative work of the original sources.
• Access and Substantial Similarity: To prove copyright infringement, access to the original work and substantial similarity must be demonstrated.
• Opinion vs. Copying: Expressing personal opinions on existing reviews is more likely to be considered an original work.
• Metacritic and Rotten Tomatoes Analogy: These platforms create their own scores based on interpretations of reviews, which is generally considered fair use.
• Example: A human trainer summarizing Wirecutter and Consumer Reports reviews and feeding that information into an LLM.

Ethical Considerations and Unintentional IP Infringement

• Technologist Perspective: Technologists should consider how they would feel if their own IP was stolen.
• AI Platform Terms: The terms of service of AI platforms can vary significantly, potentially allowing the platform to use user prompts for training purposes.
• Data Privacy: Prompts containing sensitive or proprietary information could be exposed to others if the AI platform's terms are not carefully reviewed.
• Enterprise Versions: Using enterprise versions of AI platforms provides greater protection for businesses.
• Example: An AI platform's terms might allow it to train its model on user prompts, potentially revealing confidential information.

Conclusion

The intersection of AI, copyright, and intellectual property is complex and constantly evolving. Startups and developers must carefully consider the legal and ethical implications of using AI to create content, ensuring they respect existing IP rights and protect their own. Understanding the nuances of human authorship, fair use, and derivative works is crucial for navigating this landscape successfully.