AI Copyright & Training Data w/ Chris Paniewski | Wilson Sonsini Startup Legal Basics

Key Concepts:

• AI Training Data: The data used to train AI models.
• Copyright: Legal protection for original works of authorship.
• Fair Use: A legal doctrine that permits limited use of copyrighted material without permission from the rights holder.
• Terms of Service: Rules and guidelines that users must agree to in order to use a website or service.
• Open Source Licenses: Licenses that grant users the right to use, modify, and distribute software or other works.
• ShareAlike Licenses: A type of open-source license that requires derivative works to be licensed under the same terms.
• Transformative Use: Using copyrighted material in a way that creates something new or different.
• Due Diligence: The process of investigating a company or investment opportunity before making a decision.

1. Copyright and Training Data

• Main Issue: The legal complexities surrounding the use of copyrighted material to train AI models.
• Two Components:
  • Training: Can AI be trained on copyrighted material? Is permission required?
  • Output: What are the copyright implications of AI-generated content?
• Legal Landscape: Currently being worked out in the courts, with outcomes dependent on specific facts.
• Geographic Considerations: Laws vary significantly between countries (e.g., US, Israel, China, Japan). Founders must consider where they launch their product, conduct training, and source data.
• Gold Standard: Obtaining explicit permission from copyright holders.
• Fair Use Confusion: The concept of fair use is often misunderstood, especially in commercial contexts. Commercial use weakens the fair use defense.
• Risk Assessment: Companies should assess their risk appetite based on their goals (e.g., aqua-hire vs. product-based exit).

2. Permissible Use and Novel Solutions

• Tension and Opportunity: The legal uncertainty is driving the creation of new solutions.
• Cloudflare Example: Offering tools for website owners to implement terms of service agreements for data usage.
• Clearing Houses: Companies are emerging to facilitate licensing of training data.
• Open-Source Data: While available, it doesn't automatically grant unrestricted usage rights.
• Open Crawl: A web crawling service that provides data but does not grant licenses on behalf of content holders.
• Misconception: The belief that anything found on the open web is free to use.

3. Rules of Thumb for Using Online Information

• Circumvention: Avoid circumventing paywalls, login access, or terms of service to access content.
• Terms of Service as Contracts: Agreeing to terms of service creates a contract that restricts data usage.
• Potential Consequences: Violating terms of service or copyright can lead to legal letters and deal-breaking issues during due diligence.

4. Image Generator Example and Data Acquisition Strategies

• Possible Paths:
  • Using copyright-free or royalty-free images (carefully reviewing terms of service).
  • Contacting content creators for permission.
• Licensing Challenges: Licensing from every content owner is impractical.
• Market Dynamics: Larger publications (e.g., Condé Nast, The New York Times) are creating their own licensing markets.
• Fair Use for Proof of Concept: Using scraped content for initial testing before licensing may be considered fair use.
• Open-Source License Caveats:
  • Creative Commons: Not always free to use without restrictions.
  • ShareAlike Licenses: Require derivative works to be licensed under the same terms, preventing protection of the output.

5. Project Scope and Risk Factors

• Dorm Room Project vs. Venture-Backed Startup: The level of scrutiny and risk increases significantly when a project transitions from a personal endeavor to a funded business.
• Experimentation Space: Some leeway exists for experimentation as long as it's not leveraged for commercial gain.
• Enforcement Triggers:
  • Being the first to "stick your neck out" and potentially setting a precedent.
  • Having "deep pockets" and being a worthwhile target for rights holders.

6. Autoblog Example and Fair Use Considerations

• Spy Photo Scenario: Using copyrighted images from Car and Driver on Autoblog led to legal issues.
• Compromise: A solution was reached by using a cropped image, linking back to Car and Driver, and providing attribution.
• Fairness vs. Fair Use: Balancing legal fair use with practical fairness.
• Fair Use as a Defense: Fair use is a legal defense, not a right.
• Four Factors of Fair Use:
  • Purpose of the copying.
  • Nature of the copyrighted work.
  • Amount and substantiality of the portion used.
  • Effect of the use on the potential market for the copyrighted work.
• Transformative Use: Using content in a way that creates something new or different strengthens a fair use argument.
  • Example: Training on books in multiple languages for a translation app (transformative).
  • Example: Training a model to write summaries of the original work (not transformative).

7. YouTube and Copyright Enforcement

• Content ID System: YouTube's system for identifying and managing copyrighted content.
• Options for Rights Holders: Take down infringing content or monetize it.
• Transformative Work and Monetization: Commentary and reaction videos may be allowed, or rights holders may choose to monetize them.
• Lack of Case Law: Many copyright disputes are settled out of court, resulting in limited case law.

8. Gambling with Fair Use

• Judge Determination: Fair use is ultimately determined by a judge.
• Alternative: Licensing Agreements: Striking deals provides certainty and defined parameters.
• Risk Assessment: Deciding whether to gamble on a judge's ruling or pursue a licensing agreement.

9. Conclusion

• The legal landscape surrounding AI training data and copyright is complex and evolving.
• Founders must carefully consider the legal implications of their data acquisition and usage practices.
• Fair use is a risky defense, and licensing agreements may be a more reliable option.
• Part two of the series will focus on the copyright implications of AI-generated output.

Main Takeaways:

• Navigating copyright law in the context of AI training data requires careful consideration of legal, ethical, and practical factors.
• A proactive approach to licensing and compliance is crucial for mitigating risk and building a sustainable AI business.
• The legal landscape is still developing, so staying informed and seeking expert legal advice is essential.