A Surprisingly Interesting Look at Intellectual Property Law & Liability in an AI World

Artificial Intelligence and especially Generative AI bring with them a maelstrom of legal ramifications. And although these seem like completely new issues, they still have a lot in common with the ones from yesteryear. We’ll use stories including Selfie Monkeys, One Dollar Trucks (no takesies backsies), and Books Stripped Naked to take a non-lawyer’s whirlwind tour of Intellectual Property law & legal liability and how they interact with advances in Information Technology. In the end, you’ll come away with both a better understanding of these waters and ideas on how you can sail your way more successfully through them (both generally and specifically).

Transcript

1. A Surprisingly Interesting Look at Intellectual Property Law & Liability in an AI World Mattias Andersson
2. Intro
3. Rough Agenda: • Intro & Foundation • AI/ML Data Flow • What Goes In (2) • What Comes Out (1) • How We Use It (3) • Practical Takeaways
4. MATTIAS ANDERSSON ‣ SFU Computing Science grad, 2001 ‣ Dev on Microsoft Office Shared Services ‣ Many hats in software development: architecture, security, leadership, etc. ‣ Taught hundreds of thousands to architect, build, & run cloud systems ‣ AWS Community Builder, Serverless
5. I AM NOT A LAWYER!
6. Rumsfeld Matrix
7. Foundation
8. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage
9. Ownership
21. Title • Notes for this slide
25. Types of Intellectual Property (IP) 1. Copyright 2. Patents 3. Trademarks 4. Trade Dress 5. Industrial Design Rights 6. Plant Varieties 7. Trade Secrets
26. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage
31. Training Inputs
32. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage
34. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage
37. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage
40. Output Usage
41. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage
50. I AM NOT A LAWYER!
52. Practical Takeaways
53. Re: Output Usage • Do threat modeling • Respect the randomness: Test deeper than anything else • Get proper legal advice
54. Training Data (e.g. Books) Input Data (e.g. Prompt) Trained Model (e.g. FM, LLM) Model Output Output Usage • Copyright Infringement • Bias • Defamation • Privacy • Fraud • Malpractice • Broken Trust • Copyright Infringement • Bias • Privacy • Injection Attacks • Regulatory Consider the Data Flow • Copyright Infr. • Trademark Infr. • Trade Dress Infr. • Confidential Info Leaking • Right to be Forgotten (GDPR) • Costs
55. I AM NOT A LAWYER!
56. Tool with Indemnification Known Requirements Azure OpenAI Service May need to follow Required Mitigations Microsoft 365 Copilot Year-long commitment at Business level, on top of Microsoft 365 suite GitHub Copilot Business/Enterprise (NOT Individual!) Must turn on duplication detection filter GCP Vertex AI Gemini for Google Workspace OpenAI API OpenAI Teams (2+ users) or Enterprise (call sales) (NOT Plus!) Adobe Firefly Enterprise Shutterstock Enterprise Content must pass Content Review process
57. I AM NOT A LAWYER!
58. Other possibilities • Secrecy • Make derivative works to re-protect: Embrace and extend. • Force manual-ness • The long game: Litigate and escalate
59. I AM NOT A LAWYER!
60. Wrap-Up & Resources
61. Rumsfeld Matrix
64. IP is for us (i.e. society)
66. https://www.linkedin.com/in/mattias-andersson Reach out! MATTIAS ANDERSSON