Understanding the Legal Context of AI
summary
Recent lawsuits name GitHub Copilot, Stable Diffusion, DALL‑E, Midjourney and the legal analysis ties directly to how models are built: models are matrices of weights (examples include ~160 billion parameters), training records statistical measurements not copies, and overtraining can cause memorization ~0.03–1% of outputs. Key points: fair use/TDM and Feist distinguish facts from expressive copying, APIs can retain web inputs for future training and forbid building competing models, trade secrets leak via model access, and liability often falls to the human user.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:51Who is this talk for
- 3:46Preprint version
- 4:16Models
- 4:50Analogy
- 7:27ML Training
- 13:41Model
- 16:19Training a Model
- 17:23Legal Issues
- 17:57Intellectual Property
- 27:41Confidentiality
- 38:28Questions
related talks
01h 05mNavigating the Security and Privacy Landscape of Modern AI
48mAI Impact on Cybersecurity and Privacy
29mResponsible AI Challenges and Recommendations
15mTraining the AI systems: who owns the data? Who bears the risk?
30mHow to Lie with AI: Understanding Bias, Ethics, and the Hidden Risks in ML
50mEthical Machine Learning
45mOverview of Risks, Vulnerabilities and Paths to Safer Use of Generative AI
34mRisks of AI Risk Policy: Five Lessons