Understanding the Legal Context of AI
summary
Recent lawsuits name GitHub Copilot, Stable Diffusion, DALL‑E and Midjourney and challenge training on copyrighted material while courts treat training, model and outputs separately. Models store weights as huge numeric matrices not expressive text, trained by statistical prediction; memorization can produce infringing outputs and researchers extracted 108 near‑duplicates from a 90,000‑image probe (far rarer in larger models). Legal arguments invoke Google Books fair use analogies but outputs that reproduce specific inputs remain actionable; indemnity and copyright office positions are unsettled.
AI-written summary · about