#55 - AI & Law & Order : Copyright in the Age of AI

Link to Folder with Everything

2026.09.23.officehours.55

Recording (Audio)

Google Meet didn't capture video this session, so here's the full audio instead — press play to listen right here.

Listen on Google Drive

Slides

Lili's full deck — 34 slides covering the case map, the Source/Use/Output/Value framework, and links to the IntelCraft articles on each case.

Download the slides (PDF)

Summary

Lili Kazemi — a lawyer with 20+ years of practice focused on valuing IP across value chains — walked through the state of AI copyright litigation in the US. Her framing question for the whole hour: when AI works, who gets paid? Who created the underlying value, who owns the resulting capability, and what does AI owe its sources?

The Wild West

There is no AI-specific copyright statute and no Supreme Court ruling. Courts are applying the 1976 Copyright Act — written twenty years before the internet — to roughly 150 active cases across the country, case by case. Outcomes turn on individual facts and judicial interpretation, and copying alone is not infringement: courts weigh purpose, amount taken, and market harm, with "transformative use" as the AI companies' central defense.

A framework, not a verdict

Lili's analytical lens: trace Source (how the developer obtained the material), Use (what was done with it), Output (what the user actually receives), and Value (who captured the return). Intermediate copying during training can matter legally even when the output looks entirely new. The doctrinal lineage runs from Napster through Google Books — which was never a blank check — to today's generative AI cases.

The cases

Bartz v. Anthropic: training held highly transformative and scanning purchased books fair, but building a permanent library from pirate sources was not excused — settled for $1.5B, so no appellate precedent. Kadrey v. Meta: courts want concrete evidence of market harm, not speculative licensing markets. ROSS: the output contained no Westlaw headnotes, but training a directly competing legal-search product still lost on key points. Doe v. GitHub: generating code from statistical patterns of open source was not a DMCA violation. News publishers are suing over paywall bypass and output reproduction, and visual-arts cases involve characters like Batman and Superman — with some AI companies already negotiating licenses instead.

What practitioners should do

Know your data sources and permissions before using them. Document provenance, workflow, and the human decisions along the way — if you ever have to defend how you got somewhere, records win. Assume AI inputs are stored, discoverable, and not privileged, so keep sensitive material out. Keep a substantial human creative role (a prompt alone isn't enough) if you want copyright in AI-assisted work. And since fair-use factors are weighed rather than checked off, nobody can promise you'd win — expect private licensing deals to settle most of this before the courts do. Lili's closing theme, echoed by Rahul: your brand is the strongest protection you have; copyright and patents support it, they don't replace it.

Next steps

• [Rahul Singh] Share the deck and recording with participants — this post and the Drive folder above.
• [Lili Kazemi] Open to hosting deeper focused sessions on specific copyright cases if there's interest — reach her on LinkedIn or Substack.
• [IntelCraft] Fix the recording setup for next session — Meet's video recording was stopped early this time, so only audio (Granola) and transcripts were captured.
• The deck links out to the IntelCraft articles on each case for further reading.

Details

• Session: officehours 55, September 23, 2026. Rahul Singh hosting, Lili Kazemi presenting. Part of the recurring IntelCraft law series.
• Speaker background: Lili is not a copyright litigator by trade — her practice centers on identifying, valuing, and pricing IP across value chains, which shapes her "who gets paid" approach to the AI copyright fight. She has been learning the case law with AI as a research partner, going to the original sources.
• Scope: US cases only (there is parallel litigation in the UK, Germany, and India). The landscape is mostly generative AI today and will get more complex with agentic and enterprise AI.
• Content types in play: books, news, code, images, music, and character likenesses — the diversity of content matters more than the case count, because each type produces different claims (ordinary infringement, fair use, DMCA 1201/1202, contract, publicity rights).
• Q&A: Rahul asked whether a unique sequence of prompts, skills, or training verbs is protectable. Lili's answer: copyright protects expression, not process — processes are patent territory, and teaching style alone isn't infringement unless someone takes the unique expression, for the same purpose, in the same market. Brand and trademark remain the most durable business protections.
• This summary was written by Claude from the Granola and Fireflies transcripts (both in the Drive folder), since Gemini notes weren't generated this session.