Copyright
Format. Break into groups of 4–5. Each group picks one of the two breakouts below (or takes both if time allows). Spend ~5 minutes skimming the prep reads, then ~10 minutes debating. A designated reporter brings the group’s position — and any dissents — back to the full class for a ~3-minute report-back.
Chapter 4.2 argues that copyright was not designed as a censorship regime but routinely functions as one: DMCA notice-and-takedown lets bad-faith claimants suppress lawful speech faster than the speaker can defend it, and automated matchers (Content ID, Rights Manager) cannot perform fair-use analysis. Both breakouts probe where the line between enforcement and censorship actually sits.
Breakout A: Automated Takedown — Should False Claims Carry Real Costs?
Motion. “Every party issuing a DMCA takedown (including automated Content ID matches) should face statutory penalties for demonstrably false claims — no more free shots at other people’s speech.”
Prep reads (5–10 min).
- YouTube Processed 2.5 Billion Content ID Copyright Claims in 2025 — TorrentFreak, 2026. YouTube’s own transparency data: rightsholders monetized 90% of Content ID claims, and >6% of webform DMCA notices were “likely false” — 10x the abuse rate of any other tool.
- YouTube Copyright Transparency Report — Google, ongoing. Google’s own dashboard: 12.8M Content ID disputes, and uploaders won 67% of the ones they contested in 2025.
- A Guide to YouTube Removals — Electronic Frontier Foundation, updated 2025. EFF’s running critique of keyword-based and purely automated takedowns and how Content ID sits outside the DMCA counter-notice regime.
- Takedown Hall of Shame — Electronic Frontier Foundation. Ongoing case file of specific abusive takedowns (Fox/”Homeland,” SmellyOctopus, etc.) that instructors can use as concrete examples in class.
Discussion prompts.
- The WSJ 2020 investigation found Google had processed over 4 billion URL removal requests. What error rate is acceptable at that scale? Is 1% false positives 40 million pieces of wrongly-suppressed speech, or is it just the cost of doing business?
- DMCA Section 512(f) technically allows penalties for knowingly false takedowns, but has almost never been enforced. Would putting real teeth in 512(f) (statutory damages, no scienter requirement, fee-shifting) chill legitimate copyright enforcement — or just chill abusive enforcement? Who bears the risk today, and who would bear it under your proposal?
- Content ID is not the DMCA — it is a private contractual system YouTube built to keep Viacom-style lawsuits away. Should platform-built matchers be regulated as if they were DMCA notices, or is that a category error? Would forcing them into the DMCA regime just push platforms to take down more, not less?
- The Cox verdict ($1B in 2019 for failing to disconnect repeat infringers) points the incentive in one direction; the SIGCOMM peer-review-takedown episode points the other. Design an intermediary-liability rule that survives both incidents.
Bring back. Your group’s proposed penalty structure for false takedowns, and one bright-line example of what should not count as a takedown-eligible claim at all.
Breakout B: Sci-Hub, Shadow Libraries, and the Ethics of Bulk Defiance
Motion. “Sci-Hub, LibGen, and other shadow libraries are a legitimate — even necessary — response to a scholarly-publishing system that paywalls publicly funded research.”
Prep reads (5–10 min).
- Publishers Ramp Up Pressure vs. Anna’s Archive, Sci-Hub, Z-Library & Libgen — TorrentFreak, February 2025. Coordinated escalation across jurisdictions — publisher suits, ISP blocks, and CDN/payment-processor pressure aimed at the shadow-library stack.
- Anna’s Archive Lawsuits: $322M Ruling and Active Cases — LegalClarity, 2026. Overview of the Spotify/UMG/Sony/Warner $322M default judgment, the separate 13-publisher $19.5M judgment, and the OCLC WorldCat ruling.
- The plausible consequences for Anna’s Archive, NVIDIA, and any other AI developer that trained on shadow libraries — Substack analysis, 2025. Traces the through-line from Bartz v. Anthropic ($1.5B settlement, September 2025) to the Meta LibGen disclosures to NVIDIA’s pending Anna’s-Archive claims.
- Anna’s Archive (Wikipedia) — updated 2026. Useful one-page overview of domain seizures (.org, .se), the Spotify metadata scrape, and the interaction with AI-training litigation.
Discussion prompts.
- Publicly funded research is often paywalled behind $30–$40 per-article fees. Is Sci-Hub civil disobedience, ordinary infringement, or a public-interest utility? Does the answer change if the researcher is at a well-funded U.S. university vs. a hospital in Nairobi?
- The 2024 Meta and Anthropic rulings held that training on copyrighted books was fair use, but settlements have hinged on whether the training corpus was pirated (LibGen/Books3). Is it coherent to say Sci-Hub is illegal but training on Sci-Hub is fair use? Who benefits from that gap? (See Bartz v. Anthropic for the underlying case.)
- The EU Article 17 approach effectively requires platforms to filter uploads against rights-holder databases. Would an “Article 17 for research” — filter every upload against Elsevier’s catalog — be a legitimate response, or does it prove the point that copyright at scale becomes a speech regulator?
- Compare shadow libraries to Napster (shut down), YouTube-DL (initially DMCA’d, then reinstated after EFF pressure), and Google Books (upheld as fair use). What actually predicts whether a system survives — the law, the users’ political sympathy, or the defendant’s legal budget?
Bring back. Your group’s ranking of these access regimes from most to least legitimate: subscription paywalls, green open-access, gold open-access, Sci-Hub. Justify the top and bottom.
Instructor notes
These breakouts map to the Section 4.2 takeaways on DMCA-as-censorship-tool and on automated matching’s inability to perform fair-use analysis. Breakout A is the natural home for students who came in wanting to argue about YouTube demonetization (redirect that energy to Lecture 13); keep them on the takedown side of the pipeline here. Breakout B tends to divide students along whether they see themselves primarily as future creators (pro-copyright), future researchers (pro-access), or future platform engineers (they will tell you both sides are unworkable). If time is short, Breakout A produces a cleaner report-back; Breakout B produces the better cross-lecture connections to Chapter 4.5 on generative-AI copyright.
Breakout C: Should Training a Model on a Work Require a Licence?
In Bartz v. Anthropic (2025), the court held that training a language model on books the company had bought was fair use, but that keeping a library of pirated copies of those same books was not; Anthropic then paid $1.5 billion to settle the piracy half. So the current US position is: the training is lawful, the way you got the books may not be. This breakout asks whether that line makes sense.
Motion. “Copyright law should treat training a generative model on a work the same way it treats copying that work: licence it or leave it alone.”
Prep reads (5–10 min).
- Court Grants Final Approval of $1.5 Billion Anthropic Copyright Settlement — Authors Guild, July 2026. Judge Alsup’s June 2025 ruling found training on lawfully bought books fair use but keeping a pirated library not; the settlement (final approval July 20, 2026) pays for the library, not the training.
- Meta wins on fair use for now, but court leaves door open for “market dilution” — Authors Alliance, June 2025. Judge Chhabria: fair use on this record, but a plaintiff who proves the model floods the market for books like theirs “will often decisively win”.
- Signed, Sealed, Affirmed: Third Circuit Backs Thomson Reuters over ROSS — Copyright Lately, October 2026. The first appellate ruling on AI training: copying Westlaw headnotes to train a competing legal-research tool was not fair use, “no more than an ordinary copyright case”.
- GEMA v. OpenAI: Europe’s Direction on AI Infringement? — William Fry, November 2025. Munich: lyrics memorized in the model’s weights are a reproduction; the text-and-data-mining exception does not cover them.
Discussion prompts.
- Alsup, Chhabria, and the Third Circuit each decided on the fourth factor (market harm) and reached different places. Which market counts: the market for the book, the market for licensing books as training data, or the market the model’s outputs compete in?
- “Sci-Hub is illegal, but training on Sci-Hub’s corpus is fair use.” Who benefits from a rule where the training is lawful and only the acquisition is punished? Who can afford a $1.5 billion settlement, and who cannot?
- Munich says a memorized lyric in the weights is a copy; London says a model that stores no copies is not an infringing copy. Both can be right on their facts. Does the answer turn on what the model contains or on what it emits?
- From the censorship side: a licensing regime means a handful of catalogue owners decide which works a model may learn from. Is that a market, or is it the filter-bubble problem at the training stage?
Bring back. Your group’s rule in one sentence, and the case above it would have decided differently.