AI and the Future of Copyright Politics, Pamela Samuelson

Pamela Samuelson · 2025 · 59:01 · ▶ Watch

Summary

Central question. Is generative AI just the latest in a long line of “disruptive copying technologies” that copyright owners attacked, and if so, what does that history teach about fair use, licensing, and the politics now?

Main thesis (Samuelson).

1. The panic is old. Photocopiers, VCRs, DAT and the internet each drew claims of “devastation”, mostly without empirical proof of market harm. Courts and Congress repeatedly left the answer to fair use, or to a compromise.
2. AI is different in one key way. Earlier technologies let users make consumptive copies. Training is mostly a non-expressive use, closer to Google Books.
3. The district court cases so far split. Ross sided with the plaintiff. Bartz and Kadrey mostly sided with the defendant, but disagreed on piracy and on “market dilution”.
4. The politics are new. This time “everybody’s upset”, and the White House and tech industry line up on one side. The Register of Copyrights was fired a day after a report that backed market dilution.
5. Collective licensing is attractive but hard to build. Samuelson does not see how to get there.
6. Do not treat “AI training” as one thing. Visual artists, coders, nonprofit researchers and romance novelists stand in different places.

Close. She declines to call the question: “I don’t think there’s a simple case either way”. The Q&A adds a hedge on Cox v. Sony, the idea of a tax-and-fund (AHRA-style) model, and a call to differentiate by class of author.

Stance to note. She is a long-time defender of a robust fair use doctrine, and she says so by the way she frames it, but she also grants the plaintiffs “have a point”. She filed an amicus brief for the defendant side in Ross.

Map

#	Point
1	Reid: two years on, 50+ suits; stakes in the hundreds of billions; AI is now a political subject
2	Reid: the 2011 “Can’t we all just get along” speech, framing today’s panels
3	Introduction of Samuelson
4	Roadmap: past technologies, training cases, politics, collective licensing
5	A 1982 Register of Copyrights warning that sounds like 2025
6	Photocopying: Williams & Wilkins, Texaco, Patton, CCC, CONTU
7	Home taping: Sony v. Universal, DAT, Audio Home Recording Act
8	Internet: Clinton white paper, DMCA grand compromise, Napster/Grokster, Cox pending
9	Mass digitization: Google Books found fair use; settlement rejected
10	Lessons: market harm, speculation, market failure, the circularity problem
11	Why AI differs: non-expressive use. Case count: 56, now 59
12	The two sides’ fair use arguments
13	Thomson Reuters v. ROSS (pending in 3d Cir.)
14	Bartz v. Anthropic: three buckets; piracy; $1.5B settlement
15	Kadrey v. Meta: market dilution
16	Politics: hearings, FTC, Copyright Office reports, Register fired
17	Everyone is upset; Trump administration and tech favor fair use
18	Collective licensing: four routes and their problems
19	Nonprofit researchers; spillover
20	Users: are they infringers too? (Betamax parallel)
21	Q&A: Cox v. Sony and what it means for AI
22	Q&A: romance novels and dilution
23	Q&A: P2P harm is unproven (audience pushback)
24	Q&A: tax on machines/services; transaction cost
25	Q&A: Google Books “pointing” function vs. AI; differentiate by author class; Ross amicus

Putting it together

Mental model. Copyright has been through the same cycle five times: a new copying machine, an alarm about market destruction, a thin record of actual harm, a court or Congress compromise. Generative AI repeats the cycle but breaks two links. The copying is mostly non-expressive, so the old “consumption” harm story fits poorly. And the claimed harm shifts from lost sales of one book to dilution of whole markets, which courts cannot yet measure. In 2025-26 the answer so far is split by fact pattern: copying from pirate libraries is costly ($1.5B), training on lawfully obtained books looks fair use in two N.D. Cal. courts, and a competing non-generative product lost on appeal. Secondary liability narrowed. Licensing and legislation stay open, with disclosure bills rather than payment mandates in the US.

Open tensions.

• Does a pirate source taint a fair use (Alsup) or not (Chhabria)? Still open at the appellate level.
• Is dilution a copyright harm?
• How far does ROSS reach beyond a direct competitor?
• Can anyone build a licensing market without an antitrust exemption or a mandate?
• Where does the Copyright Office stand after the leadership fight?

Reading path.

• Start: the Bartz and Kadrey summary judgment opinions; Samuelson’s own short commentary on them (not checked which).
• Then: the Copyright Office Part 3 pre-publication report; the Cox opinion; the ROSS Third Circuit opinion.
• Deeper: Litman, Digital Copyright; Lemley and Casey, “Fair Learning”; Sag on non-expressive use; Sobel on the fair use crisis; Leval 1990; the DPIIT working paper for a non-US contrast.

What changed since the talk (searched 2026-10-30)

WebSearch worked (standard mode). Only secondary sources (law-firm alerts, trackers, news) were found. I did not read court opinions. Each item says what the talk said was pending.

Pending at the talk	What happened	Source
Bartz v. Anthropic: $1.5B settlement “sought”	Final approval, 20 Jul 2026, Judge Martinez-Olguin (N.D. Cal.). Class = owners of books Anthropic downloaded from LibGen/PiLiMi (works list 482,460). All 53 objections overruled. Release covers conduct to 25 Aug 2025 only, not outputs or future conduct. Anthropic must destroy the torrented files. Fee request of 12.5% was cut; sources disagree on the final figure. Effective date depends on appeals.	Jurist, AAP, Writer Beware
Thomson Reuters v. ROSS: “now pending before the Third Circuit”	Affirmed. Reports date the decision 29 Sep 2026: headnotes copyrightable; ROSS’s training use not fair use. Described as the first US appellate ruling on fair use in AI training. ROSS was not a generative AI company, so it does not decide the Bartz/Kadrey question. Reports differ on factor-by-factor detail; I could not confirm the reasoning from the opinion.	Legal Technology, Plagiarism Today, Let’s Data Science
Cox v. Sony: argued 1 Dec	Decided 25 Mar 2026 for Cox. Contributory liability needs inducement or a service tailored to infringement; knowledge plus failure to act is not enough. Thomas wrote the opinion; sources disagree on the vote and on one calling it unanimous. The $1B award was reversed on contributory liability. Commentators say this helps AI and platform defendants on secondary-liability claims.	Mayer Brown, Cooley
Kadrey v. Meta: “whether that would be upheld on appeal, we don’t know” (; she is speaking of Alsup, but the same holds for Chhabria)	No Ninth Circuit ruling found. In June 2026 the plaintiffs asked Judge Chhabria to certify issues for interlocutory appeal (shadow-library downloading); Meta opposed. Ruling on that motion not confirmed. A new suit, Elsevier v. Meta, was filed in S.D.N.Y.	ChatGPT is eating the world, McKool Smith tracker
Register fired; “now before the Supreme Court”	On 30 Jun 2026 the Supreme Court denied the government’s stay application (Blanche v. Perlmutter, 25A478), so Perlmutter stays in office for now. This is not a merits ruling. A law-firm alert says the “final” Part 3 training report could now be formally published; I could not confirm it has been.	Lowenstein
NYT v. OpenAI/Microsoft (not named in the talk, but part of the “50+ suits”)	No 2026 merits ruling found. A coalition of about 400 local newspapers sued OpenAI and Microsoft in S.D.N.Y. on 24 Jun 2026.	Law360 case page (headline-level only), OpenClassActions (plaintiff-side tracker)
Authorship (Reid frames “copyrightability”)	Thaler v. Perlmutter: cert denied 2 Mar 2026. Human-authorship requirement stands; how much human input is enough is open.	Jones Day, Baker Donelson
”Congress is completely dysfunctional” on a compulsory license	Congress has moved on transparency, not licensing. TRAIN Act (H.R. 7209, 22 Jan 2026): subpoena route for owners. CLEAR Act (S. 3813, 10 Feb 2026, Schiff/Curtis): training-data notice to the Register. Neither advanced beyond committee referral in what I found. California AB 2013 took effect 1 Jan 2026.	Dean/Moran release, IP & Legal Filings on CLEAR
Collective licensing “I just don’t know how we get there”	The Bartz settlement is a de facto class payment for the piracy claim only. Licensing for training itself is still voluntary and deal-by-deal in the US (no source found for 2026 licensing deals)†.	as above

Not found / not checked. Outcomes in Concord v. Anthropic and Disney v. Midjourney (search returned nothing). The UK Getty v. Stability AI ruling is reported as a loss for Getty, with training-lawfulness left undecided (one aggregator, not verified). The EU position was not searched.

India (only where a source supports it). The DPIIT committee’s first working paper, “One Nation One License One Payment” (8 Dec 2025), proposed a hybrid model: a mandatory blanket licence for AI training on lawfully accessed works, with a statutory remuneration right and government-set royalties. This is a policy route Samuelson calls unlikely in the US. Reports differ on details; the paper has a second part on outputs. Sources: Medianama, Bar & Bench, Open Future copy of the paper. I did not check what happened after the consultation.

Part 1: Before you read

Key terms

• Fair use (17 U.S.C. §107). A defence to infringement judged on four factors: purpose and character of the use, nature of the work, amount used, effect on the market. The first and fourth carry most weight in practice.
• Transformative use. A use that adds a new purpose or meaning rather than substituting for the original. Watch out: “transformative” in Alsup’s and Chhabria’s rulings turns on purpose (“a very different purpose”,), not on whether the output looks different. After Warhol v. Goldsmith (2023, from memory) it is tied to whether the use shares the original’s purpose.
• Consumptive vs. non-expressive use. Her split. Consumptive: you get a whole copy and enjoy the expression. Non-expressive: you “analyze the relationship among words in text and images”. The idea comes from Matthew Sag’s “non-expressive use” (from memory).
• Market harm / market substitution. Whether the use supplants demand for the original.
• Market dilution. A newer theory, from the Copyright Office report and Chhabria: AI floods the market with competing works, even non-infringing ones, and undercuts incentives for human authors. Watch out: this is a different harm from “your book was copied, so I sold fewer copies”.
• The circularity problem. “You didn’t license it from me, so you harmed my licensing market.” Courts call this circular.
• Contributory infringement. Secondary liability for helping someone else infringe. Direct infringement is doing the copying yourself. Samuelson notes most AI cases plead direct infringement.
• Statutory damages. Up to $150,000 per work for willful infringement; the source of the “hundreds of billions” in Reid’s stakes.
• Collective licensing. A body licenses many works at once. Four routes in the talk: voluntary; legislated compulsory; class settlement; court order.
• Section 1201 / DMCA. The anti-circumvention rules; the safe harbours (§512) are the “grand compromise”.

People and cases named (reading list)

Name	Why it comes up	Start with
David Ladd	1982 Register of Copyrights: copyright “under stress” from new machines	His 1982 speech†
Williams & Wilkins v. U.S.	Photocopying by NIH librarians held fair use; 4-4 at the Supreme Court	Court of Claims opinion
American Geophysical Union v. Texaco	Commercial photocopying unfair; birth of CCC licensing	2d Cir. 1994
Cambridge UP v. Patton	Course reserves mostly fair use	11th Cir. 2014/2018
Sony v. Universal (1984)	Time-shifting fair use; 5-4 (she says “one vote away”)	The opinion; Litman on its politics
Audio Home Recording Act (1992)	Levy-and-fund compromise	17 U.S.C. ch. 10
Authors Guild v. Google	Mass digitization fair use; settlement rejected	2d Cir. 2015
Thomson Reuters v. ROSS	First AI fair use ruling (D. Del.)	Judge Bibas, Feb 2025 (from memory)
Bartz v. Anthropic (Alsup); Kadrey v. Meta (Chhabria)	The June 2025 rulings	Both opinions; Samuelson’s own commentary
Cox v. Sony Music	ISP secondary liability	SCOTUS opinion, 25 Mar 2026
Gigi Sohn	2011 “Can’t we all just get along” speech	Reid’s framing
Copyright Office Part 3 report (May 2025)	Endorsed market dilution; Register fired the next day	The pre-publication PDF

Samuelson’s own frame

• Look at history before you predict. Each wave of copying technology produced the same alarm and the same lack of data.
• Separate the technology from the dataset. Piracy of 7 million books is a different question from training on lawfully acquired works.
• Do not treat everyone the same. “I think what we see just in these two or three decisions so far is differentiation”.
• Admit the plaintiffs’ point. See.

Current debate (as of 2026-10-30)

The three fault lines Samuelson draws now have partial answers.

1. Is training fair use? The district courts in N.D. Cal. said mostly yes (2025). The first appellate ruling, in ROSS, said no for a non-generative competitor (Legal Technology). Whether the Third Circuit’s reasoning reaches generative models is disputed. The Ninth Circuit has not ruled. See the table above.
2. Is piracy the real liability? The Anthropic settlement approved in July 2026 prices it at $1.5B, but releases only past conduct and not outputs (Jurist). Anthropic’s own counsel points to the court’s fair use finding as the law that still stands.
3. Who answers for users and platforms? Cox narrowed secondary liability in March 2026 (Mayer Brown). That weakens any contributory theory against AI developers or their users, which is the spillover Samuelson was asked about.
4. Who runs the Copyright Office? The leadership fight continues, and the Part 3 report remains unfinalized as far as I could confirm (Lowenstein).
5. What will Congress do? Disclosure bills, not licensing mandates (IP & Legal Filings).

The ideas

This is the depth layer. Samuelson names few theorists; most of what follows is the background she assumes. Nothing in this part is her claim unless a timestamp says so.

How the traditions meet

 incentive theory (1) ----> "is training harm to incentives?" ---------+
                                                                       v
 fair use / transformative (2) --> purpose test --> Alsup, Chhabria, Bibas
                                         ^                |
 non-expressive use (3) -----------------+                |
                                                          v
 market harm / dilution (4) ------> 4th factor <---- piracy (Bartz) 
                                         |
 secondary liability (5) <--- users, platforms (Cox) 
                                         |
 collective licensing (6) <-- "who pays?" --> political economy (7)
 authorship (8) ---- separate branch (outputs, not training)

---

Where each school of thought is covered:

• Why copyright exists: incentive theory and the “technology panic” cycle: idea 1
• Fair use doctrine and the transformative turn: idea 7
• Non-expressive use and “fair learning”: idea 2
• Market harm, circularity and dilution: idea 3
• Secondary liability and the platform question: idea 5
• Collective licensing, levies and statutory schemes: idea 4
• The political economy of copyright: idea 6
• Copyrightability and authorship: idea 2

1. The 50-year cycle of copying panics

In short. Samuelson opens with a 1982 quote that copyright is under stress and under attack, and notes that it sounds familiar. Photocopying, the VCR and the internet each set off the same kind of panic. Congress mostly kicked the can down the road or struck a compromise, as with the AHRA and the DMCA. Her lessons are that harm to the market is what matters, that speculation is not enough, and that circular arguments about lost licensing fees are baloney.

🎙 In the talk.

• She opens with a 1982 quote and reveals it was Register David Ladd. “Copyright today is under stress. It’s also under attack.”. Her comment: “I just think that sounds familiar, right?“.
• Photocopying: the Court of Claims upheld NIH library copying 4-3, and the Supreme Court split 4-4. Texaco went the other way for commercial researchers and the CCC became the licensing body. Congress “kicked the can down the road”.
• VCRs: “We were one vote away from even making time shift copies of television programs was unfair”. The AHRA followed as a compromise: a noncommercial exemption plus a tax into a fund.
• Internet: the Clinton white paper wanted strict liability for service providers; the treaty route failed; the DMCA traded broad anti-circumvention rules for safe harbours.
• Lessons: harm to the market is what matters; speculation is not enough; market failure may mean no harm; circularity is “baloney”.

📚 Deeper theory.

• (Roots) The pattern fits Litman’s account: legislative deals are struck between industries, often after a court gives a signal.
• Samuelson’s “one vote away” framing stresses contingency. The law could have gone the other way in Sony and in Williams & Wilkins.
• The AHRA hard-drive exemption shows technology routing around levy schemes. That is evidence that a tax model can miss the real devices.
• The lessons are doctrinal (tradition 4) not just historical.

📚 Roots: Why copyright exists: incentive theory and the “technology panic” cycle

Problem. If copying is cheap, who pays for creation? Copyright answers: a temporary monopoly. The US Constitution frames it as promoting “progress” (Art. I, §8, cl. 8). Chain (dates from memory)†.

• Statute of Anne (1710) and the Copyright Clause (1787): the incentive rationale, as opposed to natural-rights authorship.
• Macaulay’s 1841 speech on copyright: a “tax on readers” justified only as far as it buys creation.
• The twentieth-century cycle: player pianos (White-Smith, 1908) to radio to photocopy to VCR to Napster. Each time owners claim catastrophe and each time Congress or courts mediate.
• Jessica Litman, Digital Copyright (2001): copyright law is made by negotiating industries; the public is absent.
• William Landes and Richard Posner, The Economic Structure of Intellectual Property Law (2003): the economic case for broad rights and limits. Feeds the talk. Her whole first half is this cycle. Ladd’s 1982 quote is the cycle’s voice. Unresolved. Whether AI is the next iteration or breaks the pattern, which is Reid’s framing question.

Where others push back. Rightsholder voices argue that the “no harm” record rests on bad data and that AI is a different order of copying. Some scholars (e.g. in the Copyright Office report) hold that generative models compete with the works they trained on in a way VCRs did not.

What they leave out. Non-US law, and the fact that the Betamax-era harm story was never tested empirically (she partly concedes this).

Check yourself.

1. Which two earlier cases did she say were decided by a single vote, and how?
2. What does the “circularity problem” say, and why do courts reject it?
3. Why does the AHRA hard-drive exclusion matter for a levy on AI?

2. The training cases: ROSS, Bartz, Kadrey

In short. She compares the training cases. The plaintiffs say it is exact copying of whole creative works, and that outputs compete. The defendants say training is highly transformative because they only care about works as data. ROSS went against fair use. Bartz and Kadrey found training transformative. They differ on piracy and on dilution. Alsup said authors are not entitled to control licensing their works as training data, period.

🎙 In the talk.

• Plaintiff’s theory: exact copying, “you’re not transforming them to make anything new”, creative works, whole works, outputs compete.
• Defendant’s theory: highly transformative because of a different purpose; “we only care about works as data”; licensing markets impossible.
• ROSS: the AI system “doesn’t embody or output Westlaw headnotes”, whose outputs are judicial opinions that are not protectable. The judge ruled against fair use; the case was then in the Third Circuit.
• Bartz: three buckets. Training copies are “quintessentially transformative”; buying and scanning books is fair use; pirated copies are not. “Authors are not entitled to control licensing their works as training data, period”.
• Kadrey: also highly transformative; Meta’s output filters mattered; plaintiffs failed to show dilution; Chhabria said a better record might change the result.
• Differences: piracy (Chhabria says it “doesn’t make any difference at all”; Alsup says “definitely not fair use”) and dilution (“science fiction” per Alsup).

📚 Deeper theory.

• (idea 7, 3) ROSS is a competing-product case: ROSS built a rival legal search tool. That fits the market-substitution reading, which the Third Circuit has now affirmed (Legal Technology). Do not read it as a ruling on chatbots.
• Alsup’s and Chhabria’s disagreement is the central open question: does the source of the copy (piracy) taint a fair use, and does indirect substitution count? (idea 3)
• The $1.5B figure is a settlement of the piracy claim, not a verdict on training (Jurist).
• District rulings do not bind other courts. Until the Ninth Circuit speaks, the Bartz/Kadrey reasoning is persuasive only.

📚 Roots: Non-expressive use and “fair learning”

Problem. Many technologies copy works to extract facts or patterns, not to enjoy them. Chain.

• Early case law on reverse engineering of software (Sega v. Accolade, 1992; Samuelson herself wrote on this, from memory).
• Perfect 10 v. Amazon (2007) and A.V. v. iParadigms (2009): copying for indexing or detection.
• Matthew Sag, “Copyright and Copy-Reliant Technology” (2009) and later work on “non-expressive use”.
• Mark Lemley and Bryan Casey, “Fair Learning” (2021): machine learning should be treated like human learning.
• Counter-view: Benjamin Sobel, “Artificial Intelligence’s Fair Use Crisis” (2017), says generative models may reproduce expression, so the data-mining analogy fails. Feeds the talk. Her key distinction and her remark that AI is “more like the Google Books cases”. Unresolved. Memorization and regurgitation of training text, and whether a model “contains” the work.

📚 Roots: Copyrightability and authorship

Problem. Can a machine-made work be owned? Chain. Burrow-Giles v. Sarony (1884, from memory), Naruto v. Slater (9th Cir. 2018), Thaler v. Perlmutter (D.C. Cir. 2025; cert denied 2 Mar 2026), Copyright Office guidance on AI-assisted works. Feeds the talk. Only at Reid’s opening. Samuelson does not discuss it.

Where others push back. Plaintiffs’ scholars say training involves wholesale reproduction and that models memorize. Defendants’ scholars say the cases are right on purpose but the piracy holding is too harsh (Chhabria’s view). Some argue Alsup’s “authors are not entitled to control training licences” overreaches.

What they leave out. Output-side claims (regurgitation, lyrics, images); non-US courts; the DMCA §1202 (copyright management information) claims.

Check yourself.

1. Why does ROSS count as a plaintiff win under her scheme, and why might it not govern ChatGPT?
2. What exactly did Alsup and Chhabria disagree on (two points)?
3. Name one fact in Kadrey that Chhabria said might change the outcome in another case.

3. Market dilution and the romance novelists

In short. Dilution is the worry that AI will flood the market with competing works and undermine incentives for human authors. The Copyright Office’s May report endorsed it, and the register was fired the next day. Samuelson asks which romance authors would be diluted, and whether everybody would be. She says there is no empirical information to answer that.

🎙 In the talk.

• Dilution defined as “a concern that AI technologies will be used to flood the market with competing works and undermine incentives for human authors”.
• The Copyright Office’s May report endorsed it, and “the register got fired the next day”.
• On romance: “which of the romance authors are getting diluted? Does everybody get diluted?“. She says “we don’t have any kind of empirical information” to answer.

📚 Deeper theory.

• (Roots) Dilution moves the fourth factor from your book to your genre. It is close to an antitrust or labour-market theory, not a copyright-substitution one.
• It strains the rule that fair use protects against substitution, not against competition from non-infringing works (an argument often made in this debate; from memory).
• Alsup’s “science fiction” remark and Chhabria’s invitation show the split is about evidence as much as law.

📚 Roots: Market harm, circularity and dilution

Problem. The fourth factor can swallow the rest: any unlicensed use could be said to destroy a licensing market. Chain.

• Harper & Row v. Nation (1985): the market effect is “the single most important element” (from memory).
• Texaco (1994): the licensing market (CCC) counts, which feeds circularity.
• Courts then limit “traditional, reasonable, or likely to be developed” markets (Texaco; Authors Guild v. Google).
• Empirical caution from the file-sharing literature, e.g. Oberholzer-Gee and Strumpf (2007), who found little effect of P2P on sales (from memory, contested). Aram’s pushback belongs here.
• Newer: “market dilution”, argued by Chhabria and by the Copyright Office in 2025. Feeds the talk.. Unresolved. How you would measure dilution; she says there is no empirical information.

Where others push back. Creator groups say waiting for proof lets the market be damaged before the law responds. Others say creators already see lost commissions (her own Berkeley logo example,).

What they leave out. Any study of output volume, e.g. the share of self-published titles that are AI-assisted.

Check yourself.

1. How does dilution differ from ordinary substitution?
2. Why did Chhabria rule against Kadrey even while endorsing dilution?
3. What evidence would you need to prove dilution for one genre?

4. Collective licensing

In short. Collective licensing appeals because 180 million registered OpenAI users show the public values the product, so it seems fair that authors get compensated. She lists four routes: voluntary collective licensing, a legislated licence, a class settlement, and a court-ordered licence. The problems are practical. Authors don’t have the data, publishers do, and AI firms think they need data but not rights.

🎙 In the talk.

• Appeal: 180 million registered OpenAI users show the public values the product, and “isn’t it fair that they get some sort of compensation?“.
• Four routes: voluntary collective licensing (backed by the Copyright Office and industry groups); legislated compulsory licence (“Congress is completely dysfunctional”); class settlement (the failed Authors Guild v. Google deal); court-ordered licence (binds only parties).
• Problems: who to ask; no ownership data; publishing contracts; “the authors don’t have data. The publishers have the data”; the AI firms “think they need data. They don’t think they need rights”; thin US collective infrastructure outside music; antitrust exemption; a licensing mandate becomes an entry barrier.
• Q&A: a tax on machines (AHRA) or on service revenue; “$0.50 to everybody or even $50” is expensive; but “it’s not very sympathetic to say… I copied billions of things… so I get to have it for free”.

📚 Deeper theory.

• (Roots) A levy on revenue sidesteps the identification problem, but not the allocation problem.
• The EU approach pairs an opt-out TDM exception with disclosure duties†.
• India’s DPIIT proposal chose the compulsory route she dismissed for the US (see “What changed”). That is a real contrast, not a refutation, because India’s legislative setting differs.
• The Bartz settlement shows a class mechanism works when the class is defined by one concrete act (downloading a known pirate library). It does not generalize to the open web.

📚 Roots: Collective licensing, levies and statutory schemes

Problem. Transaction costs are too high for one-by-one licences of billions of works. Chain.

• ASCAP/BMI and the consent decrees (music; from memory), the strongest US collective regime. Samuelson says the US is otherwise “undeveloped”.
• Copyright Clearance Center (1978) for photocopying, after CONTU.
• Audio Home Recording Act (1992): levy on devices and media, paid into a fund.
• Nordic “extended collective licensing” (from memory): a society’s licence extends to non-members.
• EU: the 2019 DSM Directive (Art. 4 text-and-data-mining exception with opt-out) and the AI Act’s transparency duties†.
• Antitrust: collective pricing needs an exemption. Feeds the talk.. India’s DPIIT paper is a compulsory-licence model of this kind (see above). Unresolved. Who is paid, and by what formula.

Where others push back. Some creators want opt-in only, not any compulsory scheme. Some AI developers argue a mandate entrenches incumbents (her entry-barrier point).

What they leave out. Licensing deals already struck between AI firms and publishers (no source checked here); non-text media.

Check yourself.

1. Why does she say publishers, not authors, hold the key data?
2. What is the “entry barrier” problem?
3. Why did the AHRA levy miss hard drives?

5. Users, Cox, and secondary liability

In short. She asks whether users also infringe by using these technologies, as in the Betamax fight, where Universal’s theory made every copier an infringer. On Cox, the question is the scope of secondary liability. She does not predict an outcome. Many claims against AI developers are contributory, but most of these cases count on direct infringement claims.

🎙 In the talk.

• The Betamax parallel: Sony’s lawyer opened by noting 5 million households had VCRs; Universal’s theory made every copier an infringer.
• Her question: “are the users also infringements by their use of the technologies?“.
• On Cox (Q&A): the question is the scope of secondary liability; “material contribution plus knowledge” supported the billion-dollar award; the Solicitor General sided with Cox and cited Taamneh. She does not predict an outcome.
• Spillover: many claims against AI developers are contributory, but “most of these cases actually are counting on direct infringement claims”.

📚 Deeper theory.

• (Roots) The Supreme Court did decide for Cox on 25 Mar 2026, in an opinion that requires inducement or a service tailored to infringement; knowledge alone is not enough (Mayer Brown, Cooley).
• That fits the staple-article logic of Sony and makes user-liability theories against AI users harder.
• It does not touch direct infringement in training, the main theory in the AI cases.

📚 Roots: Secondary liability and the platform question

Problem. When do you answer for what others do with your tool? Chain.

• Sony (1984): a “staple article” with substantial non-infringing uses is safe.
• Grokster (2005): inducement.
• DMCA §512 (1998): safe harbours.
• Twitter v. Taamneh (2023): “substantial assistance” required in the aiding-and-abetting setting.
• Cox v. Sony (2026): see “What changed” above. Feeds the talk.. Unresolved. Whether output-side claims against model makers, or users, survive Cox.

Where others push back. Commentators note the opinion is short and skips earlier circuit law (Bloomberg Law commentary found by search). Rightsholders worry about piracy platforms.

What they leave out. Vicarious liability. The Court denied Sony’s petition on that point, so it was not decided (Mayer Brown summary; not verified in the opinion).

Check yourself.

1. Why does the Betamax household count matter to her point about users?
2. What two things did the Cox jury finding rest on, per her account?
3. Why is Cox less relevant to the main AI training claims?

6. Politics: “everybody’s upset”

In short. Politics so far are less intense than in earlier fights, with a few hearings and some reports. The Trump administration and tech are gung ho about fair use. Earlier fights had one aggrieved industry, but now everybody’s upset. Nonprofit researchers may be hurt by broad rulings, and the defendants have an incentive to push for a for-profit versus nonprofit distinction.

🎙 In the talk.

• Politics so far are “less intense” than in earlier fights: three or four hearings, an FTC report, three Copyright Office reports.
• Trump administration and tech are “gung ho about fair use”.
• Earlier fights had one aggrieved industry. “Now, everybody’s upset”.
• Nonprofit researchers: students and graduate researchers train models on in-copyright works; broad rulings could hurt them; the defendants “have any incentive to push for a distinction between for-profit and nonprofit”.

📚 Deeper theory.

• (Roots) The White House position matches Reid’s quote about “a race to achieve global dominance”, a framing that moves copyright into industrial policy.
• The Perlmutter episode (firing, D.C. Circuit order, Supreme Court stay denial in June 2026) is a separation-of-powers fight as much as a copyright one (Lowenstein).
• Since the talk, Congress has put forward transparency bills, which sit between the two camps (IP & Legal Filings).

📚 Roots: The political economy of copyright

Problem. Who writes the rules? Chain.

• Litman again: copyright is bargained among industry groups.
• Lawrence Lessig, Free Culture (2004): the cost of regulation by incumbents.
• The 2012 SOPA/PIPA protests: the public as a political actor.
• Reid’s 2011 Gigi Sohn story: the rhetoric of the “copyfight”. Feeds the talk. The shift from a two-sided fight to “everybody’s upset” and a three-way mix of White House, tech and creators. Reid’s list of divides: “creative laborers and industrial rightsholders”, “tech companies and their users”, “branches of the US government”. Unresolved. Whether this realigns the coalitions on non-AI fights (the second panel’s question).

Where others push back. Some say the politics are as intense as ever, just fought in courts and agencies. Others say the real split is between creators and distributors, not creators and tech.

What they leave out. International competition on AI rules; state laws (e.g. California AB 2013).

Check yourself.

1. Why is “everybody’s upset” new?
2. What point does she make about nonprofit developers?
3. Which branch conflict does Reid name?

7. Differentiate, don’t generalize

In short. Samuelson says not to treat everybody the same. All fair use or all unfair use is probably not exactly right. Visual artists are more at risk than coders who posted source code free, and software has a thinner scope of protection. Headnotes in ROSS, if protected at all, are super thin. Asked about the Berkeley logo, she says AI does a pretty good job, but asks whether there is a causal claim.

🎙 In the talk.

• Audience: Authors Guild leaned on the “pointing function”. Why is it missing from the AI analysis?.
• She answers by asking whether to “treat everybody the same”, and says all fair use or all unfair use… is probably not exactly right.
• Visual artists are “more at risk” than coders who posted source code free; software has “a thinner scope of protection”.
• ROSS amicus: headnotes, if protected at all, are “super thin”; she rejects the judge’s marble-sculptor comparison.
• Berkeley logo example: AI does “a pretty good job”, but is there a causal claim?.

📚 Deeper theory.

• (Roots, 4) Copyright’s “thin” protection for factual or functional works (merger, originality limits) means the second factor varies by genre.
• The Third Circuit has now affirmed on headnote copyrightability, which cuts against her amicus position on that point (Let’s Data Science; coverage differs on reasoning). Treat this as a fact to re-read from the opinion.

📚 Roots: Fair use doctrine and the transformative turn

Problem. A flexible defence needs a test that predicts. Chain.

• Folsom v. Marsh (1841): the origin of fair use.
• Pierre Leval, “Toward a Fair Use Standard” (1990): the transformative-purpose test.
• Campbell v. Acuff-Rose (1994): parody; the test becomes law, and commercial use is not presumptively unfair.
• Authors Guild v. Google (2d Cir. 2015): scanning to build a search index is fair use; the snippet and pointing function matter.
• Warhol v. Goldsmith (2023): transformativeness is weighed against the original’s purpose and commercial substitution. Feeds the talk. Both sides’ arguments are this test applied. The audience question about the “pointing function” is a challenge to whether Google Books still fits. Unresolved. Whether “different purpose” alone, without a search-like benefit, is enough.

Where others push back. Rule-of-law concern: a fact-by-fact fair use regime is unpredictable for developers. Others say any categories would be gamed.

What they leave out. How to set categories in practice; she offers no scheme.

Check yourself.

1. Why does she call visual artists more exposed than coders?
2. What was the amicus argument about headnotes?
3. What did the later Third Circuit ruling do to that argument (per reports)?

---

About the source and how this page was made

• Speaker: Pamela Samuelson (keynote); Blake E. Reid (welcome, Q&A host)
• Event: “AI and the Future of Copyright Politics”, Silicon Flatirons conference, University of Colorado Law School, Boulder
• Date: 2025-11-07 (event page date; video uploaded 2025-11-21)
• Uploaded: 2025-11-21
• Length: 59:01
• Channel: Silicon Flatirons
• Video: AI and the Future of Copyright Politics | Welcome & Keynote

Speakers

Label	Who	Role	Confidence
Reid	Blake E. Reid	Associate Professor, Colorado Law. Welcomes, introduces, hosts Q&A	High (named in the description and the talk)
Samuelson	Pamela Samuelson	Richard M. Sherman Distinguished Professor of Law and Information, UC Berkeley. Keynote	High (named by Reid; description)
Audience 1	”Adam” (first name from Reid,)	Student; asks about Cox v. Sony	Medium
Audience 2	”Aram” (first name,)	Pushes back on P2P harm claim; says he has published on it	Medium
Audience 3, 4	Unnamed	One asks about user liability and enforcement; one is a teacher of Authors Guild, Bartz, Kadrey	Low (names not given)

Caption warnings (Apify auto-captions, English). Corrections used here:

• “Cahn case” = Cahn v. Sony (the digital audio tape suit; from memory).
• “Cambridge University Press versus patent” = Cambridge University Press v. Patton.
• “Williams and Wilkins versus United States” is right.
• “Audio Home Recording Act was packed in 1992” = enacted.
• “Twitter versus Taamneh” is right (Twitter v. Taamneh, 2023).
• “Ross Intelligence”/“Ross” = ROSS Intelligence. “Thomson-Reuters” = Thomson Reuters.
• “Cox v. Sony” is written both ways by the audience; the case is Cox Communications v. Sony Music Entertainment.
• “[INAUDIBLE]” is in the captions.
• “Gigi Sohn… Fred Von Lohmann… Preston Padden” match the 2011 episode Reid describes.

Possible factual slips in the talk (all minor; flagged for checking, not asserted):

• Samuelson dates Williams & Wilkins to “1974”. From memory, the Court of Claims ruled in 1973 and the Supreme Court split 4-4 in 1975. Not checked.
• “25,000 Westlaw headnotes”. From memory, ROSS’s training material (the “Bulk Memos”) drew on about 25,000 headnote-based questions, but the court’s fair use ruling covered about 2,200 headnotes. She may mean the larger set. Not checked.
• “three reports” from the Copyright Office. The Part 3 training report was a pre-publication version in May 2025 (see search results below).
• Her case counts (“56… 59 now”) are a snapshot from Nov 2025. Reid says “more than 50”. Both are now out of date.

---

🎙 marks what the speaker says. 📚 marks added background and the wider debate, not the speaker’s claim. † marks background from general knowledge that was not checked against a source.