Talks

Generative AI Meets Copyright — Pamela Samuelson

42 min read

Pamela Samuelson · 2023 · 52:05 · ▶ Watch

Summary

Central question. Does generative AI break copyright law, or does copyright law break generative AI? She splits it into three questions: (1) is copying works as training data infringement; (2) when are outputs infringing derivative works; (3) who, if anyone, owns copyright in outputs.

Main thesis.

  • Training: a “pretty good fair use case” for data on the open internet, but not settled. It will take years to decide.
  • Outputs: usually not infringing, because they are not “substantially similar” to any one input. The exception is memorised, much-duplicated material (her Mickey Mouse example).
  • Ownership: the oldest question (since the mid-1960s). Now the answer is that purely AI-generated output has no human author, so it has no US copyright. A human who adds selection, arrangement or text can still get protection for that part.
  • Bigger frame: copyright is “the only law that’s already in existence that could bring generative AI systems to their knees”. She does not think that will happen, but wants the audience to see it as a live risk.

Close. Years before courts answer. Global AI governance (EU AI Act, NIST framework, US AI Bill of Rights) does not yet reach this. The Copyright Office is under-equipped: no technologists, no economists.

Stance. An expert who tries to be even-handed. She lists both sides, but her own leanings show: she expects training to be fair use, expects Thaler to lose, and thinks “co-pilot” is a good metaphor for human authorship in AI-assisted works.

Map — talking points

#Point
1Host intro: series, speaker bio
2Why it matters: hundreds of millions use ChatGPT; lawyers’ practice is “overwhelmingly” generative AI advice
3Four lawsuits: Copilot ($9B asked) and two against Stability AI. Two claims: ingestion infringes; outputs are infringing derivatives
4Moral panic: the Future of Life letter, “Marxist nightmare”, the “stolen art” cartoon
5Three questions for the talk
6Copyright in two minutes: fixation, term, exclusive rights, idea/expression, fair use factors
7Training data and fair use: Field v. Google, Authors Guild v. Google, the counterarguments, licensing markets
8Derivative works: statute, “substantially similar”, Copilot 1% match, Mickey Mouse, dedup and output filters
9Getty v. Stability: 12 million photos, captions, logo mangling, licence claims
10Ownership of outputs: Thaler, Kashtanova, the Office policy statement
11Models and training data as property
12Conclusion: years to resolve, AI Act, NIST, “existential threat”
13Q&A: CGI and Disney, Copyright Office capacity
14Q&A: “Mona Lisa playing basketball”, prompts as authorship, co-pilot metaphor
15Q&A: private data, human authorship standard
16Q&A: user liability (strict liability)
17Q&A: Copilot class action, open-source developers’ complaint
18Q&A: lawmakers, fair use view, stretching “derivative”
19Q&A: licence conflicts, Creative Commons, data not copyrightable
20Q&A: ethics; host’s close (patents, other disciplines)

Putting it together

A mental model. Copyright guards expression, not ideas, facts or data. A generative-AI system has four touchpoints with it: the input (copies made to train), the model (a statistical artefact of doubtful copyrightability), the output (usually new, occasionally a near-copy), and the user (who may be an infringer and may or may not be an author). In April 2023 she judged inputs probably fair use, outputs mostly non-infringing, and pure machine output unowned. By October 2026 the third has hardened into law (no authorship without a human), the second has hardened at the edges (memorised items and logos still bite), and the first has split into a fact-heavy contest: how the data was obtained and whether the model competes with the works it learned from now decide cases. Settlement and licensing, which she mentioned as an emerging market, have become the main resolution mechanism.

Open tensions.

  • Transformative learning vs market substitution: the Bartz and Kadrey rulings both favoured the AI firms but for different reasons, and Chhabria warned his own result is fragile.
  • Human authorship vs assisted creation: the line between “tool” and “author” is still case-by-case.
  • A court-made rule vs a legislative one: she doubts the Office and courts are equipped; the 2025 dispute over the Register shows the institutional risk.
  • Source matters: lawful acquisition (the Bartz split) was not in her talk but now carries much of the weight.
  • No appellate ruling yet on fair use for training a generative model.

Reading path.

  • Start: the Copyright Office, Part 2 press note, and a short account of Bartz and Kadrey.
  • Then: Authors Guild v. Google (2d Cir. 2015); Lemley and Casey, “Fair Learning”; Sag, “Copyright Safety for Generative AI”†.
  • Deeper: Campbell and Warhol opinions; Leval 1990; Carlini et al. on extraction; Samuelson’s own papers on AI and copyright†; the unsealed 3d Cir. Ross opinion (LawNext).

Unverified in this page. Allen v. Perlmutter outcome; Disney/Universal v. Midjourney; NYT v. OpenAI after June 2026; Getty US after April 2026; the Supreme Court’s June 2026 order on Perlmutter and whether Part 3 is final; all items marked†.

Then and now (the quick answer)

Question the talk raisesState in April 2023 [talk]State on 30 Oct 2026 [later]
Is training on in-copyright works fair use?Open. She leans yes for open-internet dataSplit, no appellate answer for generative AI. Two N.D. Cal. rulings (June 2025) for AI firms on training; one appellate ruling (3d Cir., Sept 2026) against an AI firm, but on a non-generative legal-search tool. Details in idea 1
Doe v. GitHub (Copilot, $9B claim)Filed; copyright claims centralCopyright claims largely gone early; 9th Cir. affirmed dismissal of the DMCA §1202(b) claim, 16 Sept 2026
Andersen v. Stability AIEarly stageStill in litigation; summary judgment motions now due 17 Feb 2027
Getty v. Stability AI (“the strongest one”)Early stageUK: Getty largely lost (4 Nov 2025); permission to appeal the secondary-infringement point. US: refiled in N.D. Cal.; copyright and trademark claims survived dismissal (April 2026)
Thaler v. PerlmutterPending in D.D.C.; “I don’t think he’s gonna win”He lost at every level. Supreme Court denied cert, 2 Mar 2026. Human authorship is required
Kashtanova / Office policy statementOne policy statementOffice report Part 2 (29 Jan 2025): prompts alone do not give authorship; human selection, arrangement and modification can
Are users liable for infringing outputs?Yes, strict liabilityStill true as a doctrine; no major test case found
Copyright Office capacityNeeds a tech officer and economistOffice issued Part 3 on training (pre-publication, May 2025); the Register was removed the next day; fight over her removal reached the Supreme Court (details unverified)
RegulationEU AI Act pending; US has noneEU AI Act in force†; US still no AI copyright statute found

Part 1 — Before you read

Key terms

  • Copyright subject matter / “original work of authorship.” The law protects expression that a human author fixed in a medium. Copyright attaches automatically, with no registration. Registration matters mainly for suing.
  • Idea/expression dichotomy. “Copyright only protects the original expression in works of authorship, not their ideas, facts or methods”. Everything in the training debate turns on whether a model takes expression or only patterns.
  • Fair use. Watch out: not the same elsewhere. A US defence with four open-ended factors (purpose and character, nature of the work, amount taken, market effect). Other countries use closed lists of exceptions (UK “fair dealing”, the EU text-and-data-mining exceptions, Japan’s Art. 30-4). “Fair use” means different things in different legal systems.
  • Derivative work. A “work based upon one or more pre-existing works” that recasts, transforms or adapts it. Courts require “substantial similarity” of protected expression, not just a causal link.
  • Substantial similarity vs “based on”. Being based on a work is not enough to infringe. Being substantially similar in expression is.
  • Memorisation / regurgitation. A model reproducing near-copies of training items. Her example is Mickey Mouse, which appears in many training images.
  • Deduplication and output filters. Two engineering defences she names: remove duplicates from training data; filter outputs.
  • Copyright management information (CMI), DMCA §1202. Author names, titles, licence terms and logos attached to a work. Removing or altering them knowingly is its own violation, separate from infringement. (Section number from memory; consistent with the 9th Cir. coverage below.)
  • Work made for hire. Thaler’s theory: the AI is his “employee” so he owns what it makes.
  • Strict liability. Copyright infringement needs no intent or knowledge.
  • Training data vs model weights vs outputs. Three separate objects in law: inputs (copied works), the model (a statistical artefact), outputs (generated items). Each has a different copyright question.
  • Text and data mining (TDM). Computational analysis of large corpora. She says it “has been widely thought to be fair use”.

People and cases named (a reading list)

NameWhy it comes upStart with
Ken GoldbergHostHis Berkeley page
Stephen ThalerClaimed copyright for “Creativity Machine” artThaler v. Perlmutter (D.C. Cir. 2025)
Kris KashtanovaComic Zarya of the Dawn with Midjourney artCopyright Office decision letter, Feb 2023†
Authors Guild v. GoogleGoogle Books scanning was fair use2d Cir. 2015 opinion (Judge Leval)
Field v. GoogleSearch caching was fair useD. Nev. 2006
Getty ImagesPlaintiff against StabilityComplaint (D. Del., 2023)
Andersen v. Stability AIArtists’ class actionN.D. Cal. docket
Doe v. GitHubProgrammers’ class action over CopilotN.D. Cal., then 9th Cir.
Future of Life InstituteSix-month pause letterThe open letter (Mar 2023)
Joey Gonzalez, Rod Brooks, John SchulmanNamed in the intro as other speakers or contactsNot relevant to the legal argument

Samuelson’s own frame

She has written on computers and copyright since the 1980s and says so (“something I wrote… it was actually pretty good”). Her habit is to trace each “new” AI question to an older one: authorship to the 1960s, TDM to Authors Guild v. Google, and software copyright to the 1980s fight over whether code is expression. Her working method is analogy to precedent plus a sceptic’s eye on what the Copyright Office can actually handle.

The current debate (searched 30 Oct 2026)

  • Training and fair use is still unsettled, but now there are rulings.
    • Bartz v. Anthropic (N.D. Cal., Judge Alsup, June 2025): training on lawfully acquired books was fair use; keeping millions of pirated copies in a central library was not. The case settled for at least $1.5 billion (announced 5 Sept 2025). Judge Martínez-Olguín, who took over the approval stage from Alsup, granted final approval on 20 July 2026 (Jurist, Kilpatrick Townsend, TNW).
    • Kadrey v. Meta (N.D. Cal., Judge Chhabria, 25 June 2025): summary judgment for Meta on fair use for these plaintiffs only, because they had no evidence of “market dilution”. The judge said dilution will often win the fourth factor in other cases (Authors Alliance, Manatt).
    • Thomson Reuters v. Ross (3d Cir.): reports date a panel ruling to late Sept 2026 that affirmed the lower court for Thomson Reuters and rejected Ross’s fair use defence. It is described as the first federal appellate ruling on fair use for AI training data. Ross is a legal-search tool, not a generative model. The opinion was issued under seal on 29 Sept and released shortly after; Plagiarism Today puts this at 30 Sept. LawNext also reports that ROSS says it will go to the Supreme Court (LawNext, ruling, LawNext, opinion, Plagiarism Today). I did not read the opinion text.
  • Output ownership has a clear answer, for now. The Supreme Court denied review in Thaler v. Perlmutter on 2 Mar 2026, leaving the D.C. Circuit’s human-authorship rule intact (Jones Day, Baker Donelson, Sterne Kessler). The open question is how much human input is enough.
  • The Copyright Office has spoken in two reports. Part 2 (copyrightability, 29 Jan 2025) and Part 3 (training, pre-publication, 9 May 2025). See idea 3 and 5.
  • Image-model cases drag on. Andersen and the US Getty case are not resolved. The UK Getty case mostly went to Stability (Cleary Gottlieb, PetaPixel).
  • Settlement and licensing have become the market answer. The Anthropic settlement works out at roughly $3,000 per work, covering up to 500,000 works (Kilpatrick Townsend). This fits her remark about emerging licensing markets.

The ideas

She spoke for 28 minutes to an engineering audience and gave the law in shorthand. Behind that shorthand sit seven bodies of thought. Read this section slowly; the deep dives refer back to it. The explanatory passages below are background, not her claims, and cases and dates in them are† unless linked.

 AUTHORSHIP & THE AUTHOR-FUNCTION ──┐                ┌── FAIR USE LINEAGE
 Statute of Anne · Sarony · Foucault│                │   Folsom · Sony · Campbell
 CONTU · Naruto · Thaler            ▼                ▼   Google Books · Warhol
                          ┌────────────────────────────────┐
 MEMORISATION & ML ─────► │ SAMUELSON: where does the law  │ ◄──── NON-EXPRESSIVE USE & TDM
 Carlini · dedup · filters│ meet the model: input, weights,│       Sag · Lemley & Casey · EU/Japan
                          │ output, user                   │
                          └────────────────────────────────┘
                              ▲             ▲             ▲
        LICENCES, CMI, DATA   │   TECH-vs-COPYRIGHT       │  POLITICAL ECONOMY & GOVERNANCE
        GPL · CC · DMCA 1202  │   HISTORY                 │  Lanier · Lessig · Office capacity
        sui generis DB right  │   photocopier·VCR·Napster │  EU AI Act · NIST · Warhol-era courts

Where each school of thought is covered:

  • The architecture of copyright: expression, originality, derivative works: idea 2
  • The fair use lineage: idea 1
  • Non-expressive use and text-and-data mining: idea 1
  • Authorship: the human author as a legal fiction: idea 3
  • Memorisation, extraction and the engineering of “non-infringing” models: idea 2
  • Licences, contracts, CMI and data: idea 4
  • Technology, copyright and political economy: idea 5

1. Training data and fair use

In short. Plaintiffs claim that training on their works infringes copyright and that outputs are infringing derivative works. Samuelson weighs the fair use precedents, Field v. Google and Authors Guild v. Google. Authors did not consent, and Stability makes new images that compete. Against that, only expression is protected, text-and-data mining has been widely thought fair use, and fair use gives breathing room for new creation. She thinks ingesting open internet data is a pretty good fair use case.

🎙 In the talk

  • The two claims: ingestion “is an infringement of the copyright in the works” and outputs “are infringing derivative works”.
  • Her precedents: Field v. Google and Authors Guild v. Google. Defendants say crawling for training is the same kind of use.
  • The case against: Google “was making easier for people to find the copyright owner’s works”, whereas Stability “make new images that compete”. Authors “didn’t consent”, cannot “easily opt out”, and the training data’s value “is what makes the generative AI turn out good stuff”. “Garbage in, garbage out”.
  • Survey: 90% of Authors Guild members “think that AI developers should pay”.
  • Markets: “there are also some licensing markets that are emerging” and that bears on factor four.
  • Her counterweights: only expression is protected; “bag of words”; TDM “has been widely thought to be fair use”; the constitutional purpose is to promote progress; fair use gives “breathing room for new creation”.
  • Her view in Q&A: “I think it’s a pretty good fair use case for ingesting data that’s on the open internet”. Breaking through a paywall “would be a factor that would cut against fair use”.

📚 Deeper theory (see Roots)

  • She tests training against factors one and four. Courts since 2023 put more weight on factor four and on how the data was acquired. That is why Bartz split lawful purchase from pirated copies.
  • Her paywall remark has been borne out in part: Alsup’s central-library holding turned on pirated sources, not on training. (link to 🆕 below)
  • The “emerging licensing markets” point is circular: if courts find fair use, there is no market to harm; if courts find harm, a market exists. Critics call it bootstrapping.†
  • Transformativeness after Warhol (2023) is narrower than the one she could assume in April 2023 (Warhol came a month later).

🆕 Since the talk (searched 30 Oct 2026)

  • Bartz v. Anthropic: training on lawfully acquired books was fair use; pirated library copies were not. $1.5B settlement; final approval 20 July 2026 (Jurist). A district-court ruling, not binding elsewhere.
  • Kadrey v. Meta: fair use for Meta on a thin record; Chhabria’s dicta say dilution of the market for human works will often defeat fair use (Authors Alliance).
  • Thomson Reuters v. Ross: 3d Cir. against the AI firm, Sept 2026, on a non-generative tool (LawNext). Not yet tested for generative models.
  • New York Times v. OpenAI/Microsoft: not in the talk (filed Dec 2023). Motions to dismiss largely denied in March 2025; in June 2026 the Times amended its complaint and narrowed claims (Hoodline, Giga Law). Sources conflict on details. No trial date or ruling found after June 2026. Unverified.
  • Net: no appellate court has yet ruled on fair use for training a generative model. Her “years before courts are able to find definitive answer” still holds.

📚 Roots: The fair use lineage

The point. Her list of reasons for and against fair use is a summary of this lineage. The Authors Guild v. Google distinction she gives (Google helped people find the books; Stability makes competing images) is the core of the later Kadrey “market dilution” idea.

The problem. Rigid copyright would stop criticism, parody, search and learning. Fair use is the safety valve, and its edges are decided case by case. Chain.

  • Folsom v. Marsh (1841, Story J.): the origin of the multi-factor test.
  • Sony v. Universal (1984): the VCR is lawful because it has substantial non-infringing uses; time-shifting is fair. It shows courts protecting a new technology from a copyright holder who would have stopped it.
  • Harper & Row v. Nation (1985): taking the “heart” of an unpublished memoir was not fair, even though small. It supports the “amount and substantiality” factor, and shows that market harm weighs heavily.
  • Judge Pierre Leval, “Toward a Fair Use Standard” (1990): the idea of transformative use. Adopted in Campbell v. Acuff-Rose (1994, 2 Live Crew parody).
  • Intermediate-copying cases: Sega v. Accolade (1992), Kelly v. Arriba Soft (2003), Perfect 10 v. Amazon (2007), Field v. Google (2006), Authors Guild v. HathiTrust (2014), Authors Guild v. Google (2015). She uses the last two as the closest precedents.
  • Google v. Oracle (2021): copying API code was fair, in part because it let programmers reuse skills. Andy Warhol Foundation v. Goldsmith (May 2023, after this talk): narrowed “transformative” when the new use shares the original’s commercial purpose. This is why post-2023 rulings spend so much time on factor one and four.

What the four factors really do. Section 107 (1976) codified judge-made law and gave no weights. In practice, two factors do the work: purpose (factor one) and market effect (factor four). From Campbell until Warhol, “transformative” became close to decisive. Critics said it had swallowed the derivative right, since nearly any adaptation can be called a transformation. Warhol pulled it back by asking whether this specific use shares the original’s purpose. So the April 2023 framing she gave came just before the doctrine shifted.

Three kinds of market harm. This is where Bartz and Kadrey part ways. Classic harm is substitution: my copy replaces your copy. A second kind is lost licensing: you would have paid me. It’s the circular one, flagged in idea 1. The third is dilution: your model floods the market with new works that are not copies of mine, but which compete with mine. Older cases did not clearly recognise that third harm, and competition from unlike works is normally lawful. Judge Chhabria put preserving the incentive for human creation at the centre, disagreeing with Alsup. If appellate courts accept dilution, generative training looks very different from search. If they reject it, Samuelson’s 2023 lean looks right.

📚 Roots: Non-expressive use and text-and-data mining

The point. This is her “bag of words” argument. The counterweight is her point that the scale and competitive output differ from classic TDM.

The problem. A program that “reads” a book to count words, or to learn grammar, does not read it as a human would. Does copying for that purpose even engage the author’s interest? Chain.

  • Matthew Sag, “Orphan Works as Grist for the Data Mill” (2012) and later “The New Legal Landscape for Text Mining and Machine Learning” (2019): the term non-expressive use.†
  • Lemley and Casey, “Fair Learning” (2021): machine learning needs access to existing works to function, so copying for learning should be fair unless the output replicates expression.†
  • Sag, “Copyright Safety for Generative AI” (2023): training is likely fair if the model is built to avoid regurgitation.†
  • Law outside the US: EU DSM Directive (2019) Arts. 3–4 allow TDM, with a rights-holder opt-out for commercial TDM; UK has a non-commercial TDM exception; Japan Art. 30-4 is broad; Singapore has a computational-analysis exception.†

The argument and its weak joint. The non-expressive theory runs as follows. Copyright protects expression as humans perceive it. A program counting words perceives none. So the copy is a technical step that never reaches the protected interest. It works cleanly for search and for digital humanities, where the output is a fact about the corpus. Generative models strain it, because the output is itself expressive, and it is of the same kind as the input. The theory’s defenders reply that the output’s expression comes from the prompt and from general learning, and the test should apply only when it reproduces specific expression. The human analogy, that people learn by reading, is rhetorically strong but legally weak. A person reading makes no copy, so the reproduction right is never engaged, whereas training does reproduce.

Two regulatory models. The US asks after the fact, through an open standard, whether a use was fair. The EU and Japan set a rule in advance. The EU route adds an opt-out, which turns the question into one about machine-readable signals. Each exports its own problem: the US has years of uncertainty, and the EU has a compliance question about whether an opt-out was respected. See mammen-ai-impact-on-ip-laws for GEMA v. OpenAI and other cases outside the US.

Where others push back. Authors, artists and publishers say the “non-expressive” label ignores that models are sold as substitutes for the training works. Fair-use defenders say that a ruling for plaintiffs would make large-scale learning impossible without a licence cartel. The Copyright Office’s Part 3 (pre-publication) took a middle line: some uses fall within fair use, but commercial models that compete with their sources may go beyond it (Patently-O, Digital Camera World).

What the talk leaves out. Piracy and source of acquisition as a separate issue; non-US law (EU opt-outs, UK, Japan); the position of the model weights as a possible “copy”; trade-secret and contract claims; foreign plaintiffs; and the effect of her own prediction that licensing markets would matter.

Check yourself

  1. Why does she think Authors Guild v. Google helps the defendants, and what does she say distinguishes it?
  2. How does the “licensing market” point cut both ways under factor four?
  3. After Bartz, why does it matter whether training data was lawfully acquired?

2. Outputs and derivative works

In short. A derivative work is one based on pre-existing works, and courts require a substantially similar amount of original expression. So most generative AI outputs will not be substantially similar to the input data. The edges differ: a model may memorize heavily duplicated works like Mickey Mouse, and one study found code matching training data in about 1 percent of outputs. The Getty case adds trademark and metadata claims.

🎙 In the talk

  • A derivative work is “a work based upon one or more pre-existing works”; courts require “a substantially similar amount of original expression”.
  • So “for the most part, generative AI outputs… are not gonna be substantially similar to the data that was input”.
  • Copilot: one study found code matching training data “in about 1% of the outputs”.
  • Mickey Mouse: with many duplicates the model “may memorize that” and Disney “might have a claim”.
  • Mitigation: dedup, output filters.
  • Getty: 12 million photos plus captions; Stable Diffusion outputs with a “mangling of the Getty logo”; trademark and CMI claims; “Getty is willing to license”.
  • Q&A: a stylistic prompt (“Mona Lisa… in the style of Pixar”) raises how to disentangle text prompt from learned images. On style: she does not answer directly there; she moves to Kashtanova.
  • Anderson [sic] complaint admits outputs are “unlikely to be substantially similar”, which helps the defence; plaintiffs will “stretch that concept of derivative work”.

📚 Deeper theory (see Roots)

  • Style is not protected expression in US law; the question of whether a trained model can reproduce a living artist’s style is separate from copying a particular work.
  • Memorisation research means her general point about substantial similarity is empirically right on average but fails for duplicated items.
  • Secondary liability (contributory and vicarious) matters for platforms; she covers only direct liability.†
  • The “output” question has a user side: see idea 5.

🆕 Since the talk

  • Doe v. GitHub (Copilot): the 9th Circuit (16 Sept 2026) affirmed dismissal of the DMCA §1202(b) claim. It held that AI-generated code is a new work, not a stripped copy, and rejected the district court’s “identicality” framing; the “input” theory (training on stripped CMI) was forfeited, so it remains open (Ropes & Gray, Gibson Dunn, EFF). The claim for $9 billion mentioned did not survive in this form. Remaining open-source licence claims: not checked.
  • Getty v. Stability, UK: High Court, 4 Nov 2025. Getty abandoned its primary infringement claims during trial; lost the secondary infringement claim (model as “infringing copy”); won narrowly on trademark for logos in outputs. It has permission to appeal on secondary infringement (PetaPixel, Cleary Gottlieb). This matches her note that the logo mangling “doesn’t bode well”, though in the UK it helped Getty on trademark.
  • Getty v. Stability, US: refiled in 2025. In a ruling reported on 24 April 2026, Judge Trina L. Thompson (N.D. Cal.) let the copyright, trademark infringement, unfair competition and dilution claims proceed. Bloomberg’s page does not show which count, if any, was dismissed. Later status not found (Bloomberg Law).
  • Andersen v. Stability AI: still pending. Older reports give a trial date of 8 Sept 2026, but the Feb 2026 schedule change moved summary judgment motions from 4 Nov 2026 to 17 Feb 2027 (Chat GPT Is Eating the World). The docket was not checked.
  • Disney/Universal v. Midjourney (filed June 2025): not searched.†

📚 Roots: The architecture of copyright: expression, originality, derivative works

The point. Every training and output argument is a dispute about where the line between pattern and expression sits. Defendants say a model learns patterns. Plaintiffs say the output is a recast of expression.

The problem. What exactly does the law own? If it owned ideas, no one could write a second detective novel. Chain.

  • Statute of Anne (1710) gave authors a short, tradeable right in copies of books. The US Constitution (1787) ties copyright to “the progress of science and useful arts” (the “constitutional purpose” she cites).
  • Baker v. Selden (1879): a book’s description of a bookkeeping system is protected, the system is not. This is the origin of the idea/expression line.
  • Feist v. Rural Telephone (1991): facts and a phone-book list lack originality; “sweat of the brow” is not enough. It underlies her point that “data as such is not within the scope” and that a bare compilation is “not copyrightable at all”.
  • Nichols v. Abie’s Irish Rose (L. Hand, 1930) and the “abstractions” test. Arnstein v. Porter (1946) set up the ordinary-observer test. These are the roots of “substantial similarity”.
  • 1976 Act §101 defines derivative work with a list of examples, closing on “recast, transformed, or adapted”. She notes courts read that closing phrase narrowly in most cases.

Why the line is hard to hold. The idea/expression split is a policy tool that looks like a fact. Learned Hand’s abstractions test admits this: take away detail from a story and you reach a level of generality that is free for all, but no one can say exactly where. For human copying that vagueness is tolerable, because a jury compares two works. A model breaks the comparison. Its weights hold no single work to compare, yet they are shaped by all of them. So the law’s main instrument, side-by-side similarity, cannot be applied to the object that plaintiffs most want to attack. That is why litigation keeps moving to the two places where a comparison is possible: the training copies and the outputs.

How the derivative right grew. The derivative right wasn’t always there. In Stowe v. Thomas (1853), a German translation of Uncle Tom’s Cabin was held not to infringe, since copyright covered the book as written. Congress then added translation and dramatisation rights, and the 1976 Act widened the list to a general right over works “recast, transformed, or adapted”. Each expansion let owners capture a new market near the work. The AI question is whether “based upon” stretches to a statistical artefact. Her point that courts still demand substantial similarity is the brake on that stretch.

📚 Roots: Memorisation, extraction and the engineering of “non-infringing” models

The point. The Mickey Mouse point and the 1% Copilot figure. It is why she says the outputs are “unlikely to infringe” in general but can infringe at the edges.

The problem. If a model can reproduce a training item, it has stored something. How much, and how do you prove it? Chain.

  • Carlini et al., “Extracting Training Data from Large Language Models” (2021) and “Extracting Training Data from Diffusion Models” (2023): shows that models, especially on duplicated items, emit near-copies.†
  • Lee et al., “Deduplicating Training Data Makes Language Models Better” (2021): duplicates raise memorisation.† This is the basis of her “remove duplicates” remark.
  • Output filters and licensing-aware code filters in Copilot.†
  • Legal link: if a model produces a near-copy, the plaintiff has a clean infringement claim on that output, and some argue the weights hold a “copy”. Defendants say weights are statistics, not copies.

Memorisation is a spectrum, and the law wants a switch. Machine learning aims to generalise, and memorising is usually a training failure. But it is not cleanly separable: the same process that learns what a mouse looks like can learn one much-repeated mouse. The law, meanwhile, asks a yes-or-no question: is a copy present? Courts have since divided on how to treat this. The UK Getty court held the model isn’t an infringing copy; the German GEMA court is reported on the Mammen page as going further for lyrics. Engineers can reduce the risk through deduplication and filtering, but these are probabilistic. That’s the deep reason her “mostly non-infringing” remains accurate and also unsatisfying.

The proof problem. Memorisation is usually shown by adversarial prompting. Defendants say a plaintiff who works hard to extract a copy shows nothing about ordinary use. Plaintiffs say a system that holds the work has the work, however hard it is to get it out. The same debate from another angle is whether an output filter cures infringement or just hides it.

Where others push back. Plaintiffs: the model itself is the derivative work, since it is “based upon” the training set; “substantial similarity” is the wrong test for a model that encodes thousands of works. Defendants: that reading would make every learner an infringer.

What the talk leaves out. Style mimicry and right of publicity; voice and likeness; output licences from vendors; indemnities that vendors offer to enterprise users†.

Check yourself

  1. Why is “based on” not enough for a derivative-work claim?
  2. Why does duplication in training data raise legal risk?
  3. After the 9th Cir. Doe ruling, which GitHub-related theory is still open?

3. Who owns the outputs?

In short. The question of who owns AI outputs is old. In Thaler, the Copyright Office said the work lacked human authorship, and Samuelson does not think he will win. In Kashtanova, the Office left copyright only in the text and in the selection and arrangement of the images. She reads these rulings as being about missing human authorship, not about infringing outputs. Whether prompts amount to authorship may end up in court.

🎙 In the talk

  • The oldest question: people have argued “since the mid 1960s”; she wrote on it in the 1980s.
  • Thaler: the Creativity Machine is the author, and he owns the copyright “because the software was like an employee” — a work for hire. The Office said no (“lacks human authorship”); he sued; “I don’t think he’s gonna win”.
  • Kashtanova: the Office first registered, then cancelled, and left copyright only in “the text and the selection and arrangement of the images”.
  • Office policy statement: AI-generated works lack human authorship; applicants must identify and disclaim AI parts.
  • Her reading: not denied “because the outputs are infringing derivatives, but just that they lack human authorship”.
  • Q&A: prompts as authorship: “I went through a series of prompts”; the Office said no; it “may end up in the courts”. The co-pilot metaphor “might be something that would support copyright”; Europe is going that way.
  • Is “human authorship” codified? She says the statute is silent and lawyers “build upon existing precedent”.
  • Film and CGI: the Office “is going to have to visit and revisit”.

📚 Deeper theory (see Roots)

  • The human-authorship rule comes from case law (Sarony, Naruto) and Office practice, not from an express statutory sentence. Her Q&A answer is correct on that.
  • “Public domain” is her inference. Legal commentators note the AI part is uncopyrightable but an unlicensed copy could still breach a platform’s terms of service.†
  • The practical frontier is degree of control: prompts are like instructions to a commissioned artist, who is not the author of what the artist makes.

🆕 Since the talk

  • Thaler v. Perlmutter: lost in district court (2023) and the D.C. Circuit (2025); cert denied 2 Mar 2026. The D.C. Circuit noted that the rule does not bar copyright in work “made by or with the assistance of artificial intelligence” and did not decide how much human input suffices (Baker Donelson). Her prediction was right.
  • Copyright Office, Part 2 (29 Jan 2025): prompts alone are not enough; AI as a tool does not bar protection; human-authored works containing AI elements stay protectable; no new sui generis right needed; case-by-case (Copyright Office, Copyright Alliance, Authors Guild). This answers her “revisit and revisit” point. It also answers the Kashtanova prompt-sequence argument in the negative.
  • Allen v. Perlmutter (Midjourney, Théâtre D’opéra Spatial, D. Colo.): not in the talk. Summary judgment motion filed 25 Aug 2025; a tracker expected a ruling in the second or third quarter of 2026; I found none (McBrayer tracker). Unverified.
  • The co-pilot metaphor has partly survived: assisted work is protectable, but only for the human share.

📚 Roots: Authorship: the human author as a legal fiction

The point.. Her “work for hire” discussion shows Thaler trying to route around the author problem by making the machine an employee.

The problem. Copyright was built around a person who creates. When a machine creates, who is the author? Why does the law insist on a human? Chain.

  • Burrow-Giles v. Sarony (1884): a photograph can be authored, because the photographer chose pose, lighting and arrangement. This is the view her abstract gives as the early consensus: AI “is just a tool, like a camera”.
  • Romantic authorship: Woodmansee (1984), Jaszi (1991), and Foucault’s “What Is an Author?” (1969) argue that “the author” is a historical construct tied to property.
  • CONTU (1978), the US commission on computer uses of copyrighted works, said the computer is a tool, so the user is the author. Samuelson wrote “Allocating Ownership Rights in Computer-Generated Works” (1986), her piece.† She argued against giving the program or its maker rights in the output.†
  • UK Copyright, Designs and Patents Act 1988 s.9(3) gives the “person by whom the arrangements necessary for the creation are undertaken” authorship of computer-generated works, a rare statutory answer.†
  • Naruto v. Slater (9th Cir. 2018, the monkey selfie): animals lack statutory standing. The Copyright Office Compendium (3d ed.) states the human-authorship rule in registration practice.†
  • Thaler v. Perlmutter: district court (Aug 2023), D.C. Cir. (Mar 2025), cert. denied (Mar 2026).

Why work for hire was a clever choice, and why it failed. Work for hire is the law’s own admission that “author” can be a fiction. A corporation, which has never held a pen, is the legal author of what staff make. Thaler’s move: if a company can be an author by law, the human-author rule can’t be absolute. The reply is that the doctrine transfers authorship from a human who created to the employer. It presupposes a creative employee, and it relies on agency concepts that do not fit software. CCNV v. Reid (1989) tied “employee” to the common law of agency. So the fiction redirects authorship but does not create it.

The commissioned-artist analogy. The prompt debate rests on Reid too: a person who commissions a sculpture and gives detailed instructions is not its author, the sculptor is. If that’s what prompting is, the prompter owns nothing. Kashtanova’s and Allen’s position is that iteration changes the picture, because the human steers each round. The Office’s reply in Part 2 is that prompts do not yet give enough control. The qualifier matters: one summary says “based on current AI technology”. That leaves room for the answer to move as tools give finer control.

Why insist on a human at all? Two rationales pull apart. The incentive theory says we need no copyright for machines, because machines need no motive. The personality theory says a work expresses a self, and a machine has none. Both exclude machines, but for different reasons, and they differ on the owner of the machine, who does respond to incentives. The UK and India answer differently, by naming a human arranger or the person who causes the work to be created. See the Part 2 traditions on the Mammen page, mammen-ai-impact-on-ip-laws.

Where others push back. Some argue prompting plus iterative selection is real authorship (the Kashtanova and Allen position). Others argue for a UK-style rule giving the arranger rights. The Office and the courts reject both for now.

What the talk leaves out. Patents and inventorship (the host flags this); moral rights; non-US rules; contract ownership between vendor and user.

Check yourself

  1. What exactly was Kashtanova allowed to keep?
  2. Why did Thaler pick a “work for hire” theory?
  3. What does the D.C. Circuit leave open?

4. Models, data and licences as property

In short. The software behind large language models is almost certainly eligible for copyright. Whether the model itself is protected is a really interesting question, since mathematical and statistical creations are not eligible. Training data may get only a thin claim, and cleaning it up is probably not enough. Many people put licences on data whether or not it is copyrightable, which leaves lawyers to decide how much risk to manage.

🎙 In the talk

  • The software that embodies LLMs is “almost certainly eligible” for copyright; whether the model gets protection is “a really interesting question”.
  • Mathematical and statistical creations “are not eligible”. The model may be like the outputs: “the product of automated processes”.
  • Training data: selection and arrangement might give a thin claim; “just cleaning up the data is probably not enough”.
  • Licences on data: “a lot of people are just slapping licenses on collections of data, whether they’re copyrightable or not”.
  • Q&A: licences in Getty and Copilot cases; combining data under conflicting licences is “decide how much risk to manage”; no copyright in the collection means “arguably no basis for imposing a license”.
  • Open-source developer: “I no longer can get the job”; her answer is the Doe class action.

📚 Deeper theory (see Roots)

  • This is the Feist gap. The EU fills it with a database right; the US leaves contract and trade secret.
  • Licence enforceability without copyright is a contract question, not a copyright one.
  • Open-source licence conditions (attribution, share-alike) are exactly the “conditions” Copilot is accused of ignoring.

🆕 Since the talk

  • The 9th Cir. Doe ruling (Sept 2026) closes this DMCA route for outputs. The court treated the training-input theory as forfeited, so it did not decide it (Haynes Boone).
  • The Anthropic settlement shows the market answer: a payment per work, in exchange for a release of liability for conduct up to 25 Aug 2025, which does not cover future claims (Kilpatrick Townsend).
  • Whether weights are protectable or “copies”: no ruling found. Unverified.

📚 Roots: Licences, contracts, CMI and data

The point..

The problem. Where copyright does not reach (data, facts, uncopyrightable compilations), people use licences, contracts and technical measures. Do these hold? Chain.

  • Open-source licences (GPL, MIT, Apache): grant copying rights on conditions, mostly attribution or share-alike. The Copilot plaintiffs’ complaint is that these conditions were ignored.
  • Creative Commons (2002): a licence layered on copyright; she says “the license sticks to the work because it’s a license to use the copyright”.
  • DMCA §1202 (1998): protects CMI. A statutory hook that lets plaintiffs sue without proving copying.
  • EU database right (1996): a sui generis right in databases, which the US rejected after Feist. Her remark that “data as such” is not protected marks the gap.
  • Contract and terms of service: Getty’s licence-breach claim. Robots.txt and opt-out signals: the audience asks about it.

Condition or covenant. Her Creative Commons remark rests on a real distinction. If a licence term is a condition of the permission, breaking it means you never had permission, so you infringe copyright. If it’s a covenant, you only breached a contract. Jacobsen v. Katzer (Fed. Cir. 2008) held that open-source attribution terms can be conditions, and that is what made such licences enforceable. This is why the Doe plaintiffs’ licence claims matter: they need a copy of protected code to be in play. If an output isn’t a copy, the condition isn’t triggered. So her point that a licence “sticks” only because of copyright shows the limit: where there’s no copyright, the licence has nothing to condition.

Contract as a substitute for copyright. ProCD v. Zeidenberg (7th Cir. 1996) enforced a shrink-wrap licence over an uncopyrightable phone database, on the logic that a contract binds only its parties. Hence the terms-of-service approach to scraping. Its limit is privity: a scraper who never agreed, or a downstream model user, is not bound. Her “slapping licenses” remark describes that gap. Licences are widely used and weakly grounded.

Where others push back. Open-source maintainers argue licences should bind AI trainers. Data vendors argue contract terms are enough. Neither view has a clear holding.

What the talk leaves out. Trade secret in weights; export controls; privacy law on personal data in training sets.

Check yourself

  1. Why would a model’s weights be hard to protect by copyright, in her telling?
  2. Why does a licence on uncopyrightable data have a weak basis?
  3. What did the 9th Circuit leave open in Doe?

In short. Copyright is a strict liability regime, so you do not have to intend infringement, and a user who generates an infringing output is as liable as the developer. She says the Copyright Office lacks technologists and is more or less setting industrial policy. Regulation is thin, and the EU AI Act does not know what to do with general-purpose AI. Copyright is the only law that could bring these systems to their knees, though she does not expect it.

🎙 In the talk

  • Strict liability: “copyright is a strict liability regime. You don’t have to intend”. If you generate an infringing output, “you are as liable… as the person who developed the software”.
  • The Office lacks “inside technologists”; she has argued for a chief technology officer and a “chief economist”. It is “more or less setting industrial policy”.
  • Regulation: EU AI Act was written for specific AI systems and “don’t know what to do” with general-purpose AI; US has the NIST framework and the AI Bill of Rights, “just a set of principles”.
  • Copyright as the “only law… that could bring generative AI systems to their knees”. “It’s an existential threat”, though she doesn’t expect it.
  • Ethics: no consensus; “there’s no law to enforce a particular ethical norm… unless there’s copyright infringement”.
  • Host’s close: the industry may be “at risk… if this goes the other way”.

📚 Deeper theory (see Roots)

  • Strict liability is real but practical risk depends on volume, damages (statutory damages for registered works), and plaintiff willingness. Enterprise vendors have responded with indemnities.†
  • The Office is part of the Library of Congress (legislative branch) but its Register is appointed by the Librarian, a structure that became the centre of the 2025 dispute.

🆕 Since the talk

  • Copyright Office Part 3 (training), pre-publication, 9 May 2025. It applied the four factors and concluded that some uses of generative AI to create competing commercial works go beyond fair use. Register Shira Perlmutter was removed one day later. A district judge rejected her challenge in July 2025 per one blog. A June 2026 Supreme Court matter, Blanche v. Perlmutter (No. 25A478, 30 June 2026), is reported by Lowenstein as denying the government’s stay application, which leaves her in office for now and may let Part 3 be formally published. It is not a merits ruling, and I could not confirm the order’s text (Patently-O, Lowenstein, Publishers Weekly). A May 2026 Senate listing also named her as Register.
  • The Office did not acquire the CTO and economist she asked for, as far as I could find. Unverified.
  • EU AI Act: in force and applying to general-purpose models in phases.†
  • US federal AI copyright statute: none found.

📚 Roots: Technology, copyright and political economy

The point. The “existential threat” framing and the ethics exchange.

The problem. Each new copying technology produced a fight: the photocopier, the VCR, Napster, Google Books. Who gets to profit when a general-purpose technology is built on others’ work? Chain.

  • Boyle and Lessig on the “second enclosure” and remix culture; Lessig, Free Culture (2004).†
  • Jaron Lanier, “data dignity”: people whose data trains models should be paid.† This is close to the “Marxist nightmare” line and the 90% Authors Guild survey.
  • Samuelson’s recurring theme that the Copyright Office “is more or less setting industrial policy” without economists.
  • Regulation: EU AI Act (proposal 2021; finalised 2024)†, NIST AI Risk Management Framework (Jan 2023), the White House Blueprint for an AI Bill of Rights (Oct 2022).

The recurring settlement. History suggests that copyright rarely stops a general-purpose technology, and also rarely leaves it free. Player pianos led to the compulsory mechanical licence (1909). Cable led to a statutory licence (1976). The VCR won outright in Sony, and then studios made more from home video than from cinema. Napster lost, and streaming replaced it on licensed terms. The usual end is neither ban nor free use, but a payment system that favours the large and organised. The Anthropic settlement fits that pattern. It’s why her “existential threat” framing may be right about the legal theory and wrong about the outcome.

Who is the beneficiary? Rights are held largely by publishers, labels and studios, not by individual creators. A win for “authors” may mostly pay intermediaries. A licensing regime would also favour the largest AI firms, who can afford it, and so could entrench them. Plaintiffs and open-source developers therefore make uneasy allies. Her open-source Q&A touches this without resolving it.

Institutions. She worries about capacity at the Copyright Office. The structure is odd: an office inside the Library of Congress whose opinions do not bind courts and carry only persuasive weight. The 2025 removal fight turned a technical report into a test of who controls that advice. Her call for an economist is a call to treat the matter as industrial policy, openly.

Where others push back. Some say the “existential threat” frame is exaggerated because settlements and licences keep the industry running (the Anthropic deal being the example). Others say the Office’s report was legitimate expert advice, and its leader’s removal politicised it.

What the talk leaves out. Congress’s role; state laws (right of publicity, digital replicas)†; international fights about jurisdiction.

Check yourself

  1. Why does strict liability matter for a company that lets staff use ChatGPT for ad copy?
  2. What two staff roles does she want at the Copyright Office, and why?
  3. What did the Office’s Part 3 say about commercial, competing models?