Talks

A Global Perspective on Copyright Law and Copy-Reliant AI — Matthew Sag

50 min read

About this talk

Speaker: Matthew Sag
Event: Keynote address for the Australian Digital Alliance (specific event name is not stated in the transcript or video details)
Date: 2024-05-02 (upload date; the talk is from early 2024)
Length: 43:15
Channel: Australian Digital Alliance

How to read this page. It has two kinds of content, always marked:

  • 🎙 In the talk. What Sag says, with [MM:SS] timestamps. Text inside quotation marks is word for word from the transcript. Only these parts are his views.
  • 📚 Deeper theory and later events. Background, case law, and what happened after the talk. It is not his claim unless a timestamp says so. Anything I could not check is marked “from memory; not checked”.

Suggested route. Summary, then Map, then Key terms, then Part 2: The intellectual map (read once, slowly), then the Deep dives in order, then “Did the predictions hold up?”, then “Putting it together”.

Dates. The talk is from early 2024. Today is 30 Oct 2026. About two and a half years of litigation and law-making have happened since, so the “Did the predictions hold up?” section matters as much as the talk.

Speakers

LabelWhoRole
SagMatthew Sag (named in the video title and description)Professor of Law in Artificial Intelligence, Machine Learning and Data Science, Emory University Law School (from the video description). Sole speaker
Audience member”Derek” (first name only, as heard in the captions)Asked, in a live demonstration, to guess the next word of a sentence [12:36]

Event. The video details name only the channel, Australian Digital Alliance. The transcript does not name the event. The talk is clearly addressed to an Australian audience (he says “the Australians will help me out here” [37:44] and refers to “the introduction”). I have not guessed a conference name. Caption warnings. Auto-captions. Corrections used on this page: “Dar 3” = DALL-E 3; “STW diffusion” = Stable Diffusion; “mid Journey” = Midjourney; “Leon 5B” = LAION-5B; “burn convention” = Berne Convention; “article 92 of the trips agreement” = TRIPS Article 9(2); “hay trust” = HathiTrust; “Tex / TCH data mining” = text and data mining (TDM); “Ian MCU” / “Y mccuan” = Ian McEwan; “Peron” = Perowne (character in McEwan’s Saturday; from memory; not checked); “conset” = consent; “Seed the field” = cede the field; “albanesi” = Albanese. The title on YouTube has a typo (“Perpsective”). Slips and loose claims by the speaker.

  1. He says copyright “was invented in England in statute of an in 1709” [03:10]. The Statute of Anne is dated 1709 in the old-style calendar and took effect in April 1710; sources give both years (from memory; not checked).
  2. He dates the HathiTrust case to 2013 [32:59]. The Second Circuit decision I remember is 2014, with a district court ruling in 2012 (from memory; not checked). Google Books at 2015 matches my memory.
  3. He says the Productivity Commission proposed fair use “in 2017” [37:44]. I remember the final report being dated late 2016 and released in early 2017 (from memory; not checked).
  4. His figure that the permissive jurisdictions hold “about 12%” of world population and “almost exactly half” of world GDP [34:35] is his own estimate, with no source given. Not checked.
  5. He says typical performing-rights organisations “only send money when an artist’s work is actually used” [39:16]. This is a simplification: many PROs also use survey and sampling methods (from memory; not checked). It does not damage his point.

Summary

Central question. Generative AI models are trained on vast copyrighted material. Is that training lawful, and what should a country with strict copyright law, such as Australia, do about it?

Main thesis. Sag builds the case in four moves.

  1. Copyright protects expression, not ideas. This is the one principle that is “universal” across countries [06:20]. So copying that exists only to extract facts, patterns and abstractions (he calls these non-expressive uses) sits uneasily with the purpose of copyright.
  2. A generative model is not a copy of its training data, and its outputs are not copies either (if it is “set up correctly”). Training is closer to “learning like a student” than “copying like a scribe in a monastery” [15:45]. Copying does happen, in huge quantity, but before training, when the data set is assembled [09:28].
  3. Two consequences. The fair use argument gets stronger. And a model lawfully trained in one country can be used anywhere, so strict countries only handicap their own universities and start-ups: regulatory arbitrage [22:05]–[23:38].
  4. Four complications keep this from being “plane sailing”: memorization, the “Snoopy problem” (characters), agentic retrieval (the New York Times examples), and users who deliberately push a model toward infringement [23:38]–[31:25].

Close. He surveys the US (fair use by case law), Japan, the UK, Switzerland, the EU and Singapore (TDM exceptions), and Israel and South Korea (fair-use-style rules). He then lists Australian options: US-style fair use, a new fair dealing purpose, a TDM exception, collective licensing, an AI-safety-conditioned safe harbour, or doing nothing. He is sceptical of collective licensing and unenthusiastic about doing nothing: “I’m not sure that that’s a great option for the clever country” [42:26].

Stance to note. He is a doctrinal optimist about training, and a candid one: he lists where the argument breaks. He says the analogies between human and machine learning are “often overstated” [15:45] and still prefers the learning frame to the copying frame. He also says copyright is not “the most important issue” in generative AI [01:36]. He argues this from the structure of copyright law, not from a claim that AI companies deserve a free pass: the right to train “should not be a blank check” [36:10].

Map — talking points

#PointTime
1Opening: generative AI is astonishing, global, and raises many fears. Copyright is “not the most important issue” but is guaranteed “a seat at the policy table”[00:03]–[03:10]
2Copyright’s origin (printing press) and its dominant metaphor. The 1990s brought copying with “no reader”: reverse engineering, plagiarism detectors, search indexes, library-scale analysis[03:10]–[04:46]
3Non-expressive use defined. The core policy question: balance owners’ interests against the public’s right to learn from ideas[04:46]
4The idea/expression distinction is “universal”: TRIPS Art. 9(2), Berne. It explains why bad poetry is protected and good ideas are not[06:20]
5What copyright is really for: not banning every technical copy, but protecting expression and its communication[07:53]
6Fact A: gathering and preparing training data means copying at the scale of hundreds of millions to billions. Hidden, machine-read copies[09:28]–[11:03]
7Fact B: the model is not a copy. Next-token prediction, weights start random. The “Derek” demonstration: why “hair”[11:03]–[14:11]
8Learning vs copying; analogies are overstated but “closer to” a student. Coffee-cup abstraction (LAION-5B)[14:11]–[17:20]
9Fact C: outputs are not copies. Latent features recombine. The teddy bear in opulent clothes[17:20]–[20:29]
10Implications: fair use is stronger; models can cross borders; regulatory arbitrage; strict laws hurt only domestic universities[20:29]–[23:38]
11Complication 1: memorization (duplicates, unique captions, model-to-data ratio; GPT-5 worry)[23:38]–[25:10]
12Complication 2: the Snoopy problem (characters protected abstractly)[25:10]–[28:17]
13Complication 3: agentic retrieval. The NYT complaint’s “astonishing” examples may be web retrieval, not memorization[28:17]–[29:51]
14Complication 4: determined users. The McEwan sequel experiment. “The same is also true of typewriter”[29:51]–[31:25]
15US: non-expressive use is “quintessentially transformative”; Sag’s 2009 article; HathiTrust, Google Books[31:25]–[32:59]
16Other countries: Japan first, then UK, Switzerland, EU, Singapore (TDM); Israel, South Korea, Singapore (fair-use-style). About 12% of population, half of GDP[32:59]–[34:35]
17Variables in TDM laws and three shared principles, including “not a blank check”[34:35]–[36:10]
18Australia today: express permission from every rights holder, or public domain only. “There could never be an open AI or an anthropic in Australia”[36:10]–[37:44]
19Options: fair use (ALRC 2013, PC 2017, “never happen”); new fair dealing; TDM exception (Japan or EU model)[37:44]–[39:16]
20Collective licensing critique: “there are no Taylor Swifts”; it is really an AI training tax[39:16]–[40:50]
21AI-safety safe harbour for training; outputs still liable. Or do nothing and cede the field[40:50]–[42:26]

Part 1 — Before you read

Key terms

TermMeaning hereWatch out
Non-expressive useCopying a work for a purpose other than letting someone enjoy its expression: indexing, plagiarism detection, reverse engineering, statistical analysis. Sag’s term from his own scholarship [04:46]A scholarly label, not a statutory term. No statute uses it
Idea/expression distinction (dichotomy)Copyright protects the particular expression, never the underlying idea, fact, method or conceptIn the US it is also tied to merger and scènes à faire; elsewhere it is a TRIPS rule only in outline (see 2.1)
Reproduction rightThe exclusive right to copy. Sag says the temporary and hidden copies in data prep “meet the statutory definition of reproduction” [11:03]Whether a transient or technical copy counts differs by country. He treats the data-set copies as real
Fair useUS open-ended defence, judged by four factors (purpose, nature, amount, market effect)Not the same as fair dealing
Fair dealingCommonwealth model: a closed list of permitted purposes (research, criticism, news, parody, etc.). Australia, UK, CanadaSag’s central problem for Australia: AI training is not on the list (from memory; not checked beyond his statement at [37:44])
TDM exceptionA statutory exception allowing text and data mining. Versions differ in scope, whether commercial use is covered, “lawful access” requirements, and whether owners can opt out”Opt-out” and “lawful access” are the two main dials [36:10]
Token / parameter / weightsToken: a word or word part. Parameters/weights: the numbers inside the model, “billions possibly a trillion” [11:03]The legal question is whether weights are a copy or contain one
Latent space / latent representationThe compressed, abstract features a model learns (what “a coffee cup” is in general) [17:20]Sag uses it as a legal argument: abstract features are uncopyrightable
Memorization / overfittingA model storing specific training items, so it can reproduce themHe calls it “an edge case” [25:10]. Whether it stays one is the biggest open question
InferenceUsing the trained model, e.g. typing a prompt into Midjourney [17:20]
Regulatory arbitrageDoing the training where the law is most permissive, then deploying everywhere [23:38]
Substantial similarityThe test for whether an output copies protected expressionA work can be protected so broadly (a character) that it fails without close resemblance to any one image [26:44]
Agentic / synthetic searchA chatbot that fetches live web pages and summarises them [28:17]Changes which act is being alleged: memorization versus retrieval

People and cases named (a reading list)

Named in the talk: the Statute of Anne (1709); TRIPS and the Berne Convention; HathiTrust (he says 2013) and Google Books (2015) [32:59]; his own 2009 law review article [32:59]; OpenAI, Anthropic, GPT-4, GPT-3.5, Midjourney, DALL-E 3, Stable Diffusion, LAION-5B; the New York Times complaint against OpenAI [28:17]; Ian McEwan’s Saturday [31:25]; Australian Law Reform Commission (2013) and Productivity Commission (2017) [37:44]; ASCAP and performing rights organisations [39:16].

Not named, worth knowing: Baker v. Selden (1879); Nichols v. Universal (L. Hand, 1930); Sony v. Universal (Betamax, 1984); Campbell v. Acuff-Rose (1994) and Pierre Leval’s “Toward a Fair Use Standard” (1990); Sega v. Accolade (1992); Perfect 10 v. Amazon (2007); A.V. v. iParadigms (2009); Carlini et al. on training-data extraction (2021–23); Sag’s “Copyright Safety for Generative AI” (Houston Law Review, 2023); Mark Lemley and Bryan Casey, “Fair Learning” (2021); Pamela Samuelson on generative AI and fair use. (All from memory; not checked, except Sag’s 2023 article, which a search confirmed: SSRN.)

Sag’s own frame

Sag works inside doctrine. He starts from what copyright law is for (the purposive reading in [07:53]) and asks whether training threatens that purpose. He is not arguing about whether creators deserve compensation in the abstract; he is arguing about which legal tool fits. That is why he names tools he thinks do not fit (collective licensing) and tools he thinks fit better (a TDM exception or a new fair dealing purpose). If you read him as “pro-AI-company” you will miss the careful carve-outs: memorization, characters, and user misuse. If you read him as “neutral”, you will miss that his default, in every close call, is to protect learning from ideas.

This is added context, not Sag’s view. Sources are linked. Where a search result was thin, I say so.

1. US courts: training is often fair use, but with limits Sag partly anticipated.

  • In June 2025, two judges in the Northern District of California granted summary judgment on fair use in AI training cases. Bartz v. Anthropic (Judge Alsup, 23 June 2025) called training “exceedingly transformative” but said keeping pirated copies in a central library was not fair use. Kadrey v. Meta (Judge Chhabria, 25 June 2025) went Meta’s way on a narrower basis (plaintiffs did not prove market harm) and warned that the ruling was limited. Neither binds other courts. (Mondaq summary; Manatt)
  • Bartz then settled for $1.5 billion (about $3,000 per book across roughly 400,000 works), reported as given final approval in July 2026. Some authors opted out and sued separately. (WSLS/AP, 21 Jul 2026) Note that the sources I found disagree on which judge approved it; I have not resolved that.
  • Thomson Reuters v. Ross (non-generative AI, legal research) went the other way. The Third Circuit affirmed that Ross’s use of Westlaw headnotes was not fair use. It ruled on 29 Sept 2026; the opinion was first sealed, so the reasoning is only partly known. (IPWatchdog; Plagiarism Today) I did not confirm the full holding.
  • New York Times v. OpenAI/Microsoft: in 2025 the court let the core claims proceed; it did not decide fair use. I found no merits ruling in 2026, only an amended complaint in June 2026. (piracymonitor; Hoodline)
  • US Copyright Office, Part 3 report (pre-publication, 9 May 2025): copying to build training sets is presumptively infringing unless a defence applies; it is sceptical of fair use where outputs compete with or dilute the market for the originals. The Register of Copyrights was dismissed the next day. I did not verify whether a final version has issued. (NatLawReview; Mass Law Blog)

2. Europe and the UK: the model-as-copy question has split.

  • GEMA v. OpenAI (Munich Regional Court, 11 Nov 2025, 42 O 14139/24): lyrics memorised in GPT-4 and GPT-4o were held to be reproduced in the model, and the outputs infringed. The court said the TDM exception covers training only so long as memorization does not occur. OpenAI said it would appeal. (Bird & Bird; Taylor Wessing)
  • Getty Images v. Stability AI (England and Wales High Court, [2025] EWHC 2863 (Ch), 4 Nov 2025): Getty dropped its main copyright claims at trial; on the secondary-infringement claim the court held the Stable Diffusion weights were not an “infringing copy” because they never stored or contained copies. Training took place outside the UK, so legality of the scraping itself was not decided. The sources I found conflict on who has leave to appeal. (AIPPI; Bird & Bird)
  • EU AI Act: obligations on general-purpose AI providers applied from 2 August 2025. The voluntary GPAI Code of Practice (10 July 2025) has a copyright chapter: respect robots.txt, make best efforts to honour other machine-readable opt-outs, publish a copyright policy, offer a complaints route. Rights holders say this is too weak. (Jones Day; Bech-Bruun)
  • UK: the government’s 18 March 2026 report under the Data (Use and Access) Act 2025 did not propose a TDM exception; the earlier “exception with opt-out” idea was reported as dropped; transparency was the area of broad support. (Mondaq; Wiggin)
  • Not covered by my searches: the Hamburg Kneschke v. LAION appeal and the CJEU reference Like Company v. Google. Marked “from memory; not checked”: both exist, and both bear on the EU TDM exception.

3. Japan and Singapore.

  • Japan: Article 30-4 is still widely described as one of the most permissive frameworks, with an “unreasonably prejudice the owner” limit. Yomiuri (Aug 2025), then Nikkei and Asahi, sued Perplexity in Tokyo District Court over its answer-engine use of articles; I found no ruling. (The Star; Law360)
  • Singapore: I found nothing current in my searches. Its 2021 computational-data-analysis exception was in place at the time of the talk (from memory; not checked).

4. Australia (the talk’s target audience).

  • On 26–27 Oct 2025 the Attorney-General, Michelle Rowland, ruled out a text and data mining exception. The government said it would instead consult the Copyright and AI Reference Group (CAIRG) on licensing models for AI training. This is a stricter position than the EU’s opt-out regime. (iTnews; McCullough Robertson; Bird & Bird). Primary text of the release not read; these are secondary reports.
  • I have not checked whether anything has happened since late 2025 (for example a licensing scheme, a transparency law, or a bill).

5. Sag since the talk. In Nov 2025 he gave a Yale lecture, “The Snoopy Solution”, on how fair use and licensing might coexist, framing safety measures against memorization as a basis for cooperation between developers and creative industries (Yale Law School event page). I have not read it. It suggests he has moved from “licensing can’t work” toward a combined model, but that is my inference from the title and blurb.

→ Scorecard against his predictions: see Did the predictions hold up? after the deep dives.

Part 2 — The intellectual map

The talk is only 43 minutes, but it compresses at least eight separate bodies of thought. Each section below says what problem it addresses, who built it, and how it feeds Sag. None of this is in the talk unless a timestamp says so.

2.1 The idea/expression distinction — copyright’s escape valve

Problem. If copyright protected ideas, one author could monopolise a genre, a method or a fact. If it protected nothing, there would be no incentive to write. The law needs a line.

Chain.

  • Baker v. Selden (US, 1879): a book describing a bookkeeping system is protected; the system is not. Origin of the US rule.
  • Nichols v. Universal Pictures (L. Hand, 1930): the “abstractions test”. Every work has levels of abstraction; at some level the pattern is free for all. Characters are protected only if “sufficiently delineated” (from memory; not checked).
  • Berne Convention (1886) leaves the line to national courts. TRIPS Art. 9(2) (1994) says copyright covers “expressions and not ideas, procedures, methods of operation or mathematical concepts as such. WIPO Copyright Treaty (1996) repeats it. So the line is genuinely international. This is the basis for Sag’s Universal” at [06:20].
  • Feist v. Rural (US, 1991): facts are not protected; the compiler’s creative selection can be.
  • Merger doctrine and scènes à faire (US): where an idea can only be expressed in one or few ways, or a stock scene follows from a genre, there is no protection.

How it feeds Sag. It is his premise. “Bad poetry” is protected and “good ideas” are not [06:20]; so the key question for a model is which of the two did it extract?

Limits. The line is blurry. “Expression” has been stretched to include structure, selection, and, for characters, a recognisable essence. Sag himself shows this with Snoopy [26:44].

2.2 Fair use, transformativeness and the logic of non-expressive use

Problem. Some copying is socially valuable but not covered by a listed purpose. A flexible defence is needed.

Chain.

  • Folsom v. Marsh (Story J., 1841): origin of the fair use factors.
  • Betamax (Sony v. Universal, 1984): technology with substantial non-infringing uses is not contributory infringement. Context for Sag’s “typewriter” line [31:25].
  • Leval (1990) and Campbell (1994): “transformative” use, one that adds new purpose or meaning, favours fair use.
  • Reverse engineering: Sega v. Accolade (1992), copying object code to learn the interface is fair use. Sag cites reverse engineering as a precedent [22:05].
  • The search-and-index line: Kelly v. Arriba, Perfect 10 v. Amazon (2007), A.V. v. iParadigms (2009, plagiarism detection) → HathiTrust (2014) and Authors Guild v. Google (2015, Google Books). These are the “plagiarism detection, book search and literary analysis” cases Sag lists at [22:05]. (All dates from memory; not checked.)
  • Sag’s own contribution: a 2009 paper (he cites it at [32:59]) argued that copying which does not communicate expression to a human audience should be treated as presumptively fair. He later labelled these “non-expressive uses” [04:46].
  • Warhol v. Goldsmith (2023): tightened “transformative” where the new use serves the same purpose as the original and is commercial. Reported to have been central to the Third Circuit in Ross (see Part 1).

How it feeds Sag. The fair use argument for training is, in his words, that the model “poses no threat of direct expressive substitution” [22:05]. The 2023 Warhol shift is where later courts push back: if outputs substitute for the originals’ market, the “non-expressive” label no longer settles things.

2.3 Machine learning in plain terms — generalisation, memorization, latent space

Problem. Law depends on a factual claim: does a model contain its training data? The answer lives in computer science.

Chain.

  • Statistical learning: a model is fit to data to generalise to new inputs. Overfitting is when it fits the training set too closely and fails to generalise (classic ML theory; from memory; not checked).
  • Neural language models predict the next token; transformers (Vaswani et al., 2017) made large-scale training practical.
  • Scaling laws (Kaplan et al., 2020): more data and parameters, better models. Fits Sag’s “the more data the better” [09:28].
  • Memorization research: Carlini and others showed that large models can regurgitate rare training strings, with extraction easier as models grow and as items are duplicated. Deduplication reduces it. (From memory; not checked.) This is the scientific basis for his three risk factors at [25:10]: duplicates, unique captions, large model-to-data ratio.
  • Diffusion and latent representations: image models learn a compressed space of visual features and generate by denoising. Sag’s coffee-cup example [17:20] is a plain-language version of this.
  • Sag’s “Copyright Safety for Generative AI” (2023) turns this into a legal argument: memorization is an edge case, and developers can reduce it by deduplication and filters. A search confirmed the paper and its main claim (SSRN).

How it feeds Sag. His whole case that the model is “not a copy” [11:03] rests on generalisation beating memorization. The GEMA ruling (Part 1) is a court deciding that for specific lyrics it did not.

Limits. “Not a copy” is a claim about the typical case. Whether weights encode a work that can be extracted is a graded, empirical matter, not a yes/no.

2.4 Substantial similarity, characters and the long reach of derivative rights

Problem. How different must a new work be before it is not a copy?

Chain.

  • Substantial similarity tests (US): “ordinary observer”, “extrinsic/intrinsic”. Courts compare protected elements only.
  • Character protection: Nichols (1930) sets the “sufficiently delineated” idea; Walt Disney v. Air Pirates (9th Cir. 1978) and DC Comics v. Towle (9th Cir. 2015) protect distinctive visual characters like Mickey Mouse and the Batmobile. (From memory; not checked.)
  • Derivative works right: authors control sequels, adaptations, translations. Novels carry very broad derivative rights.

How it feeds Sag. This is the logic of the Snoopy problem [26:44] and the McEwan sequel [29:51]–[31:25]. A character is protected “so broadly” that an image can infringe without matching any particular earlier image. It is the main place where his “outputs are not copies” argument is weakest, and he says so.

2.5 Secondary liability and the “tool” question

Problem. When a user misuses a general-purpose tool, who is responsible?

Chain.

  • Betamax (1984): a technology capable of substantial non-infringing use does not make its maker liable.
  • Grokster (2005): inducement, if a maker promotes infringing use, is liable.
  • Safe harbours and notice-and-takedown (DMCA, 1998).
  • Recent US cases about whether service providers who knew of infringement are contributory infringers (a Supreme Court case on ISP liability, Cox v. Sony Music, was pending as of my last knowledge; from memory; not checked).

How it feeds Sag. His complication 4 [29:51]–[31:25]: “not every Act of infringement committed using generative AI should fairly be attributed to the entities that host or develop these models” but “some probably should”. That is a Betamax/Grokster balance in one sentence. (The second half, “some probably should”, is Sag’s too; I add that the Grokster inducement line is the tool I would reach for to draw it, which is mine, not his.)

2.6 Territoriality and regulatory arbitrage

Problem. Copyright is national, but data and models move freely.

Chain.

  • Berne’s national treatment and the territoriality principle: infringement is judged by the law of the country where the act occurs (from memory; not checked).
  • Practical consequence: training in Tokyo or Tel Aviv is judged under that country’s law, not Sydney’s.
  • The question is then about import and use of the resulting model. Sag’s argument is that if the model is not an infringing copy, there is nothing to stop at the border [22:05]. Australian law has provisions on importing articles whose making “would have” infringed if done in Australia (from memory; not checked), which a future court might test against model weights. This is my observation, not his. Getty v. Stability addressed the same “is it an infringing copy?” question under the UK equivalent (Part 1).
  • Comparable arbitrage patterns: tax residence, data-protection transfers, offshore pharmaceutical trials.

How it feeds Sag. [22:05]–[23:38]. His policy conclusion (a strict law only harms domestic universities and start-ups) depends on the “not a copy” premise. If GEMA-style reasoning wins (weights can contain works), that conclusion weakens in the EU, and potentially elsewhere.

2.7 The comparative law of TDM exceptions

Problem. How should statute handle large-scale analysis of works?

Chain (all from memory; not checked unless noted).

  • Japan: 2009 research exception, then Article 30-4 (2018 amendment, in force 2019), allowing use for “non-enjoyment” purposes, limited where it would unreasonably prejudice the owner. Sag confirms “limited by definition to non-expressive uses which is the case in Japan” [36:10].
  • UK: 2014 non-commercial research exception; a 2022 plan for a broad commercial exception was withdrawn; the 2026 report did not revive it (Part 1).
  • EU: Digital Single Market Directive (2019/790): Art. 3 research; Art. 4 commercial TDM with a machine-readable opt-out (“reservation of rights”). Sag mentions “directives three and four” [39:16].
  • Switzerland and Singapore: later TDM exceptions (Sag lists them [32:59]).
  • Israel (a 2022 Ministry of Justice opinion that training is fair use) and South Korea: fair-use-style rules adopted partly to compete with the US, which Sag notes [34:35].
  • Australia: fair dealing closed list. ALRC (2013) recommended fair use; Productivity Commission recommended it (he says 2017); neither was adopted [37:44].
  • The Berne/TRIPS “three-step test” (Art. 9(2) Berne, Art. 13 TRIPS) limits exceptions to “certain special cases” that do not conflict with normal exploitation or unreasonably prejudice the owner. Opponents argue broad TDM breaches this; Sag does not raise it. (From memory; not checked.)

How it feeds Sag. The three shared principles at [36:10] are his synthesis across these regimes. The dials he names (scope, lawful access, “contractual and technological overrides”) are exactly where the 2025 EU Code of Practice and the UK debate now live.

2.8 Licensing, collective management and the economics of attribution

Problem. If training uses works, can owners be paid fairly?

Chain.

  • Collective management organisations (ASCAP/BMI in the US, PRS in the UK, APRA AMCOS in Australia, GEMA in Germany): blanket licence, usage-based distribution.
  • Extended collective licensing (Nordic countries): a collecting society can license on behalf of non-members, with opt-out (from memory; not checked).
  • Statutory/compulsory licences: a fixed rate, no consent (the US mechanical licence for music is the classic example; from memory; not checked).
  • Attribution problem in machine learning: no practical way to say how much any single training item contributed. Research on “data attribution” or “influence functions” exists but is not settled (from memory; not checked).
  • Evidence since the talk: the Bartz settlement paid a flat roughly $3,000 per work (Part 1). That is a negotiated, one-off figure and a piracy remedy, not a licence, but it is the equal-division logic in practice.
  • Alternative: bilateral licensing deals (news publishers with AI companies) and “pay for access/retrieval” markets. These are market deals, not a statutory scheme (from memory; not checked).

How it feeds Sag. [39:16]–[40:50]. His objection is mechanical, not moral: usage-based distribution is what makes collective licensing efficient, and training has no usage signal. See Deep dive 7.

How the traditions meet

                 IDEA / EXPRESSION (2.1)  ───────────────┐
                 "facts, ideas, methods are free"          │
                          │                                 ▼
 NON-EXPRESSIVE USE (2.2) ┼──►  Is the model a COPY?  ◄── MACHINE LEARNING (2.3)
 Sega, iParadigms,        │        │      │                generalise vs memorise
 HathiTrust, Google Books │        │      │
                          │   no   │      │ yes (memorised lyrics, Harry Potter)
                          ▼        ▼      ▼
                  Training = fair   OUTPUTS (2.4)         GEMA v OpenAI
                  use / TDM         characters, sequels   Getty v Stability (UK: not a copy)
                          │             │
                          ▼             ▼
            TERRITORIALITY (2.6)   SECONDARY LIABILITY (2.5)
            train in Tokyo, use     typewriter / Betamax / Grokster
            in Sydney
                          │
                          ▼
           POLICY MENU for Australia (2.7, 2.8)
   fair use | new fair dealing | TDM exception | licensing | safe harbour | nothing

Part 3 — Deep dives

🎙 In the talk.

  • Sag starts from the printing-press story: “the printing press remains our dominant metaphor for how copyright is supposed to work” [03:10]. In that story the reproduction of a work is where author and reader meet.
  • Then the twist: “what if there is no reader” [03:10]. From the 1990s on there were economically important acts of copying (reverse engineering software, plagiarism detectors, search-engine indexing, statistical analysis of whole libraries) that had nothing to do with communicating expression. “I call these non-expressive uses” [04:46].
  • The tension: copyright’s focal right is reproduction, but its purpose is to let the author control communication of expression “while still allowing facts ideas abstractions artistic methods and techniques to be freely copied and shared” [04:46].
  • He anchors this in the idea/expression distinction, which he calls “Universal” [06:20], with the line that “bad poetry really bad poetry is copyrightable but good ideas really good ideas are uncopyrightable” [06:20].
  • His reading of the purpose: it is “not to Simply prohibit every technical Act of reproduction for its own sake regardless of purpose or effect” [07:53]. Learning from existing expression and converting it to useful metadata is “just as important” as the right to object to expressive substitution [07:53].

📚 Deeper theory (see 2.1, 2.2).

  • This is a purposive argument. Its opponent is a literal reading: the statute says reproduction is the owner’s exclusive right, with listed exceptions, so every copy needs a licence or an exception. Both are legitimate styles of legal reasoning. Sag chooses purpose, and in the US that choice is built into the fair use factors; in a closed fair dealing system (Australia) it is harder to act on.
  • “Universal” is accurate for the principle (TRIPS Art. 9(2)), but countries disagree on where the line falls for characters, styles, structure and compilations. So universality of the principle does not give a universal answer to training.
  • A second, older argument sits behind this: copyright is a utilitarian bargain (incentive for authors, benefit for the public). Authors’ rights traditions (France, Germany) emphasise the author’s personal bond to the work. Those systems also have the idea/expression line, but are less comfortable with an unwritten fair use defence. That helps explain why the EU chose a statutory TDM exception with an opt-out and not US-style fair use (from memory; not checked).
  • Critics treat “non-expressive” as a category error for generative AI: a model trained to produce expression in the style of its training data is, they argue, ultimately an expressive tool, even if the training step itself shows no expression to a human. Sag’s reply is the output-side argument in Deep dives 2 and 3.

Where others push back.

What he leaves out.

  • Market harm. In a 43-minute talk he does not discuss whether outputs substitute for human-made work in the market. Courts and the Copyright Office now treat that as central (Part 1). He mentions “expressive substitution” [07:53] but only as the line training does not cross.
  • Moral rights and consent as values, not just rules.
  • Unlawful sourcing: he assumes data is scraped from “the open internet” [03:10], and does not discuss pirated libraries. The Bartz ruling made that distinction decisive (Part 1).

Check yourself.

  1. What is a “non-expressive use”, and why does Sag say it “poses a real question” for a law built around the right to reproduce?
  2. Which two “universal” instruments does he cite for the idea/expression distinction, and what does the distinction explain?
  3. Why would a literal reading of the Copyright Act lead to a different answer from his purposive reading?

2. Learning versus copying — the model is not a copy

🎙 In the talk.

  • Three facts frame the analysis [07:53]–[20:29]. The first: gathering data “inevitably involves copying” [09:28], “not just in the hundreds not just in the thousands but almost inevitably in the hundreds of millions and the billions” [09:28]. These are “hidden copies invisible copies that are only ever read by machines” [11:03], yet they “meet the statutory definition of reproduction” [11:03].
  • The second: the model is something else. It is “an absurdly large statistical model that attempts to predict the next word” [11:03]; at the start its weights are random.
  • The “Derek” demonstration [12:36]: asked to complete “the girl with the dark”, the audience member says “hair”. Sag asks “which book did he plagiarize it from” and calls it “obviously an absurd question” [12:36]. Derek guessed from what he knows about language and the world. A guess of “galaxy” or “semicolon” would be poor [14:11].
  • Training “effectively updates the weights”, and those weights “don’t reflect any single text” [14:11]. So: “it’s not copying the text it is in a very real sense learning from the text” [14:11].
  • The caveat, immediately: analogies to human cognition “can be and are often overstated” [15:45]. But between “copying like a scribe in a monastery or learning like a student it’s definitely closer to the second one” [15:45].
  • Abstraction example [15:45]–[17:20]: images labelled “coffee cup” in LAION-5B versus a newly rendered cup. It shares round, white, ceramic, one handle, coffee-coloured liquid; “all the material that’s been abstracted away” [17:20] (sunsets, Pikachu, cakes, ocean vistas).
  • Condition: “if these models are set up correctly they are forced to learn at a somewhat abstract and uncopyrightable level rather than simply memorizing the training data” [15:45].

📚 Deeper theory (see 2.3).

  • The Derek demonstration is a teaching analogy to generalisation. It works because “the girl with the dark ___” is a high-frequency pattern. It does not show what happens with rare strings, where a model has seen one example and can reproduce it. That is the memorization complication, which Sag raises next [23:38].
  • “Learning” is doing two jobs. In machine learning it names an optimisation process. In law and everyday talk it implies a mind that understands. The legal argument needs only a weaker claim: that nothing like the original’s expression is retained. The human-analogy is rhetorical help, and he flags its limits himself.
  • Whether weights are a “copy” in law turns on how a statute defines it. In the UK Getty case the High Court reasoned that an “infringing copy” must be, or have contained, a copy; weights never did, so they were not one. The Munich court in GEMA looked at memorised lyrics and found the model embodied a reproduction (Part 1). The same technology, the same question, two answers: because the facts (image model v. lyrics memorised by a text model) and the legal tests differ.
  • A relevant fact: training involves copying before the model exists, so even a “learning” model starts with the legal exposure Sag admits at [11:03]. A TDM exception or fair use is what covers that step, which is why the global survey in Deep dive 5 matters.
  • “Hidden copies” connects to the older debate on transient and incidental copies (the EU’s Article 5(1) exception for temporary copies; the US Cartoon Network case on buffer copies). From memory; not checked.

Where others push back.

  • Memorization researchers show extraction is possible in specific cases, and larger models memorise more (see 2.3).
  • Some critics argue that a model’s capacity to generate in the style of an author is evidence it has captured that author’s expression, not only ideas. Style is not protected by copyright, but some artists claim their distinctive works are being absorbed.
  • Others argue that “learning like a student” masks that the student here is a commercial product built by copying at industrial scale.

What he leaves out.

  • Compensation for the data pipeline’s copying even if the model is not a copy (e.g. a levy). He deals with licensing later [39:16]–[40:50].
  • Differences between model types: next-token text models, diffusion image models and video models memorise in different ways.

Check yourself.

  1. Where in the pipeline does Sag say copying definitely happens, and why does that not undermine his “model is not a copy” claim?
  2. What does the “Derek” exercise show, and what does it fail to show?
  3. Why did two courts (Munich and London) reach different conclusions about whether a model can “contain” a work?

3. Outputs are not copies — latent features and novel instances

🎙 In the talk.

  • “The digital artifacts that are produced by generative AI models are not copies of the training data either” [17:20], meaning that if made any other way “no one would think that you met copyright’s criteria of substantial similarity” [17:20].
  • Reason 1: models “only learn abstract and uncopyrightable associations” [17:20]. If information “hasn’t been stored it can’t be retrieved” [17:20]–[18:57].
  • Reason 2: at inference, outputs “combine multiple uncopyrightable latent features” [18:57], further “attenuating or Breaking the Chain between the training data and the model output” [18:57].
  • The teddy-bear example: a Midjourney image for “a teddy bear in Rich opulent clothing with ultra realistic textures and a hypnotic stare reading a newspaper” [18:57]. His conclusion: “this image does not exist in the training data” [18:57]; models “generate novel instances” [20:29].
  • A joke that carries a point: the bear wears sunglasses, “so we don’t actually know whether the bear is staring or asleep” [20:29]. The output is surprising, hence not a retrieval.

📚 Deeper theory (see 2.3, 2.4).

  • Copyright infringement needs (1) copying in fact (access plus connection) and (2) substantial similarity in protected expression. Sag’s argument attacks both: the information was never stored, and the recombined output is not similar to any single source. Both are testable per output.
  • The prompt matters. Prompts can name artists, characters or titles, steering the model to a specific source. A generic prompt is easier to defend than “in the style of [living artist]“. Style alone is not protected, but the evidence of intent differs. (From memory; not checked.)
  • Novelty is not the legal test for infringement: a new picture can still infringe if it copies protected parts. The teddy bear example is clean because the content is generic. The Snoopy example is the contrast.
  • Authorship of the output is a separate question (US Copyright Office has said purely AI-generated material lacks human authorship; the Thaler line). Sag’s talk is about input-side copyright, not output ownership. From memory; not checked.

Where others push back.

  • Output similarity can be shown for some prompts; artists have produced near-copies by prompting with their names or titles (2.4).
  • The “latent space” is not a neutral summary. Which features it keeps depends on what was in the data, so there is a causal chain from training data to output even if no single work is retrievable. Critics say that is why “attenuating or breaking the chain” overclaims. Sag says “attenuating”, not “severing”, in the first half of that phrase; the second half goes further.

What he leaves out.

  • Market-level effects of many novel outputs, where individually “new” images collectively compete with illustrators and stock photography.

Check yourself.

  1. What are the two reasons Sag gives for outputs not being copies?
  2. Why does “this image does not exist in the training data” not settle infringement by itself?
  3. What kind of prompt would make the teddy-bear argument weaker?

4. The four complications — memorization, Snoopy, retrieval, determined users

🎙 In the talk.

  • Sag introduces them as the limits of his simple story: “I don’t want to suggest that all of this is or should be plane sailing” [23:38].
  • (1) Memorization. “sometimes generative AI models will overfit” [23:38]. He says “memorization is very much an edge case” [25:10], but more likely with duplicates (“so easy to get GPT to reproduce Harry Potter before they put filters on” [25:10]), with unique captions on images, and when the model is large relative to the data. Forward-looking worry: “it is interesting to think about whether it will become more problematic as we go from GPT 4 to GPT 5” [25:10].
  • (2) Snoopy. Text-to-image models “are prone to produce potentially infringing images when text descript descriptions are paired with relatively simple reoccurring images” [25:10]–[26:44] (he is describing something like Snoopy and baby Yoda). The legal sting: “you can infringe the copyright in Snoopy with images that are not actually substantially similar to any particular pre-existing Snoopy image” [26:44]. Preventing that is “a much harder challenge” than stopping the reproduction of specific items [26:44]–[28:17].
  • (3) Agentic retrieval. The New York Times gave “a hundred examples of significant and I would say quite frankly astonishing memorization” [28:17]. But some “are not actually about memorization they’re about something else” [28:17]: a chatbot with web access can retrieve text. With the right prompting “it can end up functioning as direct text retrieval” [29:51]. This is still potentially a problem for OpenAI but “a much much smaller copyright problem” than memorizing vast swathes of data [29:51]. He notes the Times “doesn’t disclose its full prompting strategy” [29:51].
  • (4) Determined users. Generative AI “can become a tool of infringement in the hands of a determined user” [29:51]. His test: asking ChatGPT to summarise McEwan’s Saturday, invent a sequel “Sunday”, and expand each chapter [29:51]–[31:25]. “In the right hands” it can be a tool of infringement “but the same is also true of typewriter” [31:25]. Verdict: “not every Act of infringement committed using generative AI should fairly be attributed to the entities that host or develop these models” [31:25].

📚 Deeper theory (see 2.3, 2.4, 2.5).

  • Memorization: the literature distinguishes verbatim memorization (strings), near-duplicate memorization, and conceptual leakage. A legal test for “contains” must pick one. GEMA picked verbatim lyrics, which are short, repeated across the web, and unambiguous (Part 1). That fits Sag’s own risk factors (duplication) [25:10].
  • Snoopy: the problem is the law’s treatment of characters, not just the technology. Character copyright protects an abstraction (the beagle), so the “abstract learning is uncopyrightable” argument runs into a doctrine in which the abstraction is exactly what is protected. Sag’s later “Snoopy Solution” talk (Part 1) suggests he treats this as the place where licensing and safety filters can coexist with fair use. Inference from the title; not read.
  • Retrieval: from a legal standpoint, retrieval-augmented generation and browsing are outputs plus a live copy and look more like search-engine snippets or news aggregation. The Japanese publishers’ suits against Perplexity (Part 1) and others are of this type. It also changes the jurisdictional and licensing picture, since it can be handled by robots.txt and paid access. The NYT litigation continues; I found no ruling testing Sag’s retrieval reading (Part 1).
  • User misuse: the Betamax/Grokster line (2.5) is the frame. Typewriters and photocopiers are the stock examples of dual-use technology. The difference some argue: a typewriter does not carry a trained-in capacity to reproduce a given novel.

Where others push back.

  • On memorization, plaintiffs argue that “edge case” is a statement about frequency, whereas liability attaches per instance. GEMA won on a handful of songs.
  • On retrieval, others argue that if the model can be induced to reproduce articles, the reason (memorised or fetched) matters less than the result.
  • On users, platforms are not passive: they design the interface and filters.

What he leaves out.

  • Remedies. Under GEMA, the court noted that training data cannot easily be removed from a trained model (Part 1), which makes injunctions hard to comply with; this “unlearning” problem is not in the talk.
  • Trademark and publicity-rights issues, which he mentions only at [40:50] in the context of the safety safe harbour.

Check yourself.

  1. Name Sag’s three risk factors for memorization and give one example of each from the talk.
  2. Why is protecting a character broader than protecting a specific image?
  3. In the NYT example, why does he think the distinction between memorization and retrieval matters?

5. The global map — who permits training, and on what terms

🎙 In the talk.

  • US. Non-expressive uses “have been recognized as highly transformative and quintessentially transformative” and “very likely to be recognized as fair use all else being equal” [31:25]–[32:59]. He says he made this argument in a 2009 law review article, then case law backed it: HathiTrust (he says 2013) and Google Books (2015) [32:59]. Because the US position comes from case law, “there is some uncertainty” [32:59]. His summary: “at least some AI training is protected by fair use and probably quite a lot” [32:59].
  • TDM exceptions. “Japan was the first movie here followed shortly by the UK and then eventually by Switzerland the European Union and Singapore” [32:59]. A parallel group adopted fair-use-like rules: Israel, South Korea and Singapore were named as “particularly important” [34:35].
  • Weight of the group. He says the group has about 12% of world population and “almost exactly half” of world GDP [34:35]. China is left out because he and a co-author are still researching it [34:35].
  • Variables [34:35]–[36:10]: breadth of exemption, a lawful access requirement, and “how to address contractual and technological overrides imposed by rights holders”.
  • Three shared principles [36:10]: (1) learning abstract information is socially valuable and does not inherently prejudice copyright’s interests; (2) the copying inherent in building AI models should be allowed “without the express consent of copyright owners at least in some circumstances”; (3) the right “should not be a blank check”, being either limited to non-expressive uses (Japan) or subject to prejudice-based considerations.

📚 Deeper theory (see 2.2, 2.7).

  • Fair use-by-case-law and statutory TDM exceptions look similar but allocate risk differently. Case law offers flexibility but uncertainty, and the burden is on the defendant to prove fair use. A statute gives predictability, but its text can be narrower than the need (the EU opt-out, the UK’s non-commercial limit until now).
  • The “lawful access” requirement means the miner must have legitimately obtained the work. This connects directly to Bartz, where piracy defeated the defence for the central library (Part 1).
  • The opt-out in the EU is not a Sag invention; it is the legislative response to “contractual and technological overrides”. The AI Act’s Code of Practice now operationalises it through robots.txt and similar machine-readable signals (Part 1).
  • Measures of “countries that allow training” depend on counting. China is a large unknown; India’s position is not covered in the talk. For India, I have not checked the current position.

Where others push back.

  • The 12%/50% framing can be read as showing that wealthy countries have already decided, and so poorer, creator-heavy countries should follow. Critics say the real driver is geopolitical competitiveness, not legal logic. Sag acknowledges the motivation: some countries adopted fair-use-style rules “motivated by the technological advantage that fair use appeared to have given the United States” [34:35].
  • The three-step test (2.7) is an argument that broad exceptions breach international obligations. He does not address it.

What he leaves out.

  • The Global South, beyond Singapore and Israel.
  • Enforcement and cross-border litigation realities.

Check yourself.

  1. List the three “broad agreement” principles.
  2. Why does the US position carry more uncertainty than Japan’s?
  3. What are the “overrides” he mentions, and who imposes them?

6. Australia, regulatory arbitrage and the options menu

🎙 In the talk.

  • The arbitrage logic: “once you train the model lawfully in one jurisdiction you’re free to export that model to other jurisdictions” [22:05]. Because “the model is not a copy of the training data”, a rights holder whose works were used in Japan, Israel or the US “could [not] object to the use or indeed reproduction of that model” [22:05]–[23:38] (the transcript has “nothing that a copyright owner whose Works were used in the training data could do to object”).
  • Result: “companies obviously can and will train their models in jurisdictions that are more flexible and then use those models worldwide” [23:38], and strict laws “will only impede their domestic academic institutions” [23:38].
  • Under current Australian law, as he understands it, a non-research model needs “express permission of each and every rights holder” or public-domain-only training [37:44]. At hundreds of millions of items, “the impracticability of this kind of work by work conset approach becomes I think fairly obvious” [37:44]. Australians can use overseas base models and fine-tune on licensed content, “but there could never be an open AI or an anthropic in Australia under current Australian law” [37:44].
  • Options [37:44]–[42:26]:
    1. US-style fair use: recommended by the ALRC (2013) and the Productivity Commission (he says 2017), “it looks like this will never happen” [37:44].
    2. An express fair dealing purpose for TDM or AI training: “this would be an elegant solution” [39:16].
    3. An express TDM exemption, modelled on Japan or the EU directives [39:16].
    4. Collective licensing: “I fundamentally don’t understand how this could ever work” [39:16] (see Deep dive 7).
    5. A safe harbour for training, conditional on an approved AI-safety proposal. Risks (copyright, trademark, privacy, Indigenous sensitivity, publicity) are “much better addressed before training begins than being patched afterwards” [40:50]. It would not shield outputs [42:26].
    6. Do nothing: “I’m not sure that that’s a great option for the clever country” [42:26].

📚 Deeper theory (see 2.6, 2.7).

  • The arbitrage claim has two parts that need separating. (a) Where training is lawful: territorial. (b) Whether the resulting model can be used elsewhere: depends on the model not being an infringing copy. Part (b) is Sag’s inference from his Fact B, and is the part courts have split on (Part 1).
  • Getty v. Stability offers partial support for (b) in the UK: weights were not “infringing copies”. But the court noted that training happened outside the UK, and did not decide the legality of the scraping. The same facts support Sag’s arbitrage point and show its limit: a court elsewhere might ask what happened in the training country.
  • Australia’s actual response (Part 1): in October 2025 the government ruled out a TDM exception and turned to licensing consultations. In Sag’s menu, that is closest to option 6 (do nothing on exceptions) plus option 4 (licensing), the option he was most doubtful about. Whether the licensing work produces a scheme is something I have not checked.
  • Evidence on the “domestic start-ups and universities” claim is thin. I did not look for Australian data on AI research relocation. Not checked.
  • A safe harbour tied to a safety plan is an unusual hybrid between copyright and AI governance. The EU’s approach (transparency summary, copyright policy, opt-out compliance in the Code of Practice) is a loose cousin.

Where others push back.

  • Rights-holder groups in Australia (music, publishing, news) argued a TDM exception would transfer value from creators to foreign tech firms, and the government agreed in Oct 2025 (Part 1).
  • Some argue that Australia’s creative sector is large relative to its AI sector, so the arbitrage point has less pull there than in countries with major AI labs. (An argument, not a measured fact; not checked.)

What he leaves out.

  • The political economy: why fair use “will never happen” (organised creative lobbies).
  • Indigenous Cultural and Intellectual Property: he lists “Indigenous sensitivity” only in passing [40:50].

Check yourself.

  1. Which of Sag’s three facts (A, B, C) does the arbitrage argument depend on?
  2. In his view, which Australian option is “elegant”, and which “politically impossible”?
  3. Which option did the Australian government end up closest to, and what did he say about it?

7. Why collective licensing doesn’t work (his argument)

🎙 In the talk.

  • He says there has been “a lot of discussion” of collective licensing “by which I mean some kind of compulsory license” [39:16].
  • His objection is mechanical. Performing rights organisations work because “they only send money when an artist’s work is actually used”, and in many countries only above a threshold [39:16]. “But with generative AI there are no Taylor Swifts” [40:50]. Nobody can say how much any work contributed, “so the royalties would have to be divided equally” [40:50].
  • Arithmetic: even a billion-dollar fund divided “into a billion equal pieces” would cost more to pay out than it holds [40:50] (he says mailing the cheques “clearly would exceed the revenue of that fund”).
  • His reading of supporters: “most people advocating for compulsory licenses in the us secretly understand this” and are really thinking of “an AI training tax that would go to fund worthy causes” [40:50]. That “might be good social policy but it’s not going to be the same as ASCAP” [40:50].

📚 Deeper theory (see 2.8).

  • The argument is about distribution, not about whether creators deserve money. A flat tax and a flat payout fix the problem he raises by giving up proportionality. Whether that is acceptable is a political choice.
  • Counter-models: (i) bilateral licensing with large rights-holders (publishers, labels, stock agencies), which skips the per-work problem for the biggest owners; (ii) opt-in/opt-out registries; (iii) output-side royalties, paid when generated content is traced to a source (for retrieval-style systems, this is feasible, since the source is known); (iv) extended collective licensing in Nordic-style systems. All from memory; not checked.
  • The Bartz settlement (Part 1) is a data point for equal division: about $3,000 per book, paid to claimants. It was a class settlement for piracy, not a licence, and not a market price for training rights.
  • Sag’s later “Snoopy Solution” framing (Part 1) suggests he sees a place for licensing alongside fair use, perhaps for the high-risk memorization zone rather than all training. I have not read it, so treat this as inference.

Where others push back.

  • Creator organisations: a tax or flat fee is better than nothing, and the administrative cost problem is solved by simple per-work or per-catalogue payments.
  • Rights holders with big catalogues prefer bilateral deals and say the per-work problem does not apply to them.

What he leaves out.

  • Whether licensing is needed for retrieval and output uses, where attribution is possible. This is where the market has moved (news deals; Part 1 notes Japanese publisher suits). His own retrieval point at [29:51] hints at it.

Check yourself.

  1. Why do performing rights organisations distribute efficiently, and why does that not carry over?
  2. What does Sag think compulsory licensing advocates “really” want?
  3. What would change his arithmetic: a smaller recipient pool, or a way to attribute contribution?

Did the predictions hold up?

Judged at 30 Oct 2026. These are my assessments against sources found by search; “partly” means mixed evidence.

What Sag said (2024)What happenedVerdict
In the US, at least some training “and probably quite a lot” is fair use [32:59]Bartz (Alsup) and Kadrey (Chhabria) both found fair use in June 2025 on the records before them. But Ross (non-generative) went against fair use and was affirmed by the Third Circuit in Sept 2026, and the Copyright Office report was sceptical where outputs competeHeld, with qualifications. He said “some uncertainty” and was right; the split is about market harm and about what the data was, which he did not stress
Unlawfully obtained data is not discussedBartz: pirated library not fair use; $1.5B settlementGap in the talk, not a failed prediction. Source of data became decisive
The model is not a copy of its training data [11:03]UK Getty (Nov 2025) agreed for weights. Munich GEMA (Nov 2025) held memorised lyrics were reproduced in the modelContested. Courts split on the same question
Memorization is “very much an edge case” but may grow with scale [25:10]GEMA and the NYT’s examples show it is litigable. Whether it grew from GPT-4 to later models: I did not checkOpen. His warning about GPT-5 was the right thing to worry about
Snoopy problem: characters are the hard case [26:44]He kept developing it (the Nov 2025 Yale talk is named after it). I did not find a decisive court ruling on character outputsPlausible, not tested by what I found
The NYT’s examples may be retrieval rather than memorization [28:17]No merits ruling found. Retrieval-based suits (Perplexity in Japan) are now a separate frontUnresolved
Regulatory arbitrage: strict laws will only impede domestic institutions [23:38]Industry did train abroad (Getty: training outside the UK shaped the result). I did not find data on Australian research impactMechanism supported, effect on Australia not checked
Australia will not adopt fair use (“it looks like this will never happen”) [37:44]Australia ruled out even a TDM exception (Oct 2025)Held, and went further than he allowed for in his options
Collective licensing won’t work as a PRO-style scheme [39:16]Government is consulting on licensing; no scheme found yet. Equal flat payments appeared in the Bartz settlementNot yet tested. His arithmetic objection is unrebutted in what I found
Country list: Japan, UK, Switzerland, EU, Singapore, US, Israel [34:35]UK: dropped any broad TDM plan (2026 report). EU: opt-out regime plus Code of Practice. Japan: still permissive, but publisher suits test its limitsDirection partly reversed in the UK and Australia; the EU and Japan largely as described
TDM rights are “not a blank check” [36:10]GEMA and the Japanese suits are about exactly that limitHeld

Net: his doctrinal map held up better than his policy hopes. Courts accepted the learning frame for training but have attached conditions (lawful source, no memorised regurgitation, market effects). The jurisdictions he thought were converging toward permissive training diverged: Australia and the UK both stepped back.

Putting it together

Mental model. Copyright is a rule about communicating expression, and a training run is mostly a machine reading works to extract patterns no one owns. If the machine ends up holding only patterns, then the copies exist only in the plumbing (the data set), the model is a separate thing, and the outputs are new. That is the case for fair use, for TDM exceptions, and for the claim that a strict country just loses the work. But the story depends on “if set up correctly”. Where the model holds a specific work (memorised lyrics, a famous character, a retrieved article) the plumbing and the product become the same thing, and the argument fails for that work. The live fight is over where that line falls and how much market effect should count, not over whether learning is ever lawful.

Open tensions.

  • “Not a copy” (London) v. “contains the lyrics” (Munich): one technology, two answers.
  • Fair use for training v. market dilution from outputs.
  • Memorization as edge case v. litigation that only needs a few instances.
  • Arbitrage v. each country asking what happened in the training country.
  • Licensing as impractical (his arithmetic) v. licensing as the market’s actual response.
  • Sourcing: scraped “from the open internet” v. pirated libraries.

Reading path.

  • Start: this page’s Summary and Map; then watch [06:20]–[23:38] and [37:44]–[42:26].
  • Then: Sag’s “Copyright Safety for Generative AI” (SSRN); the Bartz and Kadrey summaries linked in Part 1; the Bird & Bird notes on GEMA and Getty.
  • Deeper: Leval (1990) on transformative use; Lemley and Casey, “Fair Learning”; Carlini et al. on extraction; the Copyright Office Part 3 report; the Australian AG announcement and CAIRG papers (primary text not read for this page). For creator-side views see gervais-human-creators-coexist-ai, newton-rex-debbonaire-creativity-ai-fairer-future and newton-rex-ai-companies-creators-rights (not read for this page).

Graph View

Backlinks