Talks

Can Human Creators Coexist With AI? — Daniel Gervais

47 min read

About this talk

Speaker: Daniel Gervais (interviewed by Ben Cashdan)
Event: Reality Check podcast, ep. 1, Geneva Centre on Knowledge Governance
Date: 2026-07-06 (upload date; recording date not stated, see warning box)
Length: 41:57

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

  • 🎙 In the talk. What Gervais or the host says, with [MM:SS] timestamps. Only these parts are his views. Quote marks mean exact words from the transcript.
  • 📚 Deeper theory. Background from the legal, economic and philosophical traditions behind the talk. It explains why the issues look the way they do and who disagrees. It is not his claim unless a timestamp says so. Items marked “from memory; not checked” were not verified.

Suggested route. Summary → Map → Part 1 (terms, current debate) → Part 2: The intellectual map (read once, slowly) → Deep dives in order → “Putting it together”.

Speakers

LabelWhoRole
HostBen Cashdan (captions: “Ben Cashton”)Hosts the first episode of Reality Check. Pushes back, plays devil’s advocate, reads science fiction
GervaisDaniel GervaisLaw professor on IP and AI. The host introduces him as a former researcher at the WTO and WIPO, former Deputy Secretary General of CISAC (the confederation of authors’ and composers’ societies), and author of a history of the TRIPS Agreement. The video description adds: AI advisor to several UN bodies, and author of the sci-fi Coexistence Trilogy

Caption and context warnings.

  • “Z AI Monae” in the clip at [02:58] is Xania Monet, the AI-made R&B act. The chart facts are in the current-debate section.
  • “Ben Cashton” is Ben Cashdan.
  • Recording date is not stated. Upload is 6 July 2026. Gervais says he spoke at a WIPO session “yesterday” and that Google is “about to be worth 4 trillion” and Nvidia is “approaching 5 trillion” [20:40]. Those market-cap figures, and the Anthropic settlement being described as pending-but-agreed, point to a recording some months before upload. From memory, Nvidia first passed $5 trillion in late Oct 2025; not checked. Read his numbers as dated.
  • The host says “a Nobel Prize awarded to two chemists” for protein structure [09:08]. The 2024 Chemistry prize went to three people: David Baker (protein design), Demis Hassabis and John Jumper (structure prediction). From memory; not checked.
  • The Beatles song count (159 or 160) is loose between the two speakers [14:30]. The point is only an example. From memory, the official studio catalogue is larger; not checked.
  • Some sentences in the captions are garbled, for example the missing words at [34:43] (“have a notion of truth the way that humans do”). This page avoids quoting broken lines.

Summary

The central question. Can human creators live alongside generative AI, and what should law do about it? The host frames it as a new “species” arriving. Gervais answers as a copyright lawyer and then as a novelist.

Main thesis, in six steps (all his).

  1. It cannot be stopped. Economics and international competition mean one country banning it changes little [07:13].
  2. The market will be crowded out soon. Machines trained on “the very best, us” will make commercially strong output. In music this becomes hard for humans within “two or three years” [06:27].
  3. Copyright is a weak tool for this problem. Training on copyrighted works is being decided case by case (fair use in the US, exceptions in the EU and elsewhere). But copyright law has “no law for” a system that takes tiny pieces from thousands of works and produces something competitive that is not similar to any one of them [25:19].
  4. Purpose and the copier matter. Research use for science is easy to justify. A pharma company, or a general-purpose model that also writes films, is not [30:24].
  5. Two policy asks beyond copyright. (a) Label things that are 100% AI-made. (b) Consider a “transition remuneration” paid by those who use AI to produce competing output, outside copyright, country by country [35:31]–[37:21].
  6. Journalism is the biggest worry. He would consider arm’s-length public funding for it, but not for music or novels [34:43].

Close. He ties it to his novels: “We need to figure out how to coexist well beyond copyright” [38:15].

Stance to note. He is neither an AI-company defender nor a creators’ maximalist. He says he is “waiting to see” on medical breakthroughs [09:08], rejects a new exclusive right for authors [37:21], yet wants money to flow back to creators and wants provenance labels. He also hedges the “species” metaphor: “I said sort of a species” [40:41].

Map — talking points

#PointTime
1Cold open: “another species”. Clip of AI singer “Z AI Monae” charting; a DJ refuses to play her[00:00]–[03:30]
2Gervais: an inflection point; creators fear for their future. AI learned from “the very best, us” and can find which patterns please[03:30]–[05:54]
3”Within two or three years” the market is crowded out. Can’t stop it. The real worry: will there be new ideas[06:27]–[08:34]
4Promise vs doubt: protein-structure Nobel, medicine (“I’m waiting to see”), climate research[09:08]–[10:42]
5Three kinds of question to ask an AI, the third being manipulation of decision-makers[10:42]–[11:48]
6Text-and-data-mining exceptions in 7–8 jurisdictions. Training-data quality. EU Art. 3 research exemption[11:48]–[13:43]
7US fair use: four factors; Beatles test; Warhol; fourth factor applies only if infringement; input vs output[13:43]–[18:18]
8Pending lawsuits are backward-looking. EU/Singapore-style rules going forward. Big firms have a head start[18:54]–[21:32]
9Can we know what was used? EU summary and its template; US discovery; did the crawler keep URLs[21:32]–[22:50]
10Anthropic settlement: pirated-source books, “probably fair use” for ingestion, list of books, no answer on similar outputs[22:50]–[24:33]
11Market-harm theory (US Copyright Office), the one word from a thousand books problem[24:33]–[26:26]
12Human + AI authorship: prompts, ideas vs expression, “who made the creative choices?”[26:26]–[28:27]
13Feasibility: Copyright Office registration disclosure; “gentle AI”[28:27]–[29:49]
14Host’s summary; Gervais adds: who copies, and mixed-purpose models[29:49]–[31:10]
15Public-interest AI. Costs are falling. Governments as data providers for small-language countries[31:10]–[33:07]
16Barcelona “public AI”. Journalism and the audience-draining loop[33:07]–[34:01]
17Journalism as method; public funding at arm’s length[34:01]–[35:31]
18Agenda: provenance labels; transition remuneration[35:31]–[36:49]
19Who pays; territoriality; not an exclusive right[36:49]–[38:15]
20Coexistence beyond copyright; synthetic training data “failed miserably”; his novels[38:15]–[40:41]
21Close; host still disputes “species”[40:41]–[41:15]

Part 1 — Before you read

Key terms

  • Input side / output side. Input: the copies made to train a model. Output: what the model produces when used. Gervais keeps them apart throughout. Different legal tests apply to each [17:36], [20:07].
  • Fair use (US). A statutory defence, 17 U.S.C. §107, judged on four factors: purpose and character of the use, nature of the work, amount used, and effect on the market. No fixed list of allowed uses. Judges weigh case by case.
  • Text and data mining (TDM) exception. A legal carve-out that lets people copy works to analyse them by machine. In the EU there are two: one for research and heritage institutions (Art. 3) and one for everyone else that rights-holders can opt out of (Art. 4). From memory; Art. 4 is not mentioned in the talk.
  • Opt-out. A rights-holder’s machine-readable “do not mine” signal. The system EU law uses for commercial TDM. The UK consulted on a similar scheme and met heavy resistance (see current debate).
  • Substitution / market dilution. The idea that an AI output need not copy a work to harm its market; it only has to compete with the class of work. Gervais discusses this at [24:33]–[25:51].
  • Transparency of training data. Duty to say what a model was trained on. In the EU, a “sufficiently detailed summary” [21:32].
  • Provenance / labelling. A right to know whether you are dealing with a machine. His first policy ask [35:31].
  • Transition remuneration. His term for money paid back to creators from revenue earned by competing AI output, as a transition measure, not a property right [35:31].
  • Synthetic data. Content made by a model and used to train a model. Relevant because it is a route around copyright [38:50].
  • Public AI. AI built or hosted by public bodies (governments, national libraries, universities), not just by the big tech firms [31:10]–[33:07].
  • Gentle AI. His phrase for AI used to check how much AI is in a work [28:27]. He does not define it beyond that sentence.

People and cases named

ItemWhy it comes upStart with
WIPO Standing Committee on Copyright (SCCR)Where he presented the day beforeWIPO’s SCCR pages (not checked)
Anthropic book-piracy caseTraining fair use vs pirated source; settlementCurrent debate below
Warhol v. Goldsmith (US Sup. Ct., 2023)His fair-use illustration; he notes it turned on factor oneOpinion text
US Copyright OfficeRegistration disclosure; market-harm viewIts AI reports (Parts 2 and 3)
EU copyright directive, Art. 3The research exemption he citesDirective 2019/790
Singapore and EU rulesHis “going forward” modelSingapore Copyright Act 2021 (from memory; not checked)
Barcelona supercomputer (host)A Spanish/Catalan “public AI”Barcelona Supercomputing Center
Asimov (host)Robot laws, “take over”I, Robot (1950)
Severance (Gervais)Analogy for his novel ForeverThe TV series

Gervais’s own frame

His day job is the international law of intellectual property: treaties, exceptions, and how law travels across borders. That explains his two instincts in this talk. First, he looks for what a court will do and keeps saying “we don’t have court opinions” [13:43]. Second, he thinks internationally: a national ban fails, a national copyright levy leaks across borders, so he proposes something that “can be done country-by-country” [37:21]. His other life is fiction, and he keeps both: Forever (book 1), Together (book 2), a third in progress [39:34]–[40:41].

The current debate (to 30 Oct 2026)

Search was run on 30 Oct 2026 (standard mode). Items below are what the sources said, not checked beyond the snippets.

  • Training and fair use: first appeals-court answer. The US Third Circuit affirmed Thomson Reuters v. Ross on 29 Sep 2026, reported as the first federal appellate ruling on fair use for AI training. It held Ross’s use of Westlaw headnotes not fair use. It is a non-generative AI case, so it does not settle generative training (Insurance Journal, 30 Sep 2026, Plagiarism Today, Legal Technology). Opinion not read by me. This is the court opinion that Gervais said at [13:43] was still two years away; it came, for a narrower case.
  • The 2025 district rulings Gervais alludes to. Bartz v. Anthropic (training on lawfully bought books “quintessentially transformative”, but a trial over pirated books) and Kadrey v. Meta (also fair use, with the judge reluctant) (Mondaq summary).
  • The Anthropic settlement is now approved. Reported final approval on 20 July 2026, $1.5 billion, about $3,100 per claimed work, claims for 440,490 of 482,460 listed works. Payment timing was inconsistent across sources and I did not check whether an appeal was filed (Authors Alliance, Authors Guild). The release covers only past acts, which matches his “backward-looking” point [20:07].
  • Europe and the UK went the other way on outputs and memorisation. GEMA v OpenAI (Munich, 11 Nov 2025): memorised lyrics inside a model counted as reproduction and OpenAI was liable (Norton Rose Fulbright). Getty v Stability AI (English High Court, 4 Nov 2025): Getty mostly lost, because the model was held not to store the training works (CMS case tracker). Appeals not checked.
  • Authorship stayed human-only. The US Supreme Court denied review in Thaler v. Perlmutter on 2 Mar 2026, leaving the human-authorship requirement in place (JD Supra). This is background for his “author of nothing” point [27:48].
  • The AI singer in the opening clip. Xania Monet debuted on a Billboard airplay chart (Adult R&B Airplay, dated 1 Nov 2025) and signed a reported multi-million-dollar deal with Hallwood Media. Her creator, Talisha Jones, used Suno to turn her own poems into songs (CTV News, the-decoder). Reports of the deal size vary. My inference: this is the very case where his “who made the creative choices” test would matter, since a human wrote the lyrics but the sound was machine-made.
  • Music went to licensing. Universal settled with Udio (29 Oct 2025) and Warner settled with Suno (Nov 2025), both to move to licensed models. A May 2026 article had Sony still litigating; later status not checked (Music Business Worldwide, Dubspot). This is “negotiate a license” [20:40] happening through litigation, not through a statute.
  • Levies are now live policy, not just his idea. Mistral’s CEO proposed an AI cultural levy in Europe (ITPro); a Dutch government culture council (June 2026) urged AI firms to pay artists (DutchNews); a European Parliament legal affairs committee report on generative AI and copyright was reported in early 2026 (Pulse); India’s trade department floated a mandatory blanket licence (Asia IP Law). Some creator groups objected that a levy could replace the right to refuse. Several of these sources are secondary; check official texts.
  • UK. Reported: the government dropped its active opt-out plan after backlash, and the Lords Communications and Digital Committee published a report on AI, copyright and the creative industries on 6 Mar 2026 (PYMNTS via search). The government’s final response was not found.
  • EU transparency rules are live. Reported: AI Act duties for general-purpose models, including the AI Office’s enforcement powers, apply from 2 Aug 2026; the Digital Omnibus (in force 27 July 2026, per one law-firm source) delayed mainly the high-risk rules (Mondaq, Gibson Dunn). Sources conflicted on details; I did not confirm the training-summary template’s current status.
  • Not found / not checked: any Universal–Suno deal; Perlmutter’s Copyright Office status in 2026; NYT v. OpenAI; the Spanish public-AI project.

Part 2 — The intellectual map

Gervais spoke for 40 minutes to a general audience, so he does not name the traditions behind his claims. Eight families of ideas meet in this talk. This section lays them out.

 COPYRIGHT THEORY ─────────┐                 ┌──── FAIR USE & TRANSFORMATION
 idea/expression, originality│                │     Campbell · Google Books · Warhol
 authorship, Berne           ▼                ▼     "fair learning" · market dilution
                      ┌──────────────────────────────────┐
 TDM EXCEPTIONS ────► │ GERVAIS: input can't be stopped, │ ◄──── COMPENSATION WITHOUT
 UK · Japan · EU ·    │ output competes, copyright is    │       PROPERTY: levies, statutory
 Singapore            │ the wrong tool, so: label + pay  │       licences, CMOs, PLR
                      └──────────────────────────────────┘
                       ▲            ▲             ▲
   ECONOMICS OF        │ JOURNALISM │ AND TRUTH   │ PUBLIC AI & DATA COMMONS
   CREATIVE LABOUR     │ verification, public      │ sovereign compute,
   superstars, Baumol, │ sphere, press subsidy     │ low-resource languages
   model collapse      │                           │
                       SPECIES, COEXISTENCE: sci-fi, Asimov, moral status

The problem it addresses. If a machine writes a song, who owns it? Answering this means going back to why copyright exists and what it covers.

The chain of ideas

  1. The idea/expression split. Copyright protects the particular form of a work, not the idea behind it. The Berne Convention (1886, many revisions) set the international frame, and TRIPS (1994) and the WIPO Copyright Treaty (1996) restate that ideas are outside copyright (WCT Art. 2, from memory; not checked). Key point for Gervais: this is why a prompt is not enough. A prompt is an idea, and the machine supplies the expression, so “you’re still the author of nothing” [27:48].
  2. Originality as human choice. In the US, Burrow-Giles v. Sarony (1884) treated a photograph as authored because the photographer made choices (pose, lighting). Feist v. Rural (1991) set the modern bar: a “modicum of creativity”, with originality shown by choices. Key point: this is where “who made the creative choices?” [27:48] comes from. The same idea runs through Europe’s “author’s own intellectual creation” test (from memory; not checked).
  3. Non-human authors. The “monkey selfie” dispute (Naruto) and Thaler v. Perlmutter (a machine named as sole author) both ended with the answer: no human, no copyright. The Supreme Court declined to hear Thaler in March 2026 (see current debate).
  4. The US Copyright Office’s line on AI-assisted works. Its guidance requires applicants to disclose AI content, and the Office has said prompts alone usually do not give enough control to count as authorship; human selection, arrangement and modification can. From memory; not checked. This is the “registration” practice Gervais cites at [28:27].
  5. The UK exception. The UK Copyright Act (1988) s. 9(3) gives “computer-generated works” an author, the person making the arrangements, for 50 years. From memory; not checked, and the government has been reviewing it. This is a counter-example to the human-only rule he describes. He talks about “pure machine-produced music” having “no copyright protection” [27:02]; that is the US position, not the global one.
  6. Gervais’s own scholarship. He has written on machine authorship; one title I recall is “The Machine as Author” (2020). From memory; not checked, and not in the talk.

What this feeds in the talk. Section 12–13 of the Map: the registration test, the “author of nothing” point, and the idea of “gentle AI” policing the amount of AI in a work.

2.2 Fair use and “transformative” copying

The problem it addresses. Copying is how learning machines work. Is copying for analysis different in kind from copying for competing?

The chain of cases and thinkers

  1. Sony v. Universal (1984), the Betamax case. Making a technology that can be used lawfully is not itself infringement if it has substantial non-infringing uses. Early template for “tools” arguments.
  2. Harper & Row v. Nation (1985). The Court called the effect on the market “undoubtedly the single most important element” of fair use. This is why factor four carries weight in Gervais’s analysis [15:42]. (Later cases softened that ranking; from memory; not checked.)
  3. Campbell v. Acuff-Rose (1994). A parody of “Pretty Woman” was fair use. The test: does the new work add “something new” with a different purpose or character, rather than just replacing the original? This is where “transformative” comes from.
  4. Authors Guild v. Google (2d Cir., 2015). Scanning millions of books to build a searchable index was fair use, because it showed snippets and did not replace the books. The best precedent for AI developers: “copying to build a tool that does something else”.
  5. Warhol v. Goldsmith (2023). The Court narrowed “transformative”: a new work that serves the same commercial purpose as the original (a magazine illustration) is less likely to be fair. Gervais is right that the case was decided on factor one [16:50].
  6. The 2025 AI rulings and the Ross appeal. Bartz and Kadrey (district level) found training fair use on their facts; Ross (3d Cir., 2026) did not, where the output competed directly with the source. The split is about what the model is for, which is where Gervais’s “purpose is relevant” [13:43] lands.
  7. Academic framing. Lemley and Casey’s “Fair Learning” (2021) argues learning from data is not “using” expression. Matthew Sag’s work on “copy-reliant technologies” argues non-expressive copying should be lawful. Opposing view: Lemley’s own later work and others point out that generative models can produce expressive output (from memory; not checked). The sister page sag-global-copyright-copy-reliant-ai covers Sag’s argument.
  8. Market dilution. The Copyright Office’s Part 3 report (pre-publication, May 2025) accepted that a flood of competing but non-copying output can harm the market for human work. The Kadrey judge also floated this. From memory; not checked. It is the legal home of Gervais’s “global impact” worry [17:36].

The thread. Fair use has always asked: does the copy replace the original, or do something else? Generative AI breaks the question because one model can do both at once. Gervais says exactly this about mixed-purpose models [30:24].

2.3 Text-and-data-mining exceptions across jurisdictions

The problem it addresses. Rather than let courts decide case by case (US), some countries wrote the answer into statute.

The chain

  1. UK 2014. A TDM exception for non-commercial research (CDPA s. 29A). From memory; not checked.
  2. Japan 2018/2019. Art. 30-4 of the Copyright Act allows use of works for “non-enjoyment” purposes, which Japan’s government has read as covering training, with limits. Often called the most permissive major regime. From memory; not checked.
  3. EU 2019 DSM Directive. Art. 3 (research and cultural heritage institutions, no opt-out) and Art. 4 (anyone, subject to a rights-holder’s opt-out). Gervais cites only Art. 3 [13:03].
  4. Singapore 2021. A computational data analysis exception. Gervais names Singapore alongside the EU as a model for going forward [20:40]; the exception’s name is added context, not his words. From memory, Singapore’s exception is broad and has no opt-out in the EU sense, so it is not the same design as Europe’s; not checked.
  5. The EU AI Act (2024). For general-purpose models, a duty to have a copyright-compliance policy and to publish a public summary of training content. The Commission’s template for that summary is what Gervais describes as “quite general” [22:17].
  6. Case law on opt-outs. A Hamburg court held in the LAION case (2024) that an opt-out written in natural language could be machine-readable. From memory; not checked.
  7. Memorisation as copying. The Munich GEMA ruling (Nov 2025) said that a model that can reproduce lyrics holds a copy. This pushes against the idea that training is “non-expressive”.

The thread. There are three designs: ask a court (US), permit with an opt-out (EU), permit broadly (Japan, perhaps Singapore). Gervais asks which design the new entrants will face [20:40].

2.4 Compensation without property: levies, statutory licences, collective management

The problem it addresses. When copying is too common and too small-scale to police, law has repeatedly replaced “ask permission” with “pay a fee”. This is the tradition behind his “transition remuneration”.

The chain

  1. Collective management. Organisations such as ASCAP (1914) and PRS (1914) license music en bloc and pay out members. CISAC, where Gervais was deputy head, is the umbrella body for authors’ societies. Source: the host’s introduction at [01:48]; history from memory; not checked.
  2. Statutory (compulsory) licences. The US mechanical licence for recording a song (1909 Act) lets anyone record a published song if they pay a set rate. From memory; not checked.
  3. Private copying levies. Many European countries tax blank media and recording devices and pay creators from the pool. Germany’s levy system dates from the 1960s. From memory; not checked. The idea: you cannot catch every home copy, so you tax the means.
  4. Public lending right. Authors are paid when libraries lend their books (Denmark 1946; UK 1979). From memory; not checked. Pays a living for a use you cannot control.
  5. Extended collective licensing in the Nordic countries: a collecting body can license a whole class of works, including works of non-members. From memory; not checked.
  6. Proposals for the internet. William Fisher’s Promises to Keep (2004) proposed a tax-funded pool shared by use. Neil Netanel (2003) proposed a levy on file-sharing tools. Neither became law. From memory; not checked.
  7. News bargaining. Australia’s News Media Bargaining Code (2021) and Canada’s Online News Act (2023) forced platforms to pay news publishers. Mixed results. From memory; not checked.
  8. The AI wave. The levy proposals in the current-debate section.

The thread. Each system arrived after a technology made the old exclusive right impractical. Gervais’s proposal sits in this line, with one twist: the fee is charged on competing output, not on devices or training [36:49].

2.5 Economics of creative work

The problem it addresses. Why would machine output “crowd out” human work [06:27]? Is that an economic law or a guess?

  1. Superstar economics (Sherwin Rosen, 1981). A few top performers capture most income because copies of the best are cheap to spread. AI adds a new supplier at near-zero marginal cost. From memory; not checked.
  2. Baumol’s cost disease. Live human arts cannot be made faster, so their costs rise against mass-produced goods. AI pushes the gap further. From memory; not checked.
  3. Pattern-finding. Gervais says AI can take the songs that reached the top 10 or 40 and “figure out which patterns are pleasing” [05:09]. That is the logic of hit-prediction in industry: optimise for known taste.
  4. Homogeneity evidence. One study (Doshi and Hauser, 2024, Science Advances) found AI help raised individual story quality but reduced variety across writers. From memory; not checked. This supports his worry about “new ideas” [07:59].
  5. Model collapse. Shumailov et al. (Nature, 2024) showed models trained recursively on model output degrade. From memory; not checked. This backs his claim that synthetic training data “failed” [38:50] — but he cites reports, not the paper.
  6. Counter-view. Past technologies (photography, sampling, synthesisers) also threatened creators and then grew the field. Some economists expect new forms of work. Gervais admits he doesn’t know [07:59].

2.6 Journalism, truth and the public sphere

The problem it addresses. News is a public good that depends on a business model AI may remove.

  1. Habermas, The Structural Transformation of the Public Sphere (1962). A functioning democracy needs a shared space where citizens get reliable information. The press is a pillar. This is the background to “how are we as citizens going to make informed decisions?” [34:43].
  2. Journalism as discipline of verification. Kovach and Rosenstiel’s The Elements of Journalism (2001) define journalism by its method of verification, not by its product. Gervais’s line, that no journalist claims “this is true”, only that they followed the method [34:01], is close to that idea. Note: that is my link, not his citation.
  3. Truth and language models. Harry Frankfurt’s On Bullshit (2005) defines bullshit as speech indifferent to truth. Some philosophers have applied it to chatbots (Hicks, Humphries and Slater, 2024; from memory; not checked). Bender et al.’s “stochastic parrots” (2021) argued similarly. Gervais says AI lacks a human “notion of truth” [34:43] (the caption is garbled; see warning).
  4. The traffic loop. When answers appear inside the AI product, fewer readers click through to the publisher, so the source that feeds the model starves. The host states this loop at [33:07]. Publisher licensing deals and bot-blocking tools (from memory; not checked) are the market responses.
  5. Arm’s-length funding. The BBC licence fee, Nordic press subsidies, and the Arts Council’s “arm’s length” principle (Keynes, 1946) all try to fund culture without letting the state steer it. Gervais names the principle in his own words: “separates the state funder from the actual work” [34:43].

2.7 Public AI and data commons

The problem it addresses. If frontier AI belongs to a few firms, whose languages and cultures does it serve?

  1. Falling costs. Gervais notes that what “used to be over a billion dollars” can be done for less, “as the Chinese examples have shown us” [32:06]. This refers to DeepSeek’s January 2025 claims (from memory; the cost figures were widely disputed; not checked).
  2. Public-sector models. Examples include the Swiss Apertus model (2025) and national projects in Europe built on public supercomputers. The host mentions a Barcelona machine trained on Spanish and Catalan [33:07]; this is probably the Barcelona Supercomputing Center, which sits in a former chapel (from memory; not checked).
  3. Open, licensed corpora. Projects such as Common Corpus and Common Pile assemble public-domain or openly licensed text. From memory; not checked. They are a way to train without the copyright fight.
  4. Low-resource languages. Small-language communities have little digital text, so models serve them badly. India’s Bhashini and the African Masakhane community are efforts in this direction. From memory; not checked.
  5. Commons governance. Elinor Ostrom’s work on commons shows communities can manage shared resources without either privatising or nationalising them. Applying it to data is an open research area.
  6. What Gervais adds. A concrete public role: governments, national libraries and others digitising paper material and providing high-quality data sets [32:06]. That is a copyright-adjacent move: build the supply, not just regulate the demand.

2.8 “Species”, coexistence and fiction

The problem it addresses. Is calling AI a species a fact, a warning or a metaphor?

  1. Fiction as thought experiment. Čapek’s R.U.R. (1920) introduced “robot”; Asimov’s laws (1942 onward) tried rule-based control. The host brings up Asimov at [38:50].
  2. Mind uploading. The premise of Forever, as Gervais describes it: people “transfer whatever makes them a particular person into a body that will live forever” [39:34]. The idea has a long line in transhumanist writing (Moravec, Kurzweil; from memory; not checked).
  3. Companion species. Donna Haraway’s “companion species” (2003) uses species for relationships between kinds of beings, not biology. From memory; not checked. A kind reading of Gervais’s hedge, though he does not cite her.
  4. Moral status debates. Whether AI systems could matter morally is now a live academic and corporate topic. Not discussed in the talk.
  5. The host’s objection. He “will argue” about the word “species” [40:41]. A fair objection: the word suggests independent reproduction and interests; models are made by firms.

Part 3 — Deep dives

1. “Another species”: inflection point, crowd-out, and no way to stop it

🎙 In the talk.

  • Gervais calls this “an inflection point in human history” [03:30]. The host’s frame, which Gervais adopts, is “another species that will be either as or more intelligent than we are” [03:30]. (The words are in the Gervais answer after the clip.)
  • The host objects: the machine is intelligent only because it ate everything we made, “the summation of our species” [04:27]. Gervais agrees: “the big AI models have ingested pretty much the entire human timeline” and “these system learned from the very best, us” [05:09].
  • What AI can do that humans do “intuitively”: take hit songs and “figure out which patterns are pleasing”, then add a persona, an image and a voice [05:09]–[05:54].
  • Timing: “within two or three years it’s going to be very difficult” for human creators to compete [06:27].
  • Stop it? “I don’t think we can stop it even if we wanted to” [07:13]. Reason: the economy and the number of countries involved.
  • His real fear: “will there be new ideas?” The kind of “breakthrough ideas” machines at present “are unable to produce” [07:59]. He adds: “We don’t know is the short answer.”
  • Doubt on benefits: protein structure and medicine are “a future use”; “I’m waiting to see” [09:08].

📚 Deeper theory

  • The inflection claim is an economic one, not a legal one. It rests on superstar and substitution economics (Part 2.5). Gervais does not give evidence for the two-to-three-year estimate; it is a forecast. Treat the date as his judgment.
  • “Can’t stop it” is a claim about enforcement. In international law, one state’s ban binds only its own territory (Part 2.3). He uses the same logic later against a copyright-based levy [37:21].
  • The “species” word. See 2.8. In the talk he hedges it. It does the job of dramatising the policy problem: the question is how to share a market with a non-human producer.
  • “Breakthrough ideas” echoes a long debate. Margaret Boden’s levels of creativity (combinational, exploratory, transformational) is the standard frame for whether machines can do the third kind (from memory; not checked). Gervais gestures at it without naming it.

Where others push back.

  • Optimists say AI tools widen the pool of people who can make music. The Xania Monet case shows a poet reaching an audience she might not have reached. Critics (e.g. Kehlani, in the reporting above) say the human is “doing none of the work”.
  • Skeptics of the crowd-out forecast note that listeners also value a human story behind the art. The DJ in the opening clip refusing to play her is one data point.

What he leaves out. Any figure for the size of the loss. Who currently earns a living from the commercial-music layer he predicts will be crowded out (library music, jingles, session work), and how fast.

Check yourself

  • What is the difference between “we can’t stop it” and “we shouldn’t regulate it”? Which does he say?
  • Why does he worry about “new ideas” rather than about quality?
  • What evidence would you need to test the “two or three years” claim?

2. Research vs entertainment: purpose, the copier, and data quality

🎙 In the talk.

  • The host asks for the “research case” for AI having all the climate research. Gervais: “Absolutely”, and that is why “exemptions” are going to “universities and research institutions in Europe in particular” [09:43].
  • The logic: machines can ingest a whole field; “they will produce something that wouldn’t exist otherwise” [09:43]. (He repeats this at [08:34] as a clip.)
  • Three questions to ask an AI about climate: the broad one, the narrow one (the Venice wall), and the third, “How do we manipulate human decision-makers into accepting to take steps?” [10:42]–[11:16]. He adds that machines “have also learned how to work with our cognitive biases”, and might “on their own use these biases and exploit us” [11:16].
  • Data quality: “one of the big issues in AI is quality of training data.” Train on Twitter comments vs the best medical journals and you get different machines [12:27]. Walls around high-value material mean models “may end up going with the low quality product or even the extremely low quality product” [13:03].
  • EU Art. 3: named institutions, “mostly not non-profit research”, get “a full exemption” [13:03], “with conditions”.
  • US: fair use turns on “the purpose” [13:43]. Mining science for research is not the same as mining the Beatles to put Beatles-sounding music on Spotify [13:43].
  • His late addition: “law tends to factor in who’s making the copying” [30:24]. Two open questions: a model trained for many purposes (medicine and film) makes the analysis “very complicated”, and what if the miner is a pharmaceutical company, not a university [30:24]–[31:10].

📚 Deeper theory

  • The research exemption has a history. Copyright has long treated scholarship favourably: quotation rights, library and archive exceptions, the “fair dealing” for research in Commonwealth law (from memory; not checked). TDM exceptions (Part 2.3) extend that tradition.
  • The “who copies” test is about the dual-use problem. One model, many uses. Courts sometimes judge the copier’s purpose at the time of copying and sometimes the product’s ultimate uses. The Ross appeal (non-generative, same-purpose competitor) shows purpose matching matters.
  • Quality of data is a real technical point. Training-data curation is a major driver of model quality. His claim that walls push developers toward lower-quality data is plausible but he gives no study; mark it as reasoning, not evidence.
  • The manipulation question links to research on persuasion by language models and “sycophancy” (from memory; not checked). It is a safety argument that sits oddly inside a copyright talk, which is part of why it is interesting: he is saying the same corpus that cures can also steer.
  • Pharma as miner. If commercial actors use research-exempt corpora, the exemption (EU Art. 3) would not apply to them; they fall under the opt-out regime (Art. 4; from memory). So his example is closer to settled EU law than he lets on.

Where others push back.

  • AI developers argue that all training is research-like and non-expressive, so the split by purpose is unworkable (Sag’s line; see sag-global-copyright-copy-reliant-ai, not read for this page).
  • Rights-holders argue the opposite: even research use should be paid, because scientific publishers invest in curation.

What he leaves out. Who pays for the high-quality corpora. Publishers hold the best science; “no walls” means no revenue for them. He does not say how to fund scientific publishing while keeping it open to mining.

Check yourself

  • Why does a model with many uses make the fair-use analysis hard?
  • Under EU law as he describes it, would a university-trained model, once released, be usable commercially? (He does not answer; think about it.)
  • What is the third climate question, and why is it surprising in a copyright talk?

3. How US fair use is being argued: input, output, piracy and settlement

🎙 In the talk.

  • Fair use: “the courts will go methodically usually through the four factors” [15:02]. Factor one, purpose; factor two, nature of the work, “less important”; factor three, amount, which “would play against the mining” since they mine 100% but is “a minor factor”; factor four, effect on the market [15:02]–[15:42].
  • The fourth factor test: “you only get to the fourth factor if there’s infringement” [16:50].
  • Warhol: he corrects the host that the court ruled “on the basis of the first factor not the fourth” [16:50].
  • Input and output differ. On the output side, a song like a Beatles song raises a normal infringement claim [16:50]. On the input side “there’s no denying of the copying for the training of the machines”, so factor four is reached, and the dispute is whether to look at “the impact on creators generally” or on “a specific human author” [17:36]. The Copyright Office’s “global impact” idea drew “a significant reaction” [17:36]–[18:18].
  • Lawsuits are “backward-looking”. Ingestion “was not authorized by the right holders”, so the question is whether “an exception like fair use” covers it, and “we’re already seeing the first settlements” [20:07]–[20:40].
  • Going forward: “if you don’t want to end up in court, you should negotiate a license” [20:40].
  • Anthropic: a judge said the pirated source was “relevant”, but also that “the ingestion of books that came from illegal source, were probably fair use”. The company settled, paying per book, with a published list [23:21]–[23:57]. The settlement “still leaves the question of what happens going forward” [23:57].
  • Market harm from non-similar output: “We don’t have a law for a situation where I take from hundreds or thousands of songs I’ve learned from and I produce something that is commercially competitive but not similar as a matter of copyright law” [25:19]. The analogy: “I’ve only taken a word from each book” [25:51].
  • Even if the AI is a tool, “If your song is like a Bob Marley song” you still do infringement analysis [26:26]–[27:02].

📚 Deeper theory

  • The four-factor map is in Part 2.2. His order of operations (infringement first, then fair use, with factor four last) is how US judges often write opinions, but practice varies (from memory; not checked).
  • Why “backward-looking” matters. Settlements and judgments price past acts; they set no rule for new entrants. This is the argument for a statute or a levy rather than waiting for cases.
  • Piracy vs lawful acquisition. The Anthropic ruling split “use” from “acquisition”: training could be fair, building a library from pirate sites might not be. The settlement price (about $3,000 per book, per the reporting above) became a benchmark. Not in the talk beyond “pay this amount per book”.
  • The “word from a thousand books” analogy is a point about copyright’s unit of analysis. Infringement is tested work by work. Aggregate harm from many small takings has no home. Compare how antitrust or consumer law handle diffuse harm (my parallel, not his).
  • Memorisation changes the picture. The Munich ruling treats a model that can output lyrics as holding a copy. If the model can regurgitate, the “tiny pieces” analogy weakens. He does not discuss memorisation.
  • Traceback technology. At [25:19] he mentions a company at WIPO claiming to trace a song back to its sources. His analogy is meant to show tracing doesn’t fix the legal problem. Interpretability research on influence tracing is an open field (from memory; not checked).

Where others push back.

  • AI companies: training is non-expressive “learning” and fair; pirate-source questions are about the acquisition, not the training.
  • Creators’ groups: the settlement figure shows that licensing is feasible, so factor four should favour them.
  • Some scholars say US courts are unlikely to adopt “market dilution”, because factor four traditionally concerns a substitute for the specific work (from memory; not checked).

What he leaves out. The Kadrey judge’s market-dilution discussion, by name; contract and terms-of-service claims; and non-US perspectives on the “pirated copy” point. Also, how a US appellate court (now the Third Circuit in Ross) would treat his analysis, since the talk predicted appeals were two years off.

Check yourself

  • Why does he say the fourth factor is reached on the input side but not (in the same way) on the output side?
  • Explain the “one word from a thousand books” analogy and what legal rule it exposes.
  • What does the settlement settle, and what does it leave open, in his words?

4. Transparency: can we know what is inside the model?

🎙 In the talk.

  • Host: “Can you know exactly whose work is in there and who deserves how much compensation?” [21:32].
  • Gervais: even big models are retrained, so “it is possible for smaller models to emerge” [21:32]. Some licensed or shared data sets “already” exist [21:32].
  • Europe requires “a quote sufficiently detailed summary”, and the template is “quite general and it does not include work-by-work list of things that were ingested” [22:17].
  • US: discovery in litigation “may get us some information” [22:17].
  • Two possible states of knowledge: developers who gave the machine a set “probably know”; if the machine crawled, did it keep “a record of every URL” [22:50]?
  • Transparency is also the subject of the first half of the WIPO session [02:22].

📚 Deeper theory

  • Why transparency is a gateway right. Without knowing what was used, a rights-holder can neither opt out, claim a share, nor sue (the Anthropic book list helped class members file claims). Transparency is the precondition for every other remedy in this talk.
  • The EU model. AI Act obligations for general-purpose model providers (summary of training content, copyright policy) are in Part 2.3. A summary is far short of a list; that is the compromise he describes.
  • Practical limits. Web-scale crawls contain billions of documents, and deduplication and filtering steps change the set. Even honest developers may not hold a clean record (my point, plausible; not checked).
  • Tracing outputs to inputs (influence functions, attribution) is a research area; it may help allocate money under a levy. He is doubtful that tracing solves the legal problem [25:51].
  • Comparison. Food and drug labels, financial disclosure and environmental reporting are all transparency regimes that came before the substantive rule. Copyright is now being asked to do the same.

Where others push back.

  • Developers cite trade secrets and security, and say a full list would reveal competitive choices.
  • Creators say a summary is useless for enforcement and want per-work disclosure.

What he leaves out. How a creator would use the summary; and what happens to models trained abroad but offered in the EU (the AI Act’s reach is a live question, from memory).

Check yourself

  • Why does transparency matter more under an opt-out regime than under a licence regime?
  • What is the difference between a summary and a list, and who benefits from each?
  • Name two places the missing record could come from, per Gervais.

5. Who is the author of AI-assisted work?

🎙 In the talk.

  • Host: if I use an AI tool “that’s been trained with all that music” and I massage and prompt it, “I’m the creator”, right [26:26]?
  • Gervais: if the song is like a Bob Marley song “you still do copyright infringement analysis” regardless of how it was made [26:26].
  • What changes: “if it’s pure machine-produced music, there’s no copyright protection on the output” [27:02]. With a human in the process, protection “depends on the nature of the involvement” [27:02].
  • Prompts: “I’m the author of nothing. I’ve just prompted. Copyright doesn’t protect ideas” [27:48]. You have copyright in the prompt “but you’re not going to sell the prompt” [27:48].
  • The test: “Who made the creative choices? The machine or the human author?” and if both, make sure “there’s enough human creative choices in the expressive output” [27:48].
  • Feasibility: the US Copyright Office “does it on a daily basis” because you must “declare and explain” AI use when registering [28:27].
  • Future: “The future is probably a gentle AI”, with “one AI checking how much AI there is” in a work; and patterns of use will “become a little bit more along certain patterns” [28:27].

📚 Deeper theory

  • Foundations in Part 2.1: idea/expression, originality as choice, the human-authorship rule, the UK exception.
  • Degrees of control. Prompting a model is like commissioning, not painting. Copyright has an old analogue: a commissioning party who gives instructions is not an author; the person who gives the expression is. Gervais’s “ideas” language is this principle (my framing).
  • Iteration blurs the line. A person who selects among hundreds of outputs, edits, composes and arranges them may add enough human expression. The US position (from memory; not checked) protects the human selection and arrangement, not the raw machine output.
  • Global divergence. China has had court decisions granting protection to AI-assisted images where a person made detailed choices; the UK statute covers computer-generated works (both from memory; not checked). So “no copyright” is not universal.
  • Disclosure as governance. Registration forms that ask about AI use create a record and a norm. “Gentle AI” would automate the check (his idea; not elaborated).
  • Competitive effect. If pure machine output is unprotected, it can be copied freely. This makes some AI-music business models depend on human elements (voice, lyrics, branding) to hold rights. This is my inference, tied to the Xania Monet case.

Where others push back.

  • Some creators argue that heavy prompting and iteration is real authorship and that the line is arbitrary. Others argue any AI output should not be protected, to avoid flooding the system.
  • The host doubts anyone wants to be “the arbiter of every work” [28:27]; Gervais answers with current practice and a hope for automation.

What he leaves out. Cross-border divergence; the cost of examination at scale; whether disclosure can be faked; and how this interacts with platforms’ own labelling.

Check yourself

  • Why can you own the prompt but not the output?
  • What would count as “enough human creative choices”? Give an example.
  • What is “gentle AI”, and what problem does he think it solves?

🎙 In the talk.

  • Two items on his agenda beyond exceptions: transparency of provenance and, because current copyright law cannot reach “human authors that are affected by these new competitive outputs”, “transition remuneration for authors going forward” [35:31].
  • The design: “out of the revenues generated by the competitive outputs”, some “will make its way back to the creators of the material” [35:31].
  • Host: “tax the large AI companies”. Gervais: “tax whoever is using AI to produce competitive output”, not scientists, and “not tax people, if you want to use that word, at home using AI” [36:49]. It is aimed at those whose business is content “meant to displace the humans that were occupying this place” (paraphrase of a garbled passage).
  • Territoriality: copyright-based demands (Brazil, Indonesia, per the host) are “very tricky because then the AI can just be done in another country” [37:21]. He says “I’m not saying there should be some new exclusive right for authors” and the scheme he means “can be done country-by-country”, paid by anyone operating such a system there “even if it wasn’t trained here” [37:21].
  • His own caution: “it is a solution to this competition we’ve never had before from non-humans producing the kind of thing that has defined us as humans” [38:15]. He prefaces this with doubt about whether it is “a great solution”.

📚 Deeper theory

  • Lineage in Part 2.4: private copying levies, public lending right, statutory licences. His innovation is the base: not devices or training, but competing output revenue.
  • Why not a property right? Exclusive rights are territorial and need enforcement against each infringing act. A remuneration right (a claim to payment) can attach to the market in which the product is sold. This is why levies work for private copying.
  • Design problems. (1) Defining “competitive output” without catching ordinary users. (2) Measuring revenue from AI-generated content. (3) Distributing money: by catalogue share, by listening data, by collective societies? Collective societies (CISAC members) are the obvious pipes, which is Gervais’s background, though he does not claim that. (4) Interaction with licensing: if firms already pay labels, does the levy double-count?
  • Current proposals in the debate section (Mistral levy idea, Dutch advisers, EU Parliament, India). Differences: some levy training; Gervais’s targets output revenue. Not the same base; compare carefully.
  • Critics’ worry (reported in the debate section): a levy can become a price for being scraped, replacing the right to say no. Gervais’s text does not say whether creators could still opt out.
  • “Transition” matters. The word implies a time-limited bridge while the market and law adapt. He does not give a duration.

Where others push back.

  • AI developers: a levy on output penalises use, and double-charges for content that was licensed.
  • Creator groups: want consent and licensing first, not a pool.
  • Economists: levy pools often reward the biggest catalogues and need heavy administration (from memory; common critique; not checked).

What he leaves out. A rate, a base definition, a collecting body, a sunset, and how it relates to licence deals that labels and publishers are striking now.

Check yourself

  • Why does he avoid creating a new exclusive right?
  • Which output revenue would be taxed under his scheme and which not?
  • How is this different from a levy on training? Give one advantage of each.

7. Journalism, public AI and the audience loop

🎙 In the talk.

  • Host raises “public interest AI”, with governments, communities and local initiatives so it is not “four or five big corporations” [31:10].
  • Gervais: it was too expensive; costs fall “like DNA sequencing”, and “what used to be over a billion dollars can be done for probably less now” [31:10]–[32:06]. Whether governments would build or “partner with one of the big technology companies” is open [32:06].
  • A big advantage: governments “providing high-quality data sets to researchers”, especially for small countries and languages few people speak [32:06]. Paper-only material cannot be used, so “there’s a role for governments, for national libraries, and others to get involved” [32:06].
  • Host describes a Barcelona supercomputer in a church training on Spanish and Catalan: “public AI” [33:07]. The host then states the news loop: AI sites suck audiences away from sources, and “nobody will be left going to the places that need the business model” [33:07].
  • Gervais: “Journalism is my biggest worry” [34:01]. Journalists get information out and analyse it, with “methods” used “at great expense” [34:01].
  • Truth: “They can only say I’ve gone through the method of journalism” [34:01]. AI, he says, lacks a human notion of truth (caption garbled at [34:43]).
  • Policy: “I would never say government should be in the business of producing music or novels. But in the case of journalism, if there is a way to do some public funding of journalism in the way that separates the state funder from the actual work so that it’s done independently, if there’s a way to do that, I think we should try” [34:43]. Alternative: “Having machines do all of this?” and “the cost to us as citizens will be enormous” [35:31].

📚 Deeper theory

  • Public sphere and verification (Part 2.6). He treats journalism differently from music because its output is infrastructure for self-government, not entertainment.
  • Why not licensing? Licensing deals between publishers and AI firms exist, but they favour large publishers and do not fund local news (from memory; not checked). Arm’s-length public funding addresses what licensing misses.
  • Public AI and data supply (Part 2.7). His point is that supply-side policy (digitisation, open datasets) gives small-language communities what regulation of demand cannot.
  • The loop is a classic platform dynamic. Aggregators capture attention but not the cost of producing content, as with search snippets before. Australia and Canada tried bargaining codes (Part 2.4).
  • For India (my gloss; not in the talk): the same questions appear for Indian-language news and for Bhashini-style public datasets. The Indian trade-department blanket-licence idea in the current debate is a remuneration scheme in Gervais’s family.

Where others push back.

  • Press-freedom groups worry any state funding invites capture, even at arm’s length. Others say the existing BBC and Nordic models show it can work with strong safeguards.
  • Tech firms argue AI summaries drive traffic and that chatbots can cite sources.
  • On public AI: critics say public bodies lag on talent, and “public AI” can mean state surveillance, depending on the country.

What he leaves out. How to define “journalism” for funding without the state deciding; how to fund non-English local news; and how public AI would obtain rights to copyrighted text, since public bodies also face copyright limits.

Check yourself

  • Why does he treat journalism differently from music and novels?
  • What is the audience loop the host describes?
  • What role can a government play in public AI that doesn’t involve writing the model?

8. Synthetic data, the book series and “sort of a species”

🎙 In the talk.

  • Host asks: what if AI is trained on AI music and the link to human creativity is “severed” [38:50]?
  • Gervais: companies that tried synthetic content to avoid the copyright issue “have so far failed miserably according to reports I’ve read” [38:50]. “It looks like there’s something to be said for human creation even if you’re a machine” [39:34].
  • Novel 1, Forever: “people can transfer whatever makes them a particular person into a body that will live forever”, like Severance; the new beings “have their own life”, and “the idea that humans can control these beings gets challenged” (paraphrase of [39:34]).
  • Book 2, Together: “the two species realize that they need to find a way to coexist” [39:34]. Book 3 in progress; “a lot of adaptation will need to happen” [40:41].
  • Host: will keep arguing about calling AI a species. Gervais: “I said sort of a species” [40:41].

📚 Deeper theory

  • Model collapse (Part 2.5): recursive training can degrade models. Gervais cites “reports”, not the research. The effect depends on how synthetic data is mixed with real data (from memory; not checked), so “failed” is stronger than the evidence.
  • Why this matters for copyright. If synthetic data worked fully, models could be trained without touching protected works and the input-side fight would shrink. If it fails, human-made material remains a scarce input and its owners have bargaining power. That is the economics of the settlements.
  • Fiction as policy prototype. Using novels to rehearse social arrangements has precedent (Asimov, Part 2.8). Gervais keeps it separate from his legal work: the novels are where “coexistence” goes beyond copyright [38:15].
  • “Species” is hedged. The strongest reading of his sentence is that it is a metaphor for needing a shared set of rules with a new kind of producer.

Where others push back. Several labs now report using synthetic data successfully for narrow tasks (from memory; not checked). The host’s objection to “species” is also a standard one: models have no biology, reproduction or independent interests.

What he leaves out. Which reports he means, and how synthetic data differs for text vs music.

Check yourself

  • Why would working synthetic data weaken the input-side copyright fight?
  • What does “so far failed” imply about the future?
  • Is “sort of a species” a claim or a way of speaking? Cite his wording.

Putting it together

Mental model. Gervais sees a one-way machine: human works go in at scale, competing output comes out at near-zero cost, and copyright is built to police copies of single works, so it catches the input in court slowly and the output almost never. Law can still do four things. (1) Let courts sort the past (fair use and settlements, backward-looking). (2) Set simple forward rules where legislatures act (research exemptions, transparency, licences). (3) Require provenance, so people know what is machine-made. (4) Pay creators back from the revenue of competing output, in a scheme outside copyright, and build public supply (data, funding for journalism) where markets fail. Throughout, he separates what machines may learn from what they may sell.

Open tensions.

  • Opt-out vs levy. Does payment replace the right to refuse? He does not say.
  • Purpose vs product. Research use is easy to justify, but one model serves both science and entertainment.
  • Tracing vs sampling. If influence cannot be traced to works, who gets paid?
  • Public AI vs capture. Public data and funding can help small languages and journalism, or entrench the state.
  • Forecast vs evidence. “Two or three years” is a prediction; the 2026 rulings show law is moving but unevenly across jurisdictions.

Reading path

  • Start: the Authors Alliance or Authors Guild note on the Anthropic settlement; a plain-English summary of the EU AI Act’s copyright duties; the Copyright Office’s AI reports (Parts 2 and 3).
  • Then: Warhol v. Goldsmith (2023); Authors Guild v. Google (2015); Campbell v. Acuff-Rose (1994); the Third Circuit opinion in Thomson Reuters v. Ross; the Munich GEMA judgment and the Getty v Stability judgment.
  • Deeper: Lemley and Casey, “Fair Learning” (2021); Matthew Sag on copy-reliant technology (see sag-global-copyright-copy-reliant-ai); Fisher, Promises to Keep (2004); Kovach and Rosenstiel, The Elements of Journalism; Habermas, The Structural Transformation of the Public Sphere; Shumailov et al., “AI models collapse when trained on recursively generated data” (Nature, 2024); Gervais’s own Forever and Together (fiction). Sister talks on creators’ rights: newton-rex-ai-companies-creators-rights, newton-rex-debbonaire-creativity-ai-fairer-future. (Not read for this page; link only.)

Graph View

Backlinks