Talks

Creativity and AI: How Can We Build a Fairer Future? — Newton-Rex and Debbonaire

52 min read

About this talk

Speaker: Ed Newton-Rex and Baroness Thangam Debbonaire
Event: UKAI House at SXSW London, fireside chat, 83 Rivington Street (filmed Wed 3 June 2026)
Date: 2026-06-03
Length: 40:24
Channel: Practical AI strategy for UK business leaders
Watch: Creativity and AI: How Can We Build a Fairer Future? | UKAI House at SXSW London

How to read this page. Two kinds of content, always marked:

  • 🎙 In the talk. What the two speakers say, with [MM:SS] timestamps. Words inside quote marks are exact transcript words. Only these parts are their views.
  • 📚 Deeper theory and context. Background, law, and counter-arguments I added. It is not their claim. 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 “Putting it together”.

Sister pages from the same batch: gervais-human-creators-coexist-ai, newton-rex-ai-companies-creators-rights, sag-global-copyright-copy-reliant-ai. I did not read them while writing this page.

Speakers

The captions are not speaker-labelled. I assigned turns from what each person says about themselves, so a few turns are inferred.

LabelWhoRole in this talk
DebbonaireBaroness Thangam Debbonaire (named in the video description; the captions spell her first name “Fangum”)Opens, and leads the questions. Says she has been a musician for most of her life, is chief executive of an opera body (captioned “UK Opera Association”; name not confirmed), and chairs the Women’s Prize for Non-Fiction. Was Labour’s shadow culture secretary before the July 2024 election. Wants a policy view from the opposition-turned-peer side
Newton-RexEd Newton-Rex (named in the description; the captions write “Ed”)Answers on the history of training data, the UK policy fight, and enforcement. Says he runs a nonprofit that certifies AI companies and their training data (“fairly trained”, captioned in lower case). Says he started a generative-AI company in 2010
Host”Tim”, from UKAI (surname not given in the captions)Thanks them at the end and keeps time
AudienceTwo questioners, not namedQ1: what can creative industries do in practice? Q2: how could rules be enforced?

Caption warnings. Auto-captions garble names. Corrections used on this page, inferred from context: “Fangum” = Thangam; “Chris France” = Chris Bryant (the speaker names him as “Chris Bryant” elsewhere); “Pete Kyle” = Peter Kyle; “Baroness SB Vancedron” = Baroness Kidron (the filmmaker peer who tabled the Data Bill amendment; the captions also write “BBA ban amendment”, likely “Beeban” or “Kidron amendment”); “sax’s licensing system” (at [34:14]–[35:32]) is probably CLA / ALCS / PLS-type collective licensing (the speaker lists PLS, DACS, PRS in the same breath); “Demisarabis” = Demis Hassabis; “11 Labs” = ElevenLabs; “Czechov” = Chekhov; “dies it” and “dissit” = DSIT (the Department for Science, Innovation and Technology); “doulets” at [07:25] is unintelligible. Section [34:14] to [35:32] is badly garbled; I quote very little from it.

Possible factual slips and loose claims in the talk (checked where marked):

  • Debbonaire says the 1710 Statute of Anne “was the first to introduce a scheme of licensing” [00:04]. The Statute of Anne gave authors (not only printers) a time-limited right to their printed works. The earlier press-licensing regime was the Licensing of the Press Act 1662, which lapsed in 1695 (from memory; not checked). The sentence in the captions is also broken, so treat the exact point as unclear.
  • Newton-Rex says Anthropic “is about to be IPO for over a trillion dollars” [02:12]. Reports say Anthropic filed confidentially for a stock-market listing on 1 June 2026, after a round at about $965 billion (The Next Web). That is just under $1 trillion on 3 June, so “over a trillion” is loose, and the listing had not happened.
  • He says Claude 3 (March 2024) was the first big Claude model not trained on pirated books [02:12]. I could not verify this. The court record in Bartz v. Anthropic describes Anthropic buying and scanning print books from 2024 (from memory; not checked). Treat his “by my reckoning” as his own estimate.
  • He says GPT-3 “added for the first time Library Genesis” [01:37]. It is widely believed that the “Books2” dataset used for GPT-3 was largely LibGen, but OpenAI has not confirmed it (from memory; not checked). His claim is an inference.
  • “one in four illustrators said they had lost work” in a “poll in January 2024” [19:54]. I recall a Society of Authors survey in 2024 with about 26% of illustrators reporting lost work (from memory; the month and poll not checked).
  • Debbonaire says the ministers “both said in January it’s coming” [35:32] (a regulation). Not checked.

Summary

Central question. Debbonaire frames it in her opening: where does AI leave writers, artists, musicians and composers, and how did the UK get to the point where policy seems to side with the AI companies? The title asks what a “fairer future” looks like.

Main thesis (Newton-Rex). The AI industry is “born of piracy”. It built its models on copied books, news and music without asking. The fair answer is not to stop AI but to build a “responsible AI ecosystem” in which: companies license what they train on, they disclose what they trained on, and governments do not use taxpayer money or regulatory silence to back companies that did not.

Main thesis (Debbonaire). A creator’s work has rights from the moment it exists, and the right to say yes must come before use (“opt in, not opt out”). Artists are the “canaries in the coal mine”: the same questions about jobs, consent, and who gets a say will reach everyone. She wants the public asked, not told.

The UK story they tell. Labour in opposition (2023) said it would protect creatives. In December 2024 the new government’s consultation preferred a copyright exception with an opt-out. A thousand musicians released a silent protest album. Lords amendments in the Data Bill failed. A promised AI bill did not arrive in the King’s Speech “two weeks ago” (they say). The Sovereign AI Fund will not say whether it will back firms that train on unlicensed work abroad.

Close. Practical: join your sector’s licensing system, keep up pressure on MPs, and demand transparency requirements, which they see as the key to enforcement. Debbonaire’s last line is an analogy: builders stealing bricks at night will eventually be arrested.

Stance to note. Both are strongly for creators, and both say plainly that they are not against AI. Newton-Rex: “AI is clearly a powerful technology”. Debbonaire says she does not want to say “shut it down”, though she admits that thought arrives daily. This is an advocacy conversation, not a debate. No one in the room argues the AI-company side. See Part 3 for the strongest counter-arguments.

Map — talking points

#PointTime
1Opening: creators are fearful; copyright exists from the moment of creation (Statute of Anne, 1710)[00:04]
2Newton-Rex: “born of piracy”. GPT-1 on 7,000 books, GPT-2 on 8m scraped documents, GPT-3 adds LibGen. We know only because of lawsuits[00:04]–[02:12]
3Anthropic: co-founder downloaded pirate libraries at OpenAI, then again at Anthropic; early models trained on them[02:12]
4Debbonaire’s background; the “artists are inspired by artists” line from “tech bros”[03:14]
5Human learning vs machine learning: copies, scale, labour displacement[04:21]–[04:56]
6Three principles from Labour’s 2024 position: intrinsic rights, opt in not opt out, recognition and reward. Artists as canaries[04:56]–[06:53]
7The policy story: Chris Bryant in 2023, Peter Kyle, the Dec 2024 document (“preferred option” = opt-out)[06:53]–[09:44]
8The backlash: Elton John, the silent album with 1,000 musicians, Kate Bush, Max Richter[10:16]–[11:48]
9”Who changed it?” Meetings lists, Number 10, the Kidron amendment, donors[11:48]–[14:09]
10The King’s Speech and no AI bill. Vacuum. Licensing systems exist (Creative Content Exchange, Publishers’ Licensing Services)[14:09]–[15:42]
11Reality check: few firms license; OpenAI, Anthropic, Runway, ElevenLabs. Loophole: train abroad, import[15:42]–[18:13]
12”Ask what they train on.” GPT-1 authors and the flood of AI books on Amazon; AI competes with its sources[18:13]–[19:54]
13Illustrators losing work; canary argument widens to all jobs; the public was never asked; UKAI as critical friend[19:54]–[21:46]
14Transparency for readers: “I still want to know where I’m reading Czechov”; what is the role of humanity; growth at what cost[21:46]–[23:13]
15Why is government afraid? The AI race with China, growth narrative, advisers with skin in the game (Hassabis, VCs, the Sovereign AI fund, AI Action Plan)[23:13]–[26:30]
16Debbonaire: who has access; the Google lunch; “I can help you cure all cancer”; a tattoo with the poet’s permission[27:08]–[29:43]
17Living out belief: the Sovereign AI Fund non-answer; “just don’t back them”[30:49]–[33:18]
18Q1: what can creative industries do? Licensing systems, sector unity, keep pressure on, ask where the promised regulation is[33:18]–[35:32]
19Q2: how to enforce? Transparency requirements; list of training sources; audits; the “secret sauce” objection; lawsuits fear[37:04]–[39:52]
20Close: the builders-stealing-bricks analogy[39:52]

Part 1 — Before you read

Key terms

  • Copyright (UK). A property right that arises automatically when an original work is fixed (written, recorded, drawn). No registration is needed. It covers copying, adapting, and communicating the work to the public. Debbonaire’s whole case rests on this automatic vesting [10:16]. (From memory: Copyright, Designs and Patents Act 1988, “CDPA”.)
  • Training data / pre-training. The large body of text, images, audio and video a model learns from. Newton-Rex says in [38:21] that the distinction between pre-training and reinforcement learning matters less than he says AI companies claim; he asks only for the list of sources.
  • Opt-out vs opt-in. Opt-out means the use is allowed unless the owner objects. Opt-in means the use is banned unless the owner agrees. They move the burden of action. Watch out: “opt-out” is not one thing. A rights reservation in the EU sense is a machine-readable signal; the UK 2024 proposal also floated this. Many creators have no website or technical means to send such signals.
  • Text and data mining (TDM) exception. A legal carve-out allowing copying for analysis. UK law has one for non-commercial research only (from memory: s29A CDPA, 2014; not checked). The EU’s 2019 Digital Single Market Directive has a broader commercial TDM exception, with an opt-out for rights holders (Article 4; from memory; not checked). The 2024 UK proposal was about adopting something like the EU model.
  • Fair use vs fair dealing. US law has a flexible “fair use” defence. UK law has a closed list of “fair dealing” purposes. This is why US training cases and UK law do not map neatly, and why Newton-Rex’s train abroad and import worry arises [18:13].
  • Licensing. Paying for permission. For a single user this is a contract. For many creators it is often collective: a licensing body (in the talk: PLS, DACS, PRS, and the Creative Content Exchange) negotiates with the user and shares the money.
  • Transparency requirement. A legal duty to disclose what data a model was trained on. Newton-Rex’s central ask. Different from attribution of single outputs.
  • Fairly Trained. The nonprofit he mentions at [15:42]. It certifies AI companies whose training data is licensed or otherwise permitted, based on a list of sources they provide (from memory: launched January 2024; the page does not need the date).
  • Sovereign AI Fund. A UK government fund described in the talk as “500 million quid” to back British AI companies [30:49].
  • Creative Content Exchange (CCE). A government-backed platform idea for licensing cultural content to AI developers. Mentioned at [15:42]. See Part 1, current debate.
  • Canary in the coal mine. Debbonaire’s metaphor [06:53, 29:43]: creative workers meet AI’s effects first, so their treatment predicts everyone’s.

People and bodies named (a reading list)

Person / bodyWhy they come upStart with
Chris BryantLabour minister; tweeted a pro-creator line in 2023, then authored the “preferred option”His 2023 posts; Hansard debates on the Data Bill
Peter KyleThen Science Secretary (DSIT); shadow DSIT earlierThe Dec 2024 consultation
Baroness Beeban KidronFilmmaker peer; tabled the Lords transparency amendmentHer Lords speeches (e.g. 1 June 2026 debate, linked below)
Elton John, Kate Bush, Max RichterProtest voicesThe silent album (see below)
Demis HassabisGoogle DeepMind head; cited as an adviser with a stakeHis public statements on AI and growth
Ben MannAnthropic co-founder; named for downloading pirate librariesCourt records in Bartz v. Anthropic
Creative Content Exchange; PLS; DACS; PRSLicensing pathsTheir own sites
UKAIThe hostIts own site

The speakers’ own frame

Newton-Rex is a former composer and AI-music founder who turned campaigner. In the talk he says he started a generative AI company in 2010 [23:13] (from memory: Jukedeck, sold to ByteDance in 2019; not checked). His frame is consent plus disclosure: licensing is the norm, and secrecy is the only thing keeping lawsuits away [38:21]. Debbonaire’s frame is rights plus democracy: creators hold rights, and the public should have a say [20:41]. She also connects it to the wider question of “the role of humanity” [22:22].

The current debate (up to 30 Oct 2026)

This is where things stand as of today. Sources are linked; my searches could not confirm everything.

  • The opt-out is dead as policy, for now. On 18 March 2026 the government published its report and economic impact assessment on copyright and AI, as the Data (Use and Access) Act 2025 required (gov.uk impact assessment). Law-firm commentary says the government dropped its preferred broad exception with opt-out and has no preferred option (Reed Smith, Bratby, City AM). The assessment itself called the evidence “limited and uncertain” on how copyright affects AI development, and estimated AI-sector GVA of £20–90bn by 2030 (as reported by secondary sources).
  • The government admitted it got it wrong. Culture Secretary Lisa Nandy was reported saying “we got it wrong” about the first approach, citing consultation results where 88% of about 11,500 respondents wanted licences required (Screen Daily). The 88% figure comes from a secondary summary; check the primary source before quoting it.
  • No replacement policy yet. The March report said ministers would not legislate until confident that reform serves both sectors. It promised to test commercial licensing through the Creative Content Exchange, with a pilot platform expected by summer 2026, plus later consultations on digital replicas and a labelling taskforce interim report due in autumn 2026 (as reported by secondary sources, see Pinsent Masons timeline).
  • The Lords sided with licensing first. A House of Lords Communications and Digital Committee report (March 2026) backed a licensing-first approach and statutory disclosure of training data; the government replied in May 2026 (HSF Kramer).
  • The Sovereign AI Fund question has not been answered. Lord Clement-Jones asked in April 2026 whether firms must show lawful access to training data to receive fund support; Baroness Lloyd of Effra replied on 28 April that the fund operates within the existing legal framework. On 1 June, two days before this talk, Baroness Kidron pressed the point in the Lords chamber, and the reply repeated “applicable UK law” (Complete Music Update, Lords debate, 1 June 2026). That matches Newton-Rex’s account at [31:42]. I found no sign the fund has an enforceable licensing condition.
  • The King’s Speech. The speakers say no AI bill was in it [14:09]. I could not find any source on the content of the May 2026 speech, so this is not verified.
  • New political leadership? One law-firm piece dated 8 October 2026 refers to a Creative Industries AI Summit in Manchester convened by “Prime Minister Andy Burnham”, with a founding document called the “Manchester Principles” on minimum conditions for using human creative work in training (Lewis Silkin). I found only this one source, and no coverage of what was announced. Treat as unverified before relying on it.
  • Courts. The Bartz v. Anthropic class settlement of $1.5 billion (about $3,000 per work) reportedly received final approval on 20 July 2026, with a release for past conduct only (Pearl Cohen). This is the “lawsuits” evidence Newton-Rex points to at [02:12]. In the UK, the High Court (Nov 2025) rejected the main copyright claim in Getty Images v. Stability AI on the ground that model weights do not store copies, per secondary sources; an appeal was reported but I could not corroborate its timing.
  • Anthropic and the listing. Newton-Rex’s remark on a coming IPO is real in outline: a confidential filing was reported on 1 June 2026 (The Next Web). One source claims a public filing on 26 October 2026; I could not confirm it.
  • What I did not find. No post-summit reporting; no sign that the transparency duty Newton-Rex wants has been legislated in the UK.

Part 2 — The intellectual map

The talk is 40 minutes of advocacy, mostly for people who already share its premises. Behind each claim sits a body of law or theory. Eight families of ideas meet here.

 COPYRIGHT THEORY ─────────┐                   ┌──── LAW OF TRAINING
 Locke · utility · Hegel   │                   │     Sag · Lemley · fair use
 Statute of Anne · Berne   ▼                   ▼     cases · TDM exceptions
                    ┌──────────────────────────────┐
 CONSENT & DEFAULTS │  NEWTON-REX / DEBBONAIRE:    │  CREATIVE LABOUR
 Calabresi–Melamed ►│  license, disclose, ask the  │◄ Superstars · strikes
 Sunstein · Coase   │  public — don't back theft   │  Luddites · data dignity
                    └──────────────────────────────┘
                      ▲            ▲            ▲
   TRANSPARENCY &      │   POLITICAL ECONOMY      │  AUTHENTICITY &
   ENFORCEMENT         │   Stigler · public choice│  LABELLING
   Brandeis · EU Art 53│   revolving door · hype  │  Benjamin · Akerlof
   audits · provenance │                          │  moral rights
                       └── DEMOCRATIC GOVERNANCE OF TECH ──┘
                           Collingridge · Jasanoff · deliberation

2.1 Copyright theory — why creators have rights at all

The problem it addresses. Debbonaire says rights exist “from the minute it exists” and goes back to the Statute of Anne [00:04]. Why should a creator hold a right at all, and what does the right cover?

The chain of thinkers and laws

  1. Statute of Anne (1710). (From memory; not checked.) The first modern copyright law. It moved the right from the printers’ guild to authors and set a fixed term (14 years, renewable once). Its title spoke of “the encouragement of learned men to compose and write useful books”. This is the utilitarian note: rights as an incentive.
  2. John Locke, Second Treatise (1689). (From memory.) Labour theory of property: you own what you mix your labour with, provided “enough and as good” is left for others. Creator-side advocates use this: the work is the creator’s labour. AI companies’ defenders use the proviso: has enough been left in the commons?
  3. Utilitarian and economic theory. Copyright as a trade: a limited monopoly in return for more works (the US Constitution’s clause speaks of “promote the Progress of Science”). Under this view the aim is the public’s benefit, and a rule is justified only if it increases creation or access. This is the ground on which AI developers argue for exceptions: training produces public goods.
  4. Personality theory (Kant, Hegel). A work is an extension of its maker, so it is entitled to protection of integrity and attribution. This is the root of moral rights, recognised in the Berne Convention (Art. 6bis, from the 1928 Rome revision; from memory; not checked), and in UK law (CDPA ss77–89; from memory; not checked). Debbonaire’s call to be recognized as by you echoes this [06:10].
  5. Berne Convention (1886). Copyright arises without registration. This is the legal base of her statement that rights exist from creation [10:16]. Because the UK is a Berne state, a change to “protect only if you register or reserve” would run against the Convention’s “no formalities” rule, an argument creator groups made in 2025 (from memory; not checked). The opt-out proposal was not formally a formality, but critics said it acted like one.
  6. Three-step test (Berne Art. 9(2); TRIPS Art. 13). Exceptions must be (1) special cases, (2) not conflict with normal exploitation, (3) not unreasonably prejudice the author’s legitimate interests (from memory; not checked). Creator groups argue that commercial training on works fails steps 2 and 3, because it competes with the works used.

The thread. Whichever theory you choose gives a different answer to the same question. Labour and personality theory favour a ban unless licensed. Utilitarian theory favours whatever grows total creation. Newton-Rex mixes the first two with a practical point: AI “competes with the people whose work it’s trained on” [19:54], which is a utilitarian argument (the system harms its own supply) as much as a moral one.

2.2 The law of training — is copying to train a copy?

The problem it addresses. Newton-Rex says training “makes copies of millions and millions of works” [04:56]. Courts and scholars split on whether that copying is infringement, fair use, or neither.

  1. The “non-expressive use” argument (Matthew Sag). (From memory; not checked.) Copying that extracts statistics about a corpus, and does not communicate the author’s expression to a reader, should not count as infringement. This is the best-known scholarly defence of training. Sag is also the author behind the sister talk title sag-global-copyright-copy-reliant-ai.
  2. “Fair Learning” (Mark Lemley and Bryan Casey, 2021). (From memory; not checked.) They argued that machine learning should be treated like human learning in the law: learning from works is not infringement so long as the system does not copy the expression. This is the intellectual home of the “inspired by artists” line that Debbonaire objects to [03:14].
  3. The reply. Several scholars and the US Copyright Office’s 2025 report on generative AI training (pre-publication, May 2025; from memory; not checked) pointed out that fair use asks about market effect. If a model produces substitutes that compete with the works it learned from, the fourth fair-use factor weighs against it. Newton-Rex’s evidence of AI-generated books crowding out human ones [19:08] is exactly this market-substitution argument.
  4. US cases, 2025. In Bartz v. Anthropic the judge found training on lawfully acquired books to be fair use, “exceedingly transformative” in his words, but held that building a central library from pirated copies was a separate matter, leading to the class settlement. In Kadrey v. Meta the judge also found for the defendant on the record, but wrote that “market dilution” could defeat fair use in future cases. (Both from memory; the settlement approval is supported by Pearl Cohen.) Note what this means for Newton-Rex’s point. His strongest factual claim is about acquisition (pirate libraries), which the courts treat as the weak point of the AI defence, not about training in the abstract.
  5. UK: Getty v. Stability AI (Nov 2025). The High Court, per secondary sources, held the model did not store copies of the training images, so the secondary-infringement claim failed. It left open the question of training done abroad. This explains the loophole Newton-Rex describes at [18:13]: UK law applies to acts in the UK, so training done in the US and then “imported” can escape it.
  6. TDM exceptions and the EU. The EU’s Article 4 exception with an opt-out gave Europe a model that the 2024 UK proposal copied. The EU AI Act (Article 53, from memory; not checked) also requires providers of general-purpose models to publish a summary of training content and have a copyright policy. This is the closest existing law to the transparency Newton-Rex asks for.

How it feeds the talk. The speakers rely on the intuition that copying at scale for a commercial product is not like a student reading. Law scholars who disagree say the legal question is narrower than the moral one. Newton-Rex also notes: “when you train an AI model, you are making copies” [04:56]. The courts’ split decisions mean the speakers cannot assume the law is plain. Debbonaire does: she says the law “is actually not unclear” (paraphrasing [09:44]).

2.3 Consent, defaults and compensation — opt-in, opt-out, and licensing

The problem it addresses. The December 2024 fight was about who must act. The default rule decides who wins in practice.

  1. Default-rule theory. (Sunstein and Thaler; Johnson and Goldstein on organ donation, 2003; from memory.) People rarely change defaults. Countries with opt-out organ donation have far higher donor rates than opt-in countries. The same applies to rights reservation: most small creators will not opt out in time. This is why Debbonaire sees an opt-out as a transfer of value from many creators to a few firms [10:16]. A fair counter-point is that opt-out donation is an intended use of inertia, so the argument cuts both ways: it tells you the default is powerful, not which default is right.
  2. Property rules and liability rules (Calabresi and Melamed, 1972). (From memory.) An entitlement can be protected by a property rule (you need the owner’s consent, at a price the owner sets) or a liability rule (you may take it and pay a court-set or statutory price). “Opt-in” is a property rule. A compulsory or statutory licence is a liability rule. A third path is a collective licence. Creator groups often prefer property rules because they keep the bargaining power. Some economists argue transaction costs (millions of works, millions of owners) make a liability rule or collective scheme more workable. The “licensing systems” in the talk (PLS, DACS, PRS, the CCE) are a middle route [15:42], [34:14].
  3. Collecting societies. (From memory; not checked.) Music’s PRS and the US’s ASCAP both date from around 1914, formed because no composer could chase every café that played their songs. Collective licensing solves the same transaction-cost problem for AI training, and also hands the collecting body bargaining weight. The risks are well known: opaque distribution, small creators paid little, and societies that favour big catalogues.
  4. Extended collective licensing. Nordic countries let a society license on behalf of non-members too, with an opt-out. It is relevant because it combines scale with an opt-out, and so it is the kind of design a government might propose to bridge the two camps. The talk does not mention it (from memory; not checked).
  5. Machine-readable reservations. robots.txt, TDM-reservation protocols and “Do Not Train” registries are the technical form of opt-out. Their weakness: they bind only crawlers that obey them, they cover only sites the creator controls, and they cannot reach a work already copied elsewhere. This is why the debate over opt-out is also a debate about enforcement.
  6. Real-world licensing markets. Newton-Rex says “there are a handful of AI companies that want to license all of the training data they use that number is unbelievably small” [15:42]. Deals between AI developers and large news or image publishers exist (from memory; not checked), but the prices and the share that reaches individual creators are rarely disclosed.

2.4 Transparency, enforcement and provenance

The problem it addresses. Newton-Rex, asked how to enforce anything, answers “transparency requirements” [37:04], because this infringement “happens behind closed doors”.

  1. Disclosure as regulation. Brandeis’s line, “sunlight is said to be the best of disinfectants” (1913; from memory), is the standard shorthand. Disclosure laws (food labels, financial prospectuses, safety data) work when a regulator or a rival can use the information. They fail when the information is too technical for anyone to act on. Newton-Rex’s design answers the second risk: ask for the list of sources only, not the weights or the recipe [38:21].
  2. Information asymmetry. A rights holder cannot know whether their work is in a dataset. Economics treats this as a market failure: one side holds all the knowledge. A disclosure rule corrects the asymmetry.
  3. Existing disclosure rules. The EU AI Act asks general-purpose model providers for a public summary of training content (from memory; not checked). California passed a training-data transparency law (AB 2013, 2024; from memory; not checked). In the UK, Kidron’s amendments to the Data Bill proposed duties of this kind; they were removed in the Commons in 2025 (as reported, see the search results around the Lords votes in January–May 2025; from memory for the final outcome).
  4. Litigation discovery as the only current transparency. Newton-Rex says “no one admits this stuff until they’re forced to in court” [02:12]. This is a known pattern: Authors Guild and Kadrey filings exposed the use of Books3 and LibGen (from memory; not checked).
  5. Audits and certification. Third-party certification (like Fairtrade labels) turns a promise into something checkable. Fairly Trained certifies against a list of sources [38:21]. The risk in any such scheme is “label inflation” and the limited capacity of the certifier.
  6. Provenance tools. C2PA “content credentials” attach signed metadata to media about how it was made. They address output labelling (see 2.7) and not input disclosure, so they are a different tool for a different part of the problem.
  7. The “secret sauce” objection. Companies say training data is a trade secret. Newton-Rex says this has “no credibility” [37:04] for a source list (his words here are a paraphrase of [37:04]–[38:21]). A reasonable middle position: disclose to an auditor under confidentiality, as banks do with regulators. He does not discuss that option.
  8. The exposure incentive. He also gives the reason he thinks companies resist: the “hundred lawsuits … a thousand the next week” point [38:21]. This is the deterrence view of transparency: it works because it makes liability visible.

2.5 Political economy — who advises, who benefits, who is heard

The problem it addresses. Both speakers ask who changed the policy between 2023 and December 2024 [12:22], and Newton-Rex names advisers with a stake [24:37].

  1. Regulatory capture (Stigler, 1971). (From memory.) Regulated industries tend to come to control their regulators, because they have concentrated interests and the public does not. The creative sector is many small producers, and the AI sector is a few large firms, which in public-choice terms favours the second.
  2. Public choice and concentrated interests (Olson, 1965). (From memory.) Small groups with large stakes organise easily; large groups with diffuse stakes do not. Newton-Rex’s comment that organising a thousand musicians was “unbelievably easy” [11:48] is evidence that creators had an unusually sharp shared grievance, a rare case where the large group did organise.
  3. The revolving door and advisory structures. The UK AI Opportunities Action Plan (January 2025, led by Matt Clifford; from memory; not checked) and the Sovereign AI Fund are the instruments Newton-Rex mentions [25:25]. His complaint is about conflict of interest in advice, not bad faith: “I’m sure they have the country’s best interests at heart”.
  4. The “race” frame. Newton-Rex says deregulation in the US is driven by “the apparent AI race with China” [23:13]. National-security framing is a known way to reduce scrutiny of an industry’s practices. The counter-case is also serious: if a rival state does not respect copyright, a restrictive rule might simply move AI development elsewhere. That is the very worry in the talk’s other direction (“they go and do it in some other country” [18:13]). It is not settled which risk is larger.
  5. The growth narrative. Debbonaire notes that “pretty much all political parties” assume AI means growth, and asks “economic growth at what cost?” [22:22]–[23:13]. The government’s own March 2026 assessment found the evidence “limited and uncertain” (gov.uk), which supports her point that the growth case rests on assumptions. The News Media Association made the same attack in March 2026 (NMA).
  6. Solutionist persuasion. Debbonaire’s story of the Google presentation, “I can help you cure all cancer” [27:08], fits what Evgeny Morozov called solutionism (2013; from memory). Offer a politician a big good, and regulation looks like a cost. She is clear that lobbying is legitimate: “all companies should be allowed to lobby” is her sense at [28:09] (paraphrase). Her demand is for balance in who is met, and the ministerial meeting lists are public.
  7. Timing and electoral cost. Newton-Rex says a change to copyright “would cause uproar like it frankly probably make them lose the next election” [30:49] (paraphrase). The 2025 Data Bill outcome fits this: the government delayed, did not decide, and in March 2026 withdrew its preferred option.

2.6 Creative labour — the “canary” argument

The problem it addresses. Debbonaire says arts and culture are “several canaries in possibly several coal mines” [06:53], and Newton-Rex says generative AI competes with the people it learned from [19:54]. What does the economics of creative work say?

  1. Superstar economics (Sherwin Rosen, 1981). (From memory.) In markets where a top performer can serve everyone at low cost (records, books), income concentrates at the top and falls steeply below. Generative AI pushes this further at the bottom: the mid-career illustrator or session musician is replaced first. This fits the survey in the talk that illustrators were losing work [19:54].
  2. Cultural-industries scholarship (David Hesmondhalgh and others). Creative work is precarious: project-based, freelance, low-paid for most. Debbonaire says as much: artists’ lives “are under huge pressure anyway” [06:10]. AI arrives as an added pressure on an already thin base.
  3. Automation and the 2023 strikes. The Writers Guild and SAG-AFTRA strikes of 2023 put AI terms (consent, compensation, limits on digital replicas) into contracts (from memory; not checked). They are the labour-law version of the policy demand here. The sister talk sag-global-copyright-copy-reliant-ai likely covers this ground; I did not read it.
  4. The Luddite reading. Brian Merchant’s Blood in the Machine (2023; from memory) retells the Luddites as skilled workers who objected to how machines were used to cut wages, not to machines as such. Both speakers fit that pattern: they accept the technology and fight about terms.
  5. “Data dignity” (Jaron Lanier). (From memory.) The idea that people whose data trains a model should be recognised and paid, the data equivalent of royalties. It is a close cousin of Debbonaire’s “recognition and reward” [06:10], with the difference that Lanier proposes it for everyone, not only professional creators.
  6. Extraction (Kate Crawford, Atlas of AI, 2021; Shoshana Zuboff). (From memory.) AI as an extractive industry: data, labour, energy. The talk’s phrase “born of piracy” is a sharper version of this.
  7. A caution. Evidence that AI has cut creative incomes at scale is still thin and survey-based. One survey in a talk is an anecdote. The Amazon example [19:08] is a plausible mechanism, but I have not seen sales data for those authors.

2.7 Authenticity, labelling and moral rights

The problem it addresses. Debbonaire does not only want payment. She wants to know: “I still want to know where I’m reading Czechov, not a novel in the style of Czechov” [22:22].

  1. Benjamin’s aura (1935). (From memory.) The “aura” is the sense of an original’s presence and history. For Benjamin it is lost in mass reproduction. Debbonaire’s wish to buy “a book by a human” is a market form of that wish for provenance.
  2. Akerlof’s lemons (1970). (From memory.) When buyers cannot tell quality, bad goods drive out good ones. If readers cannot tell human from machine writing, labelled human work loses its price premium. A labelling rule restores the signal. The government’s labelling taskforce, due to report in autumn 2026 (secondary sources, see Part 1), is aimed at this.
  3. Moral rights. Attribution (to be named) and integrity (not to have your work distorted) exist in UK law (CDPA ss77–89; from memory; not checked). Debbonaire’s call for work to be “recognized as by you” connects to attribution. AI outputs complicate it: a style is not attributable.
  4. Style and imitation. Copyright does not protect a style. This is why “in the style of” outputs are legal in many cases even when the training was disputed. The speakers do not address this gap directly; the sister talks may.
  5. Her personal practice. Debbonaire’s tattoo story, in which she asked the poet’s permission before inking his poem [29:43], is a small enactment of consent culture. Whether most people would do the same, she says, she doubts.

2.8 Democratic governance of technology

The problem it addresses. “When have we the public been asked to get involved in this?” [20:41].

  1. The Collingridge dilemma (1980). (From memory.) Early in a technology’s life it is easy to shape and its effects are unknown; later the effects are known but it is hard to change. Debbonaire’s worry that governments say “we don’t know what to do, so let’s not talk about it” [22:22] is the first horn; the opposite, rushing to adopt, is the second.
  2. Technology assessment and public deliberation. Citizens’ assemblies and consensus conferences (Denmark’s model; Sheila Jasanoff’s work on “technologies of humility”; from memory) are tools to involve the public before lock-in. The UKAI approach she praises as “constructive critical friend” [20:41] is an industry-side version.
  3. Technology as religion. She says early pioneers “are like a religion” [20:41]. Scholars of the “California ideology” and of AI futures (e.g., Timnit Gebru and Émile Torres’s critique of the “TESCREAL” bundle, 2023; from memory) make similar arguments. This is a charged claim; treat as one view.
  4. Hidden uses. She says she does not think people know how much AI is already involved in decisions, including weapons and democracy [29:43]. She gives no evidence in the talk. The wider record on military AI is real but varied; I did not check specific claims.

2.9 How the eight meet

Copyright theory (2.1) says why a right exists. The law of training (2.2) says that the right may not reach training, so the speakers move to politics. Consent theory (2.3) shows why a default rule matters more than a principle. Transparency (2.4) is the practical lever that makes any rule enforceable. Political economy (2.5) explains why that lever is not being pulled. Labour economics (2.6) says who loses. Authenticity (2.7) says what consumers lose. Democratic theory (2.8) asks who should decide.

Part 3 — Deep dives

3.1 “Born of piracy” — the acquisition argument

🎙 In the talk

  • Newton-Rex: “the AI industry today as we know it is born of piracy” [00:04]. He walks through GPT-1 (7,000 books taken from an ebook platform by another group), GPT-2 (8 million scraped documents, including news), and GPT-3, which he says added Library Genesis, “a vast pirate library” [01:37].
  • On how we know: “no one admits this stuff until they’re forced to in court” [02:12].
  • He says an Anthropic co-founder, Ben Mann, “had been one of the people who was downloading these pirate libraries at OpenAI”, admitted this in court, and “one of the first things he did at Anthropic was download them again” [02:12].
  • On the effect: “they raised almost a billion dollars in financing” while training on such models (his estimate) [02:12].
  • On what pirated authors say: the GPT-1 authors “had no idea” and AI is “directly competing with them” [19:08].

📚 Deeper theory (not the speakers’ words)

  • This is the strongest ground in the debate because it avoids the unsettled question in 2.2. Courts in the US treated acquiring pirated copies as separate from training on them, and the Anthropic class settlement put a price on it (Pearl Cohen). If the model of the future is “buy or license the books, then train”, the acquisition problem is solved and the argument moves to the harder question of whether training itself needs permission.
  • Note the difference between infringement (a legal finding) and piracy (a moral and rhetorical label). Newton-Rex uses the second. In places he is making claims about particular datasets that rest on court filings, which is where he is on firmest ground.
  • The GPT-1 detail matches the “BookCorpus” dataset of about 7,000 unpublished books (from memory; not checked). The GPT-2 detail matches “WebText” (from memory; not checked).
  • Link: 2.2, 2.4.

Where others push back. AI developers argue that the legal question is training, not the library, and that a model does not contain or reproduce the books (the Getty reasoning, 2.2). Some also point out that datasets such as Common Crawl were open to everyone, and that scraping public web pages is a long-standing norm. Newton-Rex’s reply is that availability is not permission, and that the same companies did their scraping when nobody was looking.

What they leave out. Open-source and academic datasets built from public sources, which are used by small teams who cannot afford licences. A strict licensing rule would hit them hardest. The talk also does not mention that some companies have since bought books and settled.

Check yourself.

  1. Why is the “acquisition” claim legally easier to win than the “training” claim?
  2. What is the difference between saying a dataset was pirated and saying a model infringes?
  3. Which of Newton-Rex’s dataset claims rest on court records, and which on his own inference (e.g. “GPT-3 added LibGen”)?

3.2 “Inspired by artists” — human learning vs machine learning

🎙 In the talk

  • Debbonaire recounts being told, about three years before, that “artists have been inspired by other artists for centuries. This is just the same” [03:14]. Her answer: “This is not the same” [04:21].
  • Her point: when artists copy each other there are laws and “they do sue” [04:21].
  • Newton-Rex: it is “a very unfortunate accident of history that we call this machine learning” [04:56]. His reasons: “you are making copies of millions and millions of works. This does not happen in human learning.” And the result is a system that can “displace the labor, the work of huge numbers of people in a way that an individual can’t” [04:56].

📚 Deeper theory

  • Their argument has three distinct parts that are worth separating: (1) copying (the machine makes copies, the human does not store verbatim), (2) scale (one system versus one person), and (3) substitution (the output competes with the inputs). Law treats (1) as technical, and (3) as central to fair use (2.2).
  • The Lemley–Casey “fair learning” view disagrees with (1): a human also stores traces, and a model does not usually store whole works. A research literature on memorisation shows that large models can sometimes reproduce text or images verbatim; how often is disputed (from memory; not checked).
  • On the metaphor: “learning” in “machine learning” is a technical term for fitting parameters. Philosophers of mind such as Dreyfus and Searle argued long ago that the word hides differences; Newton-Rex’s complaint about the name is a version of that (from memory).
  • Link: 2.2.

Where others push back. Defenders say that analogies to human learning are not meant as identity but as a principle: reading and analysing are not infringement, and the law has never required permission to learn from a book. They also say copyright has never protected ideas or style.

What they leave out. The strongest middle position: even if training is not infringement, outputs that reproduce a work are, and the market harm can be remedied without banning training. The talk does not separate training, output and style.

Check yourself.

  1. Which of the three parts (copying, scale, substitution) does the law weigh most in fair use?
  2. Why might a rule about output leave Debbonaire’s complaint unsolved?
  3. What would count as evidence that a model “stores” a work?

3.3 The UK policy story — opt-out, backlash, and the vacuum

🎙 In the talk

  • Debbonaire’s three principles from the 2024 Labour position: intrinsic rights from creation, “the right to opt in, not opt out”, and recognition and reward [06:10].
  • Newton-Rex on Chris Bryant: in 2023 he publicly warned Rishi Sunak about threats to creators’ copyright, then authored the opposite approach [07:25]–[09:44]. (The transcript has Newton-Rex making the “screwing the arts” jibe; the captions do not label the speaker.)
  • The December 2024 document had a “government preferred option” [08:36]: Debbonaire tweeted that the document’s statement that the law is unclear was wrong (“The law … is actually not unclear”, her gist at [09:44]). The proposal as Newton-Rex puts it: “the entirety of the life’s work of this country’s creatives should be handed to AI companies for free” unless the creator opts out [10:16]. It “basically entirely flips the concept of copyright” [10:16].
  • Backlash: protests from musicians and authors; a silent album with “a thousand musicians”; “how unbelievably easy it was to find a thousand musicians who wanted to protest this” [11:48].
  • The shift: “that change from a position of protecting creatives to one that frankly was just repeating talking points from big tech” [11:48].
  • Debbonaire on process: creators “did not feel that they were getting the hearing that the tech bros were getting” [12:22], and having checked ministerial meeting lists she thinks that was fair. Her view of the cause: “the tech industry had said to number 10 we don’t want to be regulated” [13:06].
  • The result: a Lords amendment (Kidron) went “backwards and forwards” between the Houses, and the government “was going to dig in its heels” [14:09]. Then: “Come the king’s speech two weeks ago AI bill is there none” [14:09]. And: “we’re left with a vacuum” [15:42].

📚 Deeper theory (with the 2026 record)

  • What actually happened next, per the sources in Part 1: the government’s own March 2026 report dropped the opt-out as its preferred option, and a minister is reported saying the first approach was wrong. So the speakers’ complaint in June 2026 is partly about now (no bill, no firm policy) rather than the 2024 proposal, which had by then been withdrawn. Their talk handles both, but a listener from the video alone could think the opt-out is still the plan.
  • The lesson from public-choice theory (2.5): the consultation showed what happens when a policy imposes visible costs on a large and vocal group.
  • The institutional detail: the Data (Use and Access) Act 2025 required the government to publish a report and impact assessment on copyright and AI, which is how March 2026 came about (Pinsent Masons timeline). The Lords amendments for transparency were removed in the Commons (as reported in 2025 coverage found in my search; the Commons ground was financial privilege, from memory; not checked).
  • Link: 2.1, 2.3, 2.5.

Where others push back. Ministers’ position, as reported in 2025, was that the existing law lacks clarity and harms both AI developers and rights holders, and that a single package of rights reservation plus transparency would give certainty to both. Tech-sector groups argue that a licensing-only rule will push AI development to jurisdictions with a TDM exception. Some legal commentators say that “the law is not unclear” is too strong: the UK has never ruled on commercial training, and Getty left the territoriality question open.

What they leave out. Any account of what the civil servants and the ministers in the department say they were weighing, and the possible role of the 2025 US–UK tech partnership (from memory; not checked). They also give no figure for how many creators would have opted out.

Check yourself.

  1. What is the difference between “the law is unclear” and “the law is unsettled”?
  2. Why might a government prefer an opt-out even if it knows most creators will not use it?
  3. Which of the speakers’ claims about Dec 2024 have been overtaken by events by June 2026?

3.4 Licensing in practice — and who actually does it

🎙 In the talk

  • Debbonaire names working systems: the Creative Content Exchange, and a Publishers’ Licensing system for generative AI that lets artists “opt in rather than opt out” [15:42]. Later: PLS for text, DACS for visual artists, PRS for music, at different stages [34:14]–[35:32].
  • Newton-Rex’s reality check: “while there are a handful of AI companies that want to license all of the training data they use that number is unbelievably small” [15:42]. He knows because his nonprofit “certifies these companies and it certifies their training data” [15:42].
  • Named examples of firms without a licence record, in his view: OpenAI and Anthropic (started with piracy and “still train on huge amounts of copyrighted work today”), Runway (media reports of YouTube scraping), and ElevenLabs (sued by voice artists) [15:42]–[17:14].
  • On loopholes: “these companies go and do it in some other country and we sort of turn a blind eye” [18:13].
  • A slogan: “The next time any AI company calls themselves responsible, I strongly recommend saying, ‘I’m curious what you train your models on.‘” [18:13]. His one stated condition for an ecosystem: legislation and regulation “insisting that people license as it does at the moment in the UK” [17:14].

📚 Deeper theory

  • Licensing sits on the property-rule / liability-rule axis (2.3). Collective licences lower transaction costs, but only if the AI developer is under pressure to use them. Newton-Rex’s remark that the legal landscape must insist on licensing is the point: without a legal duty, a licensing market is voluntary and thin.
  • Weaknesses in a collective approach that critics raise: distribution to individual creators; coverage gaps (what about works with no society?); the risk that the licence price is set by the buyer’s bargaining power; and whether a licence for training also covers output. The talk does not discuss pricing at all.
  • Claims about firms (Runway, ElevenLabs) are the speaker’s report of media coverage and lawsuits; I did not verify them. Treat as allegations.
  • Link: 2.3, 2.5.

Where others push back. Developers say a licence for every training work is impractical because datasets hold billions of items and ownership is often unknown (“orphan works”). The answer from the licensing side is that large rights holders and societies already cover the bulk of commercial value, and a rule for the remainder can be built (e.g., extended collective licensing).

What they leave out. Open-weight models and small developers; a number for what training licences should cost; and whether individual creators would get paid in practice.

Check yourself.

  1. Why does a voluntary licensing market stay thin without a legal duty?
  2. What is the “orphan works” problem and why does it matter for a licence-everything rule?
  3. If a company licenses training data but trains abroad, what does UK law say about it, per Newton-Rex?

3.5 Competition, the “canary”, and why the stakes are wider than art

🎙 In the talk

  • “Generative AI competes with the people whose work it’s trained on” [19:54]. Evidence given: a “huge influx of AI generated books on Amazon, which is making their books impossible to find” for the GPT-1 authors [19:08], and “one in four illustrators said they had lost work to generative AI already” [19:54].
  • Debbonaire widens it: “we’re also the canary in the coal mine” for everyone’s questions about jobs [19:54]–[20:41]. She says: “when have we the public been asked to get involved in this” [20:41].
  • On tone: some early AI pioneers talk “like a religion” [20:41], but she wants to be a “constructive critical friend” [20:41].
  • On human purpose: “what is the role of humanity?” [22:22].
  • She worries about unseen uses: “how much AI is involved in our democracy” [29:43].

📚 Deeper theory

  • The competition argument is the heart of the market-substitution test (2.2). It also answers a standard defence: “the model doesn’t copy your work, it learns patterns.” If the patterns are used to out-produce you, the harm is the same.
  • The canary argument belongs to creative-labour economics (2.6). It is a prediction: what happens to illustrators will happen to other knowledge workers. That may be right, but it is a forecast, not evidence in the talk.
  • The Amazon example has a known context: in 2023 Amazon limited the number of books a self-publisher could upload per day after floods of AI-generated titles (from memory; not checked).
  • Link: 2.2, 2.6, 2.8.

Where others push back. Economists note that new tools often raise demand for human work in complementary roles (taste, curation, performance) and that early job-loss surveys overstate effects. Others say the substitution is real but is mostly of low-end commercial work (stock imagery, copywriting), while high-end art is unaffected.

What they leave out. The upside side of the ledger for creators who use these tools; the fact that some artists have licensed their work to AI firms by choice; and any number on net employment in the sector.

Check yourself.

  1. Why does the substitution evidence matter more to a fair-use analysis than the “piracy” label?
  2. What would convince you the “canary” prediction is wrong?
  3. Name one creative role AI might complement rather than replace, and why.

3.6 Who is advising government — and the Sovereign AI Fund

🎙 In the talk

  • Newton-Rex on the cause: UK government sold the idea that AI is “the route to growth” [24:37]. “the main problem in the UK is who is advising the government” [24:37]. He names Hassabis (and notes Google is “being sued in the US for alleged copyright infringement” [25:25]), VCs who run “the sovereign AI fund” and write “the AI action plan” [25:25]. His rule: “you cannot solely base your policy on the approaches of VCs, on the approaches of AI execs” [25:25], and calls it “just like basic democracy 101” [25:25].
  • Debbonaire: “I do not think that the heads of publishing houses or writers or musicians or composers have quite the same level of access” [27:08], then the Google away-day lunch with “I can help you cure all cancer” [27:08]. She calls the pitch “a cult technique” [28:09] (her words about how it works on politicians) and adds that companies should be allowed to lobby, but “you can judge for yourselves who were the ministers meeting” [28:09] (“Have a look. It’s out there.” [28:09]).
  • The Fund: “500 million quid” [30:49]. Asked repeatedly “what is your policy on backing companies that train on copyrighted work without a license?” [31:42], the government gives “a non-response”: it will back firms that adhere to UK law [31:42]. His objection: “UK law doesn’t say anything about what happens if you’re a company and you go train your models in the US and then you bring them back here” [31:42].
  • His inference: they are “not only open, but probably actively keen to invest in companies who do go and train on copyrighted work abroad” [32:43]. His ask: “just don’t back them” [32:43] and “Let’s incentivize responsible AI development” [33:18]. And: “If they truly believe that creatives rights are important … they have to live that out” [33:18].

📚 Deeper theory

  • The Fund question is a clean test of procurement as policy: even when a state cannot regulate foreign conduct, it can set conditions on its own money. This is how governments use grants and contracts to set standards (from memory).
  • The official record in Part 1 supports his description: written answers repeated “applicable UK law”. His inference about motive (“actively keen”) is his reading, not a documented fact. A plainer reading is that the Fund wants no hard condition because it would shrink its pool of deals.
  • The wider argument is about conflicts of interest in advice (2.5). Newton-Rex says plainly that the advisers may be sincere. This is the distinction between corruption (bad faith) and capture-by-perspective (good faith, narrow view).
  • Debbonaire’s evidence is a mix of her own experience and an invitation to check public meeting lists. I did not check the lists.
  • Link: 2.5.

Where others push back. Government replies that supported firms must comply with law and that conditions on an investment fund are a matter for its commercial mandate. Industry replies that an investor cannot audit the training history of every foreign model. And an adviser with industry experience is also a source of expertise the civil service lacks.

What they leave out. Any named example of a Fund investment that trained on unlicensed work; and what a workable condition would be (disclosure to the Fund? a warranty in the investment contract?).

Check yourself.

  1. Why is “adhere to UK law” a weak condition for a firm that trains abroad?
  2. What is the difference between a conflict of interest and an honest but narrow view?
  3. Write one sentence of a Fund condition that Newton-Rex would accept and a firm could actually meet.

3.7 Enforcement — transparency, audits, and the “secret sauce”

🎙 In the talk

  • Questioner 2 asks how any rule could be enforced when taking is hidden [37:04].
  • Newton-Rex: “transparency requirements” [37:04]. The reason: this is “one of the first major types of copyright infringement that happens behind closed doors” [37:04], unlike DVDs sold on the street.
  • He says the government “has not accepted multiple demands for transparency requirements” [37:04].
  • On the objection that training sources are a secret: “In my mind there is no credibility to this” [37:04]. What is needed is “a list of sources of training data” [38:21], which is what his nonprofit asks for in certification. “You implement audits” [38:21].
  • His explanation of resistance: reveal sources and “there’ll be a thousand the next week” in lawsuits, “in my mind rightly” [38:21]. “It’s not like using library genesis was this moment of innovation. No, it’s just theft … and it’s theft that everyone is doing” [38:21]. His conclusion: any refusal “is an absolute failure of of the government’s responsibility” [38:21].
  • Debbonaire’s closing analogy: builders who steal bricks “might do it in the middle of the night so you didn’t know but eventually you have to arrest them” [39:52].

📚 Deeper theory

  • See 2.4 for the full frame. The key design choice is what to disclose: the narrowest useful item is a list of source domains and datasets, not individual files. The EU AI Act’s “sufficiently detailed summary” model sits in this space (from memory; not checked), and its template was the subject of much argument over how detailed is detailed enough.
  • A real tension: a list of sources lets a rights holder suspect infringement but may not prove their specific work was copied. A stronger scheme would give rights holders a query right (“was my work used?”). Newton-Rex does not say which he wants.
  • Compare the brick analogy: it implies detection and arrest, i.e. a penalty regime. A disclosure rule alone creates information, but someone must bring claims. In the UK, that is the rights holder, which brings the cost problem back for small creators (collective actions or a regulator would help).
  • The lawsuit-wave claim is a prediction. It also reads as an admission that companies know the sources include unlicensed works. He offers no proof for it beyond inference.
  • Link: 2.4.

Where others push back. AI developers say disclosure invites strategic lawsuits and reveals competitive information; that sources lists are meaningless for web-scale scrapes; and that the EU template shows the compliance burden. Privacy lawyers add that datasets include personal data, so disclosure has to be managed.

What they leave out. Penalties, who enforces, whether disclosure should be public or to a regulator, and how to treat models trained before the rule starts.

Check yourself.

  1. What is the minimum disclosure that would let a creator test whether their work was used?
  2. Why does disclosure raise the odds of lawsuits, and is that a reason against it?
  3. Who would bring the claims in the UK, and what would make that realistic for a freelance illustrator?

3.8 What creators can do — Q1, sector unity and licensing now

🎙 In the talk

  • Questioner 1 asks what the creative industries can do, given that unity a year ago has become harder [34:14].
  • Debbonaire: “two things that have to happen in parallel” [34:14]: join the licensing systems that exist, “work with them if you’re in that sector” [34:14], and settle positions inside each sector (she cites film and TV needing to “battle it out” in an internal forum, as publishing had to) [35:32].
  • The second: “do not stop keeping up the pressure” [35:32], using licensing as a “name and shame device” [35:32], and asking MPs “Where is the regulation we were promised?” [35:32].

📚 Deeper theory

  • This is a standard collective-action answer (Olson, 2.5): small groups need selective incentives and institutions. The collecting societies are those institutions.
  • A licensing scheme has a double role she names: a payment route and a public record of who does and does not pay. That role echoes 2.4 (disclosure by other means).
  • The cost: if a sector licenses on terms that suit insiders, it may lock out smaller creators and may weaken the case for a statutory rule (“the market is handling it”). This tension between voluntary schemes and law is not discussed in the talk.
  • Link: 2.3, 2.5.

Where others push back. Some creators fear that joining a licensing scheme legitimises training. Others want a ban, not a market.

What they leave out. Practical steps for the individual freelancer, and how smaller or international creators fit.

Check yourself.

  1. Why might a voluntary licence undercut the demand for a statutory duty?
  2. What does the “name and shame” function rely on?
  3. Who is left out of sector-by-sector licensing?

Putting it together

Mental model. A system was built by copying a great deal of human work, often without asking, and it now competes with the people it copied. The speakers’ fix is three levers: consent (rights arise at creation; get permission, and pay), disclosure (say what you trained on, so claims can be checked), and integrity of government (do not let the advisers, or the state’s own money, favour firms that skip the first two). Policy has so far produced a vacuum: the 2024 opt-out was withdrawn in March 2026, but no replacement has been enacted, and in the speakers’ account no bill and no fund condition have followed. The honest weak point of their position: the law on training itself is unsettled, and the strongest parts of their case (pirated acquisition, market substitution, transparency) are exactly the parts that do not depend on resolving it.

Open tensions

  • Licensing market (voluntary, fast) vs statutory duty (slow, complete).
  • Opt-in as a principle vs transaction cost at scale (orphan works, small developers).
  • “Train abroad and import” vs the risk that a strict UK rule just moves AI work elsewhere.
  • Transparency as the key lever vs trade-secret and privacy limits on what can be shown.
  • The canary prediction (wider job losses) vs thin evidence so far.
  • Moral claim (“it’s just theft”) vs the legal reality that courts split on training and only agree on piracy.

Reading path

Graph View

Backlinks