Talks

Will AI Companies Respect Creators' Rights? — Ed Newton-Rex

54 min read

About this talk

Speaker: Ed Newton-Rex
Event: Future of Life Institute Podcast
Date: 2025-06-20 (upload date; recorded a few days after the US Copyright Office’s pre-publication AI report of 9 May 2025, so about mid-May 2025)
Length: 1:27:15
Channel: Future of Life Institute
Watch: Will AI Companies Respect Creators’ Rights? (with Ed Newton-Rex)

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

  • 🎙 In the talk. What Newton-Rex or the host says, with [MM:SS] timestamps. Words in quotation marks are exact words from the transcript. Only these parts are his views.
  • 📚 Deeper theory / outside context. Law, economics and history that explain why the argument works or fails, and what has happened since. It is not his claim unless a timestamp says so.

Facts I could not check are marked “from memory; not checked”.

Suggested route. Summary → Map → Part 1 (key terms and what has happened since June 2025) → Part 2: The intellectual map (read once, slowly) → Deep dives in order → “Putting it together”.

Speakers

LabelWhoRole
HostGus Docker (captions: “Gus Ducker”; surname from memory; not checked)Host of the Future of Life Institute podcast. Asks the sceptical questions: why are there still human musicians, can a certificate scale, is it too late, why is automation bad
Newton-RexEd Newton-RexComposer. Founded a music-generation startup (Jukedeck) [01:05], worked at TikTok’s For You feed and Snapchat, then ran audio at Stability AI. Resigned in 2023 over the fair-use stance. Founded the non-profit Fairly Trained

Caption warnings. Auto-captions garble names. Corrections used here: “Bite Dance” = ByteDance; “Juke Deck” = Jukedeck; “Markoff chains” = Markov chains; “Gus Ducker” = Gus Docker; “null Gallagher” = Noel Gallagher [01:01:05]; “Open AI” = OpenAI; “Overton window” is correct. Slips and simplifications in the talk (checked or flagged).

  • Dates for Jukedeck. He says he started the company “in 2010” [01:05] and later that it launched publicly around 2014 [16:46]. My memory is that Jukedeck was founded in 2012 and launched in 2014 (from memory; not checked). The 2010 date may mean when he began the work.
  • “Training on copyrighted work in the UK is straight up illegal” [35:58]. This is a simplification. UK law (CDPA s29A, from memory; not checked) permits text and data mining for non-commercial research, and bars it otherwise. So commercial training is not covered by an exception, but “straight up illegal” skips the research carve-out, and (see Part 1) the recent Getty ruling shows the territory of the training matters.
  • “Training is not a trade secret”. Fine as opinion. Not a factual claim to check.
  • Mechanize. He says a company called Mechanize announced about a month earlier that its mission is to “automate all work” [52:44]. The company launched in April 2025 (from memory; not checked). I did not verify the investor list he gives.
  • Spotify study. He says Spotify researchers in 2020 studied July 2019 data on about 100 million users [01:08:04]. This matches my memory of Anderson et al., WWW 2020 (from memory; not checked).

Summary

Central question. Will the people building AI respect the rights of the creators whose work trained it, and if not, what follows for creators, for culture and for the wider economy?

Main thesis. Newton-Rex says the industry’s standard position is that training on copyrighted work without a licence is fair use. He argues that this is a break with the industry’s own earlier practice, that it is a choice and not a necessity, and that it pushes out the companies that do license. He wants three things: licensing (consent and payment), transparency (lists of training data that a third party can check), and time pressure (“time is of the essence” [45:03]). He is not against AI. He calls himself “not an AI booster” [07:22] and expects AI to keep improving.

How the talk moves.

  1. AI music is already good, and the damage is in the long tail of working musicians, not the stars [04:46]–[11:58].
  2. Why he left Stability AI, and how licensing became the exception when it had been the norm [12:15]–[21:03].
  3. Open-weights models and why they weaken fair-use defences and cannot be recalled [21:03]–[26:35].
  4. Fairly Trained, the certification, and how far it can scale. Transparency, synthetic data “laundering”, audits and red teaming [26:35]–[45:49].
  5. Is it too late? Model collapse, and why he thinks the technology will keep improving [41:54]–[50:10].
  6. The wider stakes. Creatives as “canary in the coal mine” for mass labour displacement, and doubts that those who ignore copyright will redistribute wealth [50:42]–[58:53].
  7. Culture. Remix, “concrete works”, winner-take-all, recommendation systems, a coming “humanist movement”, and why efficiency is the wrong frame for art and for UK policy [58:53]–[01:24:03].
  8. Priorities: shift the Overton window; marketplaces, public-domain datasets and attribution research; no ban on AI in the creative industries [01:24:03]–end.

Stance to note. He argues as an insider, as a builder and as a creator. His strongest moral word is “theft”; his most careful distinction is between revealing training data and listing training data [37:12]. He is explicit that he does not want to stop AI: “I think that’s unrealistic” [01:26:28].

Map — talking points

#PointTime
1Background: Jukedeck (2010 start), rule-based then neural music, ByteDance, Snapchat, Stability AI audio[00:00]–[03:54]
2AI music is “pretty indistinguishable” from human pop; classical still safe[04:46]
3Why musicians survive yet are hurt: stars survive, the long tail is eroded (songwriters, jingles)[06:21]–[11:58]
4Resigning from Stability: the US Copyright Office submission saying training is fair use[12:15]–[15:09]
5The industry attitude: fair use is the default; it was not always so (Google Magenta)[15:49]–[18:39]
6The gold rush: everyone copies the first movers; investors punish licensors[18:39]–[21:03]
7Open-weights models: “commercial”, no downstream limits, irreversible[21:03]–[26:35]
8Fairly Trained: certification, 19 models, self-certification, trust-based[26:35]–[31:33]
9Transparency: the House of Lords amendments, why “trade secret” fails, UK government refuses[33:07]–[37:12]
10Revealing data vs listing data. Synthetic data as “laundering”. Audits and red teaming[37:12]–[41:54]
11Is it too late? Recall by law; model collapse is temporary[41:54]–[45:49]
12Technology keeps improving; “can’t do X” is a bad argument; pace of change[46:25]–[50:10]
13Labour: creatives are the canary; Mechanize; “automate all work”[50:42]–[55:49]
14Why not just relax? Because the builders do not respect rights now, so why would they redistribute later[56:55]–[58:53]
15Culture: harder entry to the industry; remix culture; “concrete works” persist[58:53]–[01:03:10]
16Avatars and personalised content: novelty with poor retention[01:03:21]–[01:04:56]
17Winner-take-all and recommendation systems (Spotify 2020 study); no homogenisation if the user cares[01:06:30]–[01:10:57]
18Backlash is strong; the camera/synth analogy is wrong; a “humanist movement” in the arts; the Beatles on a guitar[01:10:57]–[01:18:12]
19Competitive pressure: creativity is not about efficiency; UK politics and the AGI race[01:18:45]–[01:22:13]
20Soft power and culture; undermining copyright undermines a strong industry[01:22:13]–[01:24:03]
21Priorities: shift the Overton window; marketplaces, public-domain datasets, attribution research; do not ban AI[01:24:03]–end

Part 1 — Before you read

Key terms

  • Fair use (US). A defence in US copyright law (17 U.S.C. §107, from memory; not checked). A court weighs four factors: the purpose and character of the use (is it “transformative”?), the nature of the work, how much was taken, and the effect on the market for the original. The fourth factor matters most in this talk, because Newton-Rex’s point at [23:50] is that what the model is used for bears on whether training was fair. Fair use is decided case by case [24:43].
  • Licensed vs unlicensed training. “Licensed” means the developer asked permission and, usually, paid. His test for Fairly Trained is “AI models that aren’t trained on copyrighted work without a license” [28:08].
  • Open source vs open weights. An open-weights model publishes its trained parameters, but not its training data. He notes this at [21:03]: many “aren’t open in the traditional sense because they’re not revealing their training data”. Do not treat the two as the same thing.
  • Transparency (of training data). A duty to disclose what data a model trained on. He separates two things: sharing the data itself, and sharing a list of sources [37:12].
  • Red teaming. Probing a model to find what it knows or does. Here, prompting it to see whether it reproduces a protected work [41:54].
  • Synthetic data. Data generated by a model, used to train another. He calls it a possible “way of laundering copyright” [39:20].
  • Model collapse. The claim that training models on model-generated data degrades them over generations. He doubts it will last [43:28].
  • Long tail. The many small earners (session players, songwriters, jingle writers) beneath the stars. His economic worry sits here [08:44].
  • Overton window. The range of ideas treated as acceptable in public debate. His aim is to move it [01:24:03].
  • “Concrete works”. His term for finished recorded songs, books and films that everyone can share and talk about, as opposed to personalised content [01:02:13].

People and organisations named (a reading list)

  • Stability AI, where he was VP of Audio. Stable Audio (Sept 2023) was trained on licensed data [12:43].
  • US Copyright Office. Its call for comments (2023) and its report on AI training (pre-publication, early May 2025) frame the talk [13:26], [23:50].
  • Google Magenta (from 2015). His example that researchers once disclosed and commissioned their training data [16:46].
  • Fairly Trained. His non-profit; 19 certified models at the time [28:08].
  • House of Lords, UK government. Amendments to force disclosure of training data were proposed twice and rejected [33:07].
  • Mechanize. A start-up he says aims to “automate all work” [52:44].
  • Spotify researchers (2020). The recommendation-diversity paper [01:08:04].
  • “Is This What We Want?” The 2025 silent protest album by about a thousand UK musicians, including Kate Bush and Max Richter [01:14:04].
  • Taylor Swift, Oasis. Rhetorical examples: the stars who survive [07:22], [09:34].
  • Marc Andreessen (“software is eating the world”, paraphrased) [52:44]. Not named in the form of a thesis he holds.

Newton-Rex’s own frame

He is a rare case in this debate: a composer, a founder who built generative music systems for about 15 years, and a former executive at a company that took the opposite path. He says his views on copyright “align with a lot of the anti-AI crowd’s views” but he is not part of it [43:28]. He rests his case on consent and payment, not on opposition to the technology. That makes his critique harder to dismiss as technophobia, and it is also why his proposals are pragmatic (certify, disclose, license) and not abolitionist. (He has also written a book, The Intelligence Heist, listed by Simon & Schuster UK. I did not read it.)

The current debate: what has happened since June 2025 (checked 30 Oct 2026)

The talk is 16 months old. Several of its predictions have been tested. Sources are secondary (law-firm notes, trade press); verify against the judgments before relying on details.

US courts

  • The talk came just before the first two fair-use rulings on generative training. In late June 2025 two Northern District of California judges found training on books to be fair use on the records before them: Bartz v. Anthropic (lawfully acquired books; piracy treated separately) and Kadrey v. Meta (a narrower win for Meta, with the judge saying “market dilution” could often win the fourth factor for plaintiffs). The two judges disagreed on dilution (Authors Alliance on Kadrey; summary of both). In June 2026 the Kadrey plaintiffs asked to take an interlocutory appeal; I did not find the outcome (ChatGPT Is Eating the World).
  • Anthropic paid $1.5 billion. The Bartz class settlement, about $3,000 for each of roughly 482,000 works, received final approval in July 2026, with a destruction order for the torrented files. The release covers conduct up to 25 Aug 2025 and not AI outputs or future conduct (Jurist; AAP). Note the pattern: the court that said training could be fair use still made the company pay for how it got the books.
  • First appellate word. On 29 Sept 2026 the Third Circuit affirmed that Ross Intelligence’s use of Westlaw headnotes to train a non-generative legal-research tool was not fair use, calling the use “minimally transformative, at best”, and distinguished generative AI (IPWatchdog; Johnson & Wales summary). Reports differ on whether the reasoning is still sealed. It does not settle the generative question.
  • NYT v. OpenAI/Microsoft is still pre-trial. The news plaintiffs moved for summary judgment on fair use on 4 Sept 2026, and no trial date is set. The briefing schedule is reported inconsistently (summary of docket reports).
  • US Copyright Office. The pre-publication report came out 9 May 2025, and the Register of Copyrights, Shira Perlmutter, was fired the next day. The D.C. Circuit restored her in Sept 2025 and the Supreme Court declined in June 2026 to let her be removed while her suit continues (Washington Examiner; Bloomberg Law). I did not find whether the report was ever finalised.

Music, the talk’s home ground

  • Suit-to-licence conversions. Universal settled with Udio on 29 Oct 2025 and Warner settled with Udio (Nov 2025) and Suno (25 Nov 2025), each with a licensed-platform deal. Sony had not signed with Udio or Suno as of those reports, and I did not find whether Universal or Sony later settled with Suno (Music Business Worldwide; The Star). This supports his claim that licensing is possible. It also differs from his model: these firms trained first and licensed later, under litigation pressure.
  • GEMA v. OpenAI (Munich, 11 Nov 2025). The court found that OpenAI’s models had memorised nine songs’ lyrics and reproduced them, and said memorisation goes beyond what the text-and-data-mining exception covers. OpenAI said it would appeal (Bird & Bird).

UK, the talk’s political target

  • Getty v. Stability AI (4 Nov 2025). The High Court rejected Getty’s secondary-infringement claim: model weights were “articles”, but not “infringing copies”, because they did not store Getty’s works. Getty had dropped its primary claims during trial. Permission to appeal was granted (reports differ on which party sought it) (Bird & Bird; AIPPI). My memory is that the primary claim fell away because training took place outside the UK (from memory; not checked). If so, “UK law bans it” [35:58] covers training in the UK, not use of a model trained elsewhere.
  • Government report (18 March 2026). Required by the Data (Use and Access) Act. It kept the status quo, no longer prefers a commercial text-and-data-mining exception with opt-out, and made no firm proposal on mandatory training-data disclosure. Commentators read it as can-kicking (Simmons & Simmons; Hogan Lovells/HSF Kramer note).

EU

  • The AI Act’s general-purpose-AI duties applied from 2 Aug 2025: a copyright-compliance policy and a public summary of training content, using a Commission template published 24 July 2025. The Commission’s enforcement powers began 2 Aug 2026, and models already on the market have until 2 Aug 2027 (Mayer Brown). That is a partial answer to his transparency demand, though a summary is weaker than the list a third party can check he asks for [40:20].

Fairly Trained keeps its certification, which does not accept a fair-use-only basis. I found no current count of certified models, so his “19” is the last number I have (from the talk).

Not checked at all (no search): current AI-music market share, the labour-market studies he cites, model-collapse results since 2025, and whether any “humanist” movement has emerged.

What the talk’s claims look like now

Claim in the talkWhereStatus by Oct 2026
”some AI training is probably fair use and some isn’t”; expect different cases to go different ways[24:43]Borne out. Bartz/Kadrey (generative, books) went one way; Ross (non-generative) the other. No appellate ruling on generative training yet
Fair-use reliance starves licensing[15:49], [19:52]Partly borne out. Settlements show a licensing market forming, but under litigation pressure, not as a result of any fair-use ruling
Open-weights models are irreversible and complicate fair use[24:43]Untested in court as far as I found
The UK government will resist transparency[35:58]Borne out so far (March 2026 report)
Training data lists are easy to produce and need not be secret[34:40]Partly adopted in the EU summary template; not as checkable lists
Synthetic data will be a loophole[39:20]Not checked
Model collapse will not save creators[45:03]Not checked
Music long tail is being eroded[08:44]Not checked

Part 2 — The intellectual map

Newton-Rex spoke in plain terms to a general audience. Behind the plain terms are eight bodies of thought. Read them as a map of why a single question (may a model train on your work?) pulls in law, economics, ML and politics.

 COPYRIGHT THEORY & FAIR-USE LAW ──┐                 ┌── LICENSING MARKETS & CERTIFICATION
 Locke · Hegel · Statute of Anne   │                 │   collective societies · Akerlof
 Sony · Campbell · Google Books    │                 │   Napster→Spotify · Fairtrade
 Warhol · input/output             ▼                 ▼
                           ┌──────────────────────────────────┐
 ML DATA & MEMORISATION ─► │  NEWTON-REX: unlicensed training │ ◄─ OPEN-WEIGHTS & IRREVERSIBILITY
 Carlini · datasheets      │  is a choice, not a necessity;   │    OSD 1998 · OSAID 2024
 model collapse            │  license, disclose, hurry        │    Kapoor et al. 2024
                           └──────────────────────────────────┘
                              ▲              ▲               ▲
   LABOUR & AUTOMATION        │   CULTURE ECONOMICS         │   POLITICAL ECONOMY OF AI POLICY
   Keynes · Luddites · Autor  │   Baumol · Rosen · Benjamin │   Stigler · Make It Fair · AI race
   Acemoglu · Bastani         │   Anderson/Spotify · Morris │   Nye · EU AI Act

2.1 Copyright theory and fair-use law — what the fight is about

The problem it addresses. Why should anyone own an expression, and what can others do with it without asking? A model does not “read” like a person. It makes copies during training and then generates new works. Law written for printing presses and photocopiers must say which step is the copying that matters.

The chain of thinkers and cases

  1. John Locke, Second Treatise (1689). A person owns what they mix their labour with. This is the moral root of the creator’s claim that “my work, my say”. It is the intuition behind Newton-Rex’s phrase “they want to be asked permission before their work is used” [19:52].
  2. Statute of Anne (1710). The first copyright statute. It framed copyright as a limited public bargain: an author’s monopoly for a term, to encourage learning. The US Constitution (Art. I §8 cl. 8) says Congress may secure rights “to promote the Progress of Science and useful Arts” (from memory; not checked). On this utilitarian view the question is not “who owns?” but “what rule best produces new works?“.
  3. Hegel, Philosophy of Right (1821) and the continental droit d’auteur tradition. Work is an extension of the author’s personality. This is why the EU and Germany give authors moral rights and read memorisation as more serious (GEMA, above).
  4. Sony v. Universal (1984), Harper & Row v. Nation (1985), Campbell v. Acuff-Rose (1994), Authors Guild v. Google (2d Cir. 2015), Andy Warhol Foundation v. Goldsmith (2023) (all from memory; not checked). In short: copying for a new function can be fair (the VCR, parody, book search); the effect on the market weighs heavily; and after Warhol, a use that serves the same commercial purpose as the original is harder to call transformative. Ross (2026) leans on that last idea.
  5. The input/output split. Courts and offices separate (a) copying to train, from (b) what the model emits. Newton-Rex mixes them on purpose: “how a model is ultimately used comes to bear on the question of whether it is a fair use of the data that trained it” [23:50]. This tracks the Copyright Office’s approach, and the “market dilution” idea in Kadrey.
  6. Territory. Copyright is national. The EU allows commercial text-and-data mining unless a rightsholder opts out (DSM Directive Art. 4, from memory; not checked); the UK allows only non-commercial research mining; Japan and Singapore are more permissive (from memory; not checked). A model trained in one country can be used in another. This is the weak point in UK training is illegal [35:58].

How it feeds the talk. The whole argument [12:15]–[21:03] is that “fair use” was a late, convenient reading. The law above shows why that is arguable on both sides. The first courts to rule [see Part 1] found training “transformative” but still penalised piracy and left market dilution open.

2.2 Licensing markets, collective management and certification

The problem it addresses. If every use needed a one-to-one deal, nothing would be licensed. How do rights markets form, and how do buyers know a product is “clean”?

  1. Player pianos, the 1909 mechanical licence. After White-Smith v. Apollo (1908) held that piano rolls were not copies, Congress created a compulsory licence for recording songs (from memory; not checked). This is the standard precedent: new technology, no deal possible, so the law set a price.
  2. Collective management organisations (ASCAP and BMI in the US, PRS in the UK, GEMA in Germany). They license catalogues in bulk, and collect and pay out. GEMA is both a rightsholder collective and the claimant against OpenAI. These bodies are the natural counterparties for AI licences at scale.
  3. Napster to Spotify (2001–2008). Newton-Rex draws this analogy himself: “before there were legal ways of streaming music, a whole bunch of people used Napster” [28:52]. The industry sued, then licensed. The 2025 label–Suno/Udio settlements repeat the shape.
  4. George Akerlof, “The Market for Lemons” (1970) and Michael Spence, “Job Market Signaling” (1973) (from memory; not checked). When buyers cannot see quality, bad goods drive out good, unless there is a credible signal. A certificate is such a signal. Newton-Rex says his certification works on trust because “no way of actually scanning” a model for its data [31:33].
  5. Fairtrade, organic and forest labels. Certification schemes (Fairtrade from 1988, FSC from 1993, from memory; not checked) shift buying by a minority of conscious buyers and give law-makers a worked example. He says this directly: legislators “can look to and … point to as an example that this is possible” [30:24]. His own words show the limits: “we’re not going to affect the public’s feelings on this” [28:52].
  6. Why unlicensed firms can undercut. In his account, investors see licensors’ “expenses are going to be higher than the people who are taking their training data for free” [19:52]. That is a classic race to the bottom on a cost the buyer does not see.

How it feeds the talk. The Fairly Trained section is a small-scale version of this history. Its weakness is also standard: the certificate can only be as good as the information supplied. Hence his turn to transparency, audits and red teaming.

2.3 Machine learning: data, memorisation, synthetic data

The problem it addresses. What does it mean for a model to “contain” training data, and can anyone check?

  1. Datasets and their documentation. The Pile and Books3 (2020), LAION (2021–22) and Common Crawl are the scraped sources behind many models (from memory; not checked). “Datasheets for Datasets” (Gebru et al., 2018) proposed standard documentation. This is the technical ancestor of Newton-Rex’s call for a list that a third party … can reliably go and check [40:20].
  2. Memorisation and extraction. Carlini et al. (2021, GPT-2; 2023, diffusion models) showed models can emit near-verbatim training examples (from memory; not checked). The GEMA court used exactly this. Newton-Rex’s red-teaming test (“if … people can get Harry Potter out of your system, then you’re obviously lying” [41:54]) is the same logic turned into an audit.
  3. No reverse-engineering from a model to its data. His claim at [31:33] is correct in the ordinary case: you cannot read off a full training list from weights. Membership-inference tests give probabilities, not proof (from memory; not checked). So disclosure has to come from the developer, and that is why he centres transparency.
  4. Synthetic data and distillation. Models trained on model outputs, as in Microsoft’s “Textbooks Are All You Need” line of work (2023, from memory; not checked), mean a clean-looking dataset may sit on a tainted one. His point that rules must cover “the whole chain of models” [39:20] is a data-lineage requirement.
  5. Model collapse. Shumailov et al. (2023, later in Nature, 2024; from memory; not checked) showed recursive training on generated data can degrade models. Newton-Rex judges this “a very temporary limitation” [45:03]. The later literature is mixed, so treat both sides as open (not checked).
  6. Attribution research. At [01:25:43] he lists work to “shine a light on the black box” and link outputs to training data. The older technical roots are influence functions (Koh and Liang, 2017, from memory; not checked). If this worked at scale, payment by use would become possible. It is still research.

How it feeds the talk. Everything he asks for depends on facts about models that are currently hard to establish: what is in them, where it came from, and what flows into outputs. His policy asks are, in effect, requests to make these facts legible.

2.4 Open source, open weights and irreversibility

The problem it addresses. Releasing a model cannot be undone. How does that interact with the law, which may find the training unlawful later?

  1. The Open Source Definition (1998) and the free-software tradition value freedom to inspect, modify and redistribute. Software’s norms carry over to models with a catch: weights are not source code.
  2. OSI’s Open Source AI Definition 1.0 (Oct 2024) requires detailed information about training data but not the data itself (from memory; not checked). Many “open” models do not meet even that.
  3. Kapoor et al., “On the Societal Impact of Open Foundation Models” (2024) weigh benefits (innovation, auditability) against marginal risk (from memory; not checked). Newton-Rex adds a copyright-specific risk: guardrails on a downloadable model “will just be removed” [24:43].
  4. Remedies in law. If a court finds a model’s training unlawful, a hosted company can switch it off. Open weights can only be made unlawful to host or use, which is harder to police. His answer at [43:28] is that this still reduces use: “most people don’t want to break the law”.
  5. Commercial motives. He argues open releases are commercial strategy by large firms: attract engineers and “build out … the ecosystem” [23:13]. This matches what economists call complements strategy (give away a complement to raise demand for your core product, from memory; not checked).

How it feeds the talk. It explains why he is “an advocate of open source in some areas” but wary here [25:48]. Note the tension with the open-source movement’s own values, which he calls “almost kind of religious adherence” [22:22].

2.5 Labour, automation and the creative “canary”

The problem it addresses. Is generative AI one more wave of automation that will create new jobs, or a general technology that breaks the pattern?

  1. Keynes, “Economic Possibilities for Our Grandchildren” (1930). He predicted “technological unemployment” and a 15-hour week by 2030. The leisure part has not arrived (from memory; not checked).
  2. The Luddites (1811–16). Skilled workers broke frames to protect wages and standards, not out of hatred of machines (from memory; not checked). Debates about “anti-AI” artists often invoke them.
  3. David Autor on why automation has not ended jobs (2015), and Acemoglu and Restrepo on automation versus the creation of new tasks (2019). The standard historical answer to the host’s “300 years ago” point [55:49] is that machines displaced some tasks and made new ones (all from memory; not checked).
  4. Brynjolfsson, “The Turing Trap” (2022). If AI imitates humans instead of augmenting them, workers lose bargaining power (from memory; not checked). This is the logic of Newton-Rex’s “models that are designed to out compete humans in the creative spaces” [01:16:39].
  5. Early evidence on writers and designers. Studies using Upwork data (Hui, Reshef and Zhou, 2023; Demirci and colleagues, 2024) found drops in demand for writing and design freelance jobs after ChatGPT (from memory; not checked). A Danish study (Humlum and Vestergaard, 2024) found little effect on earnings in the wider labour market (from memory; not checked). Newton-Rex cites the first kind [51:41]. The evidence is early and contested.
  6. “Automate all work”. The Mechanize founders have long argued this in their earlier research writing (from memory; not checked). Aaron Bastani’s Fully Automated Luxury Communism (2019) is the optimistic cousin he names at [54:17] (without the book’s title).
  7. Baumol and Bowen, Performing Arts: The Economic Dilemma (1966). Live performance cannot get more productive (a string quartet takes the same four players as in 1800), so its cost rises relative to everything else (from memory; not checked). It helps explain his claim that live music will remain valued while cheap recorded and stock music is replaced.

How it feeds the talk. The “canary” claim is an argument by position: creative work sits first in line for substitution because its outputs are digital and its training data are public. The deeper bet is political: that those who will not pay for data will not redistribute the gains [58:53].

2.6 Culture economics — stars, long tails, recommendation, authenticity

The problem it addresses. If AI makes creation cheap, does culture become more diverse, more homogeneous, or just more crowded?

  1. Sherwin Rosen, “The Economics of Superstars” (1981). Small differences in talent produce large differences in earnings when one performer can serve a whole market (from memory; not checked). Newton-Rex’s point that “creativity is already pretty democratized” [01:06:30] while stars still dominate is this effect.
  2. Chris Anderson, The Long Tail (2004/2006). Cheap distribution would lift niche products. Later data suggested the head stayed strong (from memory; not checked). His “long tail” in the sense of working musicians is a related but different use: tail earners, not tail listeners.
  3. Spotify’s 2020 study (Anderson et al.). Users who followed recommendations listened to a narrower range of music than those who searched [01:08:04] (from memory; not checked, besides the transcript). It supports his claim that distribution, not production cost, drives concentration.
  4. Walter Benjamin, “The Work of Art in the Age of Its Technological Reproducibility” (1935–36). Reproduction removes the “aura” of the unique work, which he links to presence in time and place. Newton-Rex’s idea that human presence (“No one would be remotely interested in robots playing an Oasis concert” [10:07]) matters for music is an aura argument in plain words, though he does not use the term.
  5. Lionel Trilling, Sincerity and Authenticity (1972) and Pine and Gilmore, Authenticity (2007). Authenticity as a social value rises as mass production spreads (from memory; not checked).
  6. Ruskin and Morris, the Arts and Crafts movement (1880s on). A reaction to industrial production that valued hand-making. His predicted “humanist movement” [01:12:28] has this precedent. Whether it holds against cheap alternatives is the question.
  7. Lessig, Remix (2008). A “read-write” culture where consumers remix needs clear licensing. His remix-culture forecast [01:01:05] depends on exactly that: “as soon as licensing gets done at scale”.

How it feeds the talk. The cultural half of the talk is a set of bets: concrete works persist; novelty apps have poor retention; a human-first movement grows; efficiency does not drive choice in art. These are testable, but not yet tested.

2.7 The political economy of AI policy

The problem it addresses. Why do governments side with AI firms, and what does it take to move them?

  1. George Stigler, “The Theory of Economic Regulation” (1971). Regulated industries tend to shape the regulation (from memory; not checked). Newton-Rex says the UK government is “very close to the big tech companies” [35:58].
  2. The UK sequence. A late-2024 consultation favoured a commercial text-and-data-mining exception with opt-out; the January 2025 AI Opportunities Action Plan backed growth; a “Make It Fair” press campaign and the “Is This What We Want?” album (Feb 2025) followed; Baroness Kidron’s amendments to the Data (Use and Access) Bill passed the Lords but were rejected in the Commons (all from memory; not checked, except the album). The March 2026 report (Part 1) is the end of that story, so far.
  3. The AI race frame. He argues ministers fear being “left behind” on AGI more than they want a particular AI company [01:19:50]. This resembles national-champion industrial policy and security framing (from memory; not checked).
  4. Soft power. Joseph Nye (1990) coined it. Newton-Rex’s claim is that US and UK cultural influence comes from “supremely talented people backed by great industries” [01:22:45]. The test is whether weakening copyright would weaken that engine.
  5. Overton window. Joseph Overton’s idea, popularised in the 1990s (from memory; not checked). His aim is not to win a specific rule but to make fairer outcomes thinkable [01:24:03].
  6. The EU route. The AI Act made training-data summaries and a copyright policy law (Part 1). It is the nearest real-world version of what he asks.

How it feeds the talk. It explains the shape of his advocacy: public protest, amendments, certification, and pressure on investors. These are tools for changing incentives when you cannot out-lobby the other side.

2.8 Two shorter threads

  • Consent and moral rights in AI. Creators’ demand for opt-in (permission first) rather than opt-out (object later) is a general rule in consent ethics: silence should not mean yes. Opt-out regimes shift the burden onto millions of small rightsholders who do not know their work was used. He does not use these words in the transcript, but this is the logic behind his objection to an opt-out for UK creators (from memory of his public campaigning; not in the talk).
  • Deep learning’s history in music. From symbolic rule systems and Markov models in the 1950s–2010s, to RNNs and raw-audio diffusion models, which he describes at [02:19]. The shift from writing notes to generating waveforms is also a shift from knowing the rules to absorbing the data, which is why data rights became the central issue.

Part 3 — Deep dives

1. Why he left Stability, and the “licensing was normal” claim

🎙 In the talk. At Stability he built an audio model, and “we licensed all of our training data” [12:43]. In late 2023 he read the company’s public submission to the US Copyright Office and found it said training “without a license is fair use”. The talk gives the gist at [14:20]: that “goes against everything I stand for”. He says “I didn’t really resign from stability so much as I resigned from the wider industry” [13:26]. He had hoped that a well-received licensed model (named among Time’s best inventions of 2023 [15:09]) would show the way, but “most of the reason you don’t see leading models trained on licensed data is because people just can’t be bothered” [15:09].

He then offers a history. “No one trained on copyrighted work without a license for a very long time” [16:46]. Google’s Magenta team, from 2015, wrote papers saying where its data came from and commissioned data, including for an AI drummer [17:23]–[17:57]. Research use gets a better argument for unlicensed data [17:57]. Then “two or three companies” released commercially, everyone copied, and “if they can get away with it, surely we can, too” [18:39].

📚 Deeper theory

  • The two-stage story is a standard pattern in technology law. Research exceptions are broad; commercial release changes the legal question. Many jurisdictions draw exactly this line (UK s29A for non-commercial research, from memory; not checked). His point is that the industry slid from the research frame to a commercial one without changing its assumptions.
  • First-mover norms. Economists call this herding or an information cascade: later firms infer legality from earlier firms’ survival (from memory; not checked). Note that survival proves nothing about legality until a court rules, as the 2025–26 cases show.
  • What the courts did. The first rulings (Part 1) went partly the way Stability’s submission argued, at least for books lawfully acquired. His claim “everyone knew it was illegal” [16:46] is a claim about industry opinion in the 2010s, not about law as decided. It can be right about opinion and wrong about the law at the same time.
  • Why the submission mattered. Public statements to a regulator are strategic. A firm that tells the Copyright Office training is fair use is also helping to create the legal environment its competitors will cite. That makes it, for him, more than a private business choice.
  • Link to 2.1 and 2.2.

Where others push back.

  • The fair-use camp (many US academics; Authors Guild v. Google is their anchor, from memory; not checked) argues training is a non-expressive, analytical use, like a search index, and that transformative use is what fair use is for. The Bartz ruling in June 2025 agreed for the books at issue.
  • Developers say licensing every source is impractical at web scale. He answers that for music “you can do it in an afternoon” [34:40] and for LLMs that it is “garbage” when firms claim they cannot [18:39]. For web-scale text, the practical case against him is stronger than for music.
  • “Research came first” is also arguable. Large web-scraped datasets (Common Crawl, LAION) were widely used in research before commercial release, so the line between “research” and “commercial” was blurry (from memory; not checked).

What he leaves out. The talk does not discuss the licensed-data models’ quality disadvantage, if any, or how many works a licensed model can reach. It also does not cover collective or statutory licences (see 2.2), which solve the “afternoon” problem for large catalogues in a different way.

Check yourself

  • What is the difference between what “everyone knew” in the 2010s and what a court decides? Why does it matter for his argument?
  • What does the Stability anecdote tell you about how firms use regulator submissions?
  • Why does the research/commercial split matter in law?

2. The investor cycle, open weights and irreversibility

🎙 In the talk. Companies that license data “are having a really tough time” raising money, because investors see that their “expenses are going to be higher than the people who are taking their training data for free” [19:52]. He says a “whole industry” now depends on fair use prevailing: “desperate for fair use to prevail in these lawsuits” [19:52]. The host adds that licensed firms also compete with open-weights models [21:03].

On open models he makes three points. They are “100% a commercial thing” for the firms behind them [23:13]. Because they have “no downstream limitation” [23:13], they weaken fair-use defences: the Copyright Office said that use “comes to bear” on fair use [23:50], and potential facilitation matters, not only actual use [24:43]. And “open models are irreversible” [24:43]: if courts find some training unlawful, a closed company can switch off the model but an open one cannot be recalled [25:48].

📚 Deeper theory

  • Cost advantage from externalising a cost. Not paying for inputs is a subsidy from creators to model builders. In economics this is an externality: a cost borne by someone outside the transaction. Law and licensing are how externalities get priced (2.2).
  • Why investors act this way. If an unlicensed rival can launch sooner and cheaper, a VC sees a timing and margin risk. The argument is cold but coherent: it is rational given the legal uncertainty, and it would reverse if unlicensed training became costly. That is why court outcomes matter so much, and why the $1.5B Bartz settlement matters (Part 1): it put a number on the risk.
  • The cost of being wrong on fair use. For a developer, the exposure is statutory damages per work (US law allows large per-work statutory damages for registered works, from memory; not checked). Classes of hundreds of thousands of works turn small per-work amounts into billions. This is the mechanism behind the settlement.
  • “Open” as a legal question. Whether releasing weights is “distribution” of copies depends on whether weights contain works. The UK Getty ruling said they did not (Part 1). Under that view, a downloadable model is not an infringing copy, which weakens one of his arguments about open models (that they carry legal risk downstream). The German GEMA court went the other way for memorised lyrics.
  • Irreversibility and safety. The same “can’t take it back” logic features in AI-safety debates about open weights (Kapoor et al., 2.4). FLI’s own audience will know this, which may be why the host did not push back.

Where others push back.

  • Open-source advocates say open models give researchers and small firms access that licensing would price out, and that access creates the competition that licensing needs. Both can be true.
  • Legal scholars note that courts often grant injunctions with exceptions and sunset periods, and that “unwinding” a model is a remedy courts are slow to order (from memory; not checked). Bartz did order destruction of the pirated files, a narrower remedy.
  • Technical pushback. Guardrails on closed models are also bypassable. The difference is one of degree.

What he leaves out. Open datasets built from the public domain and permissively licensed work (he mentions public-domain datasets only at [01:25:03]). Also non-US open models trained under other legal systems.

Check yourself

  • Why does an externality create a competitive advantage? What would end it?
  • How does the Getty weights ruling cut against his “open models” argument?
  • What does a settlement do to the investor logic he describes?

3. Fairly Trained and the limits of certification

🎙 In the talk. After he left Stability, the story made news, and journalists told him “I thought AI could only be built by stealing people’s work” [26:35]. Fairly Trained is “a very simple certification for AI models that aren’t trained on copyrighted work without a license” [28:08]. It had certified 19 models across music, voice, a large language model and other modalities [28:08]. It is a non-profit, and he does not pay himself [28:52]. It is “a self-certification scheme” in the sense that companies submit information and lists of training data and he and his team check them by drilling down until they have “a high level of confidence” [31:33]. He admits “there is at the moment no way of actually scanning” a model to recover its data [31:33], so it is trust-based. He does not expect the public to care, because they use the best and easiest model, but expects companies, legislators and creators to use it [28:52], [30:24].

The host asks whether it can scale to OpenAI, Google DeepMind and Anthropic. He answers that “the biggest thing stopping that is transparency” [33:07]: the checking is not the problem; disclosure is.

📚 Deeper theory

  • Certification as signalling (2.2). The scheme gives buyers who care a cheap way to find clean models, and gives clean builders a premium. It works best where buyers are organisations with procurement rules, which is the group he targets.
  • Auditing without access. Real audits of complex systems use access (to books, code, data). Without access, an auditor can sample and cross-check but cannot prove a negative. He recognises this at [41:54].
  • The standard has a definition problem. “Not trained on copyrighted work without a licence” has to deal with works under open licences, public-domain works, data licensed by an intermediary who may not hold the rights, and fair-dealing material. Certifiers need rules for each; I did not find his current criteria (not checked). Trade press says the badge was tightened to make clear which parts of a system are certified (from memory of one article’s headline; date not seen).
  • Why the scheme does not need scale to matter. It is evidence for an argument rather than a market: “this is possible” [30:24]. This is the existence proof function of a pilot, as with early organic or Fairtrade labels.
  • A moral claim inside a technical one. He is not neutral about fair use. Fairly Trained does not certify fair-use-only models, which means it encodes his legal and ethical view as a standard. Supporters call this a clear line; critics call it a contested legal position dressed as a quality mark.

Where others push back.

  • Developers say a model trained partly on fair-use material can still be responsible, and that a binary mark ignores opt-outs, robots.txt compliance and output filters.
  • Certification sceptics note that self-reported lists can be wrong, that certificate bodies can be captured by those they certify, and that a non-profit with one founder has limited capacity for large firms.
  • Market-minded critics say that collective licensing or statutory schemes would do more than marks, since most users never look at a mark.

What he leaves out. What happens when a certified model is fine-tuned on someone else’s base model, a case he handles for synthetic data [39:20] but does not discuss in this section. Also how he checks that a data provider had the rights to license.

Check yourself

  • What can a certifier verify, and what can it only take on trust?
  • Why does he say the public will not be moved but legislators might be?
  • What is the difference between a market tool and an existence proof?

4. Transparency, synthetic data and how to check a list

🎙 In the talk. Disclosure is the issue: “if these companies would just reveal their training data publicly then of course you could check it” [33:07]. In the UK the House of Lords twice proposed a short amendment requiring disclosure and the government rejected it, “using arguments that are basically based on procedure” [34:40]. His reasons for thinking “trade secret” is weak: “there is no secret to where you get data” [34:40], and a music company can list who to license from “in an afternoon”. The government rejects it, he says, because “they are very close to the big tech companies” [35:58].

When the host raises special data (commissioned research data, synthetic data), he separates sharing the data from listing sources. Copyright holders “need enough information that lets them know if their work is in the training set” [38:28]. Without it they must red-team models, which “stops them exercising their rights” [38:28]. Synthetic data “can be a way of laundering copyright” [39:20], so the whole chain of models must be named. The test: “can a third party looking at this list you provide reliably go and check it themselves” [40:20]. He notes most regulation “missed the synthetic data problem” [40:20]. To the host’s point that firms could omit data, he answers audits plus red teaming [41:54].

📚 Deeper theory

  • Transparency is a precondition for enforcement. Copyright is enforced by the owner, not the state. If the owner cannot find out about the use, the right is hollow. Discovery in a lawsuit is one route (hence the NYT case’s fights over logs and datasets, Part 1), but it is slow and expensive.
  • The EU template (Part 1) is the first mandatory format. It asks for a summary by categories and top domains (from memory; not checked) rather than a work-level list. Compare that with his bar, which is a list a third party can check. The distance between the two is the live policy gap.
  • Data lineage. The synthetic data point is the same issue as supply-chain due diligence: a firm is responsible for its inputs’ inputs. Product law uses this for materials and conflict minerals (from memory; not checked). He proposes it for data.
  • What audits can and cannot do. An audit can confirm that a declared source set matches a sample, test for undeclared sources by probing, and check processes. It cannot prove no other run exists. Red teaming gives one-sided evidence: a hit proves use, a miss proves little. His own phrase “you can’t prove a negative” is the host’s, and he accepts it [41:54].
  • Trade-secret law. Trade secrets protect information with economic value from being known, subject to reasonable secrecy. Courts balance this against disclosure duties in many fields, such as drug ingredients and financial reporting (from memory; not checked). His argument is that source lists are not valuable because secret.
  • A counterargument on privacy and safety. Data lists can reveal personal data or security-relevant sources, and bulk lists of URLs may expose scraping practices that raise separate issues. This is a reason for careful design, not for refusing.

Where others push back.

  • Developers say lists at web scale run to billions of items and have no stable identity (duplicates, mirrors). Summaries are what is feasible.
  • Some rightsholders want more than lists: payment, opt-in, and the ability to remove their work. A list alone does not give them these.
  • Policy sceptics say transparency without a liability rule only produces paperwork.

What he leaves out. How a list handles data bought from brokers, scraped mirrors and user-uploaded content, and who pays for audits.

Check yourself

  • Why does he separate sharing data from listing sources? Who gains?
  • How does synthetic data open a loophole, and what does he require to close it?
  • What is the gap between the EU template and his test?

5. Is it too late? Recall by law, model collapse and the pace of change

🎙 In the talk. The host asks if the open-source cat is “out of the bag”. Newton-Rex says “I don’t think so” [41:54]. First, you can make hosting, distribution and use illegal; “most people don’t want to break the law” [43:28]. Second, he does not rely on model collapse: “a very temporary limitation” [45:03], so synthetic-only training will likely work. Hence “I don’t think it’s too late but I do think time is of the essence” [45:03]. He then argues from a broader view of technology: intelligence “has already come about once”, so there is no reason to think machines cannot do what people do [46:25]. “Can’t do this yet” is “always a bad argument” because the problem is “going to be solved” within years, perhaps from “some new startup” [48:00]. He recalls getting his timeline wrong in 2010 (too slow) and being wrong that music would come before art [46:25]. The host agrees and gives reasoning models as an example of synthetic data working [45:49].

📚 Deeper theory

  • Model collapse, in more detail. The Shumailov result concerned recursive training on a model’s own unfiltered output. In practice labs mix real and curated synthetic data and add verification (for example for code and maths), which is why synthetic data has worked in reasoning models (from memory; not checked). Newton-Rex’s prediction agrees with that practice.
  • The laundering problem is real only if provenance is hidden. A model trained on synthetic data derived from protected works may or may not infringe, depending on whether protected expression passes through. The legal test is still open (not checked).
  • Legal recall in practice. Orders can bind hosts and platforms (Hugging Face, app stores) and big users. They cannot reach a copy on a private disk. The relevant precedent is software piracy and file-sharing enforcement (from memory; not checked), which reduced mass use but never reached zero.
  • Technology forecasting and the “AI effect”. His argument matches a familiar pattern: tasks once thought uniquely human are redefined as “not real intelligence” once done by machines (the “AI effect”, associated with Larry Tesler, from memory; not checked). He uses it as an argument against complacency.
  • A counterweight: forecasting is not the same as inevitability. The jump from “machines can in principle” to “soon” is the weakest link in any such argument. He takes it as given; others dispute it.
  • Policy lag. If law takes five years and models turn over every year, remedies based on specific models arrive late. This supports his “time is of the essence” and also supports rules aimed at practices (disclosure, licensing) rather than specific models.

Where others push back.

  • Technologists say recall is unrealistic and that the real contest is over future training, not past models.
  • “Collapse” optimists (some creators) hope for a data wall; he warns against relying on it. The host’s point on reasoning models supports him.
  • Critics of his “pace of change” point to slow areas (self-driving cars, medical AI) where forecasts were missed (from memory; not checked).

What he leaves out. Compute and energy limits, and the question whether capabilities that arrive in music will arrive on the same schedule in other fields. He also does not discuss the possibility that the best models need less web data over time, which would change the licensing economics.

Check yourself

  • Why is “recall” different for closed and open models?
  • What would have to be true for model collapse to protect creators?
  • Why does a forecast of rapid progress change what regulation should look like?

6. Labour, “automate all work” and the creatives as canary

🎙 In the talk. He says one of the biggest near-term risks is “the potential mass displacement of labor” and the creative sector is “kind of the canary in the coal mine” [51:41]. He cites data from Upwork and papers on freelance writing and design tasks that “fell … and never recovered” [51:41], and says people he knows lost income because former clients now use AI [51:41]. He cites Mechanize, whose mission is to “automate all work”, backed by prominent names, and says this is “not some fringe idea in Silicon Valley” [52:44]. He expects a huge amount of work could be automated, not all (“not going to be politicians … priests … sports people” [52:44]).

The host asks why this is not like automating agriculture [55:49]. His answer: the technology is “much more general” and the intent is to be general-purpose [56:55]. Then the moral turn: the copyright fight “showed me maybe more than anything that a lot of the people at the forefront … seem willing to trample on people’s rights” [57:30]. If they do not pay copyright holders now, “why will they ultimately pay anyone?” [58:53]. He sees no political will for mass redistribution.

📚 Deeper theory

  • The agriculture analogy, in economics terms. Historical automation (Autor, Acemoglu, 2.5) displaced workers from tasks and created new tasks in which humans had an advantage. The question is whether a general-purpose technology leaves any such tasks, or whether the new tasks are also automatable. This is an open empirical question, not settled by either history or forecast.
  • Transitions are the problem even if the endpoint is fine. Even optimists about the long run accept large distributional costs in the transition (the English handloom weavers, from memory; not checked). His worry about unpreparedness is about transition costs.
  • The redistribution argument is an inference about character, not an economic result. It reads present behaviour (not paying for data) as evidence of future behaviour (not sharing gains). This is plausible as a political judgment but it is a different kind of claim from the labour-market data. The Anthropic settlement (Part 1) is a counter-example: a firm paid, though under court pressure. It supports his claim that pressure is what works.
  • Creative work is not the only canary. Translators, illustrators, junior coders and customer-service staff are also exposed (from memory; not checked). His reason for choosing creators is that they are first in line and that their work is the training data, so they bear the cost twice.
  • The measure matters. “Freelance task postings fell” is not the same as “creative incomes fell” or “creative employment fell”. Platform data capture one segment. His personal anecdotes [51:41] are illustrations, not evidence.
  • Link to Baumol (2.5). Live and relational work may gain relative value as routine production is automated. That would shift, not remove, creative income.

Where others push back.

  • Labour economists often find modest aggregate effects so far and large uncertainty about the future.
  • AI-optimist economists argue lower costs raise demand for creative products and for complementary human roles.
  • Some in the AI safety community share his worry about displacement but disagree about the route, preferring policy on AI deployment to copyright as the lever.

What he leaves out. Concrete proposals for what should replace lost income (licensing pools, taxes, public funds). He frames copyright as a mechanism for sharing gains, but does not argue that it would be enough.

Check yourself

  • What distinguishes this wave from the farm-machinery wave, in his view?
  • Why is “they don’t pay now, so they won’t later” an argument about character and not about economics?
  • What would count as evidence that creatives are, or are not, a canary?

7. Culture: remix, concrete works, winner-take-all and homogenisation

🎙 In the talk. Asked about the future of culture, he first points to harder entry into the creative industries as “long tail” jobs go [58:53]–[01:00:26]. He expects a rise of remix culture once licensing is done at scale: a fan could say can I hear Noel Gallagher singing it [01:01:05]. But he has “quite high faith that concrete works … will for a long time remain the norm” [01:02:13]: shared songs and books that are “part of the public consciousness”. On interactive avatars he is more open; on personalised content he is not. “Write me a song about X” is “an absolute moment of magic the first time you hear it” and then “you never use it again” [01:04:56], with, he guesses, “atrocious” retention.

On winner-take-all: “creativity is already pretty democratized” [01:06:30] and abundance already exists. The concentration comes from recommendation systems; he cites the 2020 Spotify study [01:08:04]. He does not expect homogenisation, since recommendation can also create niche pockets and “a lot of it comes down to the user” [01:10:57].

📚 Deeper theory

  • Why shared works matter. Culture is partly a common reference: songs and books give people something to talk about. Sociologists describe this as shared cultural capital (Bourdieu) or “water-cooler” value (from memory; not checked). Hyper-personalised content loses it. His bet is that this value is large.
  • The retention claim is an empirical hypothesis about a particular use case. He gives no numbers, just “my bet” [01:04:56]. The evidence on current AI music apps is not something I checked.
  • Rosen and the long tail (2.6). Production cost is not the bottleneck for attention, so cheaper production does not change who wins. This is a strong argument, though it ignores a second effect: AI-generated catalogue can flood platforms and reduce the share of royalty pools for human artists (a point he raises at [01:18:12]).
  • Recommenders both narrow and widen. The Spotify result shows narrowing when users follow suggestions. His TikTok experience points the other way: fast feedback finds niche tastes [01:09:23]. Both can hold: different people, different systems, different design.
  • Authenticity and aura (2.6). The idea that people will pay for human presence is plausible for live music. For recorded music, listeners may care less about origin than he supposes, as with many listeners of stock or library music. He says the long tail of production music is exactly where this fails [01:00:26].
  • Remix and consent. A licensed remix economy needs a way to pay many rightsholders cheaply. That is the old collective-licensing problem (2.2) in a new form.

Where others push back.

  • AI-music proponents argue personalised music will find uses (background, games, therapy) and that retention figures will change as tools improve.
  • Cultural critics worry less about homogenisation of taste than about the flattening of what is made by models trained on the average of past output (a theme in many discussions of “mean images”, from memory; not checked).
  • Platform researchers note that recommendation effects depend heavily on interface design, so a general claim is risky.

What he leaves out. Other cultures and languages, where training data are thinner and the “average” imposed by a model is different. Also non-commercial culture (community, ritual, folk practice), which may be affected differently.

Check yourself

  • What is a “concrete work”, and why does he think it survives?
  • How can recommendation systems both narrow and widen listening?
  • Why does cheap creation not change who wins, on his view?

8. The “humanist movement”, efficiency and the UK government

🎙 In the talk. “The backlash to generative AI is just massive”, and tech people “underestimate” it [01:10:57]. The camera and synthesizer analogies fail because “the strength of feeling is much much greater with AI” [01:12:28]. He predicts “a new kind of humanist movement in the arts” [01:12:28], perhaps acoustic and live, which will begin with “a rejection of a few specific practices” [01:15:39]. He cites the silent protest album “Is This What We Want?” (about a thousand UK musicians) [01:14:04]. He admits “I’m not sure this movement like exists” [01:15:39]. The contrast: a song on a guitar in a room versus one “composed on GPUs by a model that’s been trained on the Beatles work without permission” [01:18:12].

The host asks whether such humanism will lose to competitive pressure. He answers “Creativity isn’t all about efficiency” [01:18:45], and “no one’s ever going to listen to that song again” [01:21:22] about a song made in ten seconds. On politics, he says the UK government is “all out favoring AI companies” [01:19:50], and that the real driver is the AGI race and fear of being left behind, not a wish for a domestic AI music firm. The host adds soft power [01:22:13]. He replies that weakening copyright will “increase the amount of slop” and “eat into real creators royalty streams” [01:22:45].

📚 Deeper theory

  • Backlash movements have a history. Arts and Crafts (2.6) and the later “slow” movements emerged against mass production. They did not stop industrialisation. They created premium niches and shifted norms. The likely lesson is segmentation: a human-made premium market alongside cheap machine output.
  • Efficiency is a poor measure of artistic value, but not of commercial value. Production music, jingles and stock imagery are sold on cost and fit. He concedes that these are where harm lands [01:00:26]. So “creativity isn’t about efficiency” is true for great art and weaker for the long tail, which is the labour question in section 6.
  • Is the backlash as strong as he says? I did not check opinion polling or usage figures. Usage of generative tools has also grown fast, so strong feeling and high use can coexist (from memory; not checked).
  • Political economy. His reading of the UK stance (AGI race, not creative-industry interest) is an explanation of motives. The 2025–26 record is consistent with a government that wanted to avoid committing (Part 1), but I did not check ministers’ stated reasons.
  • The strongest version of his policy claim. The UK creative industries are a large export sector. Weakening their rights for uncertain returns in AI is a bad trade (from memory; not checked, for the figures). He argues this at [01:22:45] without numbers.
  • Soft power (2.7). Nye’s concept fits cultural industries well. The counterview is that AI could extend soft power by exporting a country’s cultural style through models, which the host hints at.

Where others push back.

  • Tech optimists say generative tools extend creativity to more people and the ‘humanist’ framing is nostalgic.
  • Pragmatists in government say the UK cannot regulate training for the world and risks driving AI work abroad.
  • Some artists use AI tools as a medium and reject the idea that using them is anti-human.

What he leaves out. A positive account of artistic AI use that he would endorse. He says he is not against AI per se, and Fairly Trained exists to support fair models, but the talk spends little time on what good use looks like.

Check yourself

  • Why does he think the camera analogy fails?
  • Where is “efficiency” a good argument against him, and where is it not?
  • What does he think motivates the UK government, and what evidence would test this?

Putting it together

The mental model in one paragraph. Newton-Rex says unlicensed training is a choice, not a technical necessity: it was not the norm until a few firms shipped commercially, and the rest followed because the first movers were rich and unpunished. The choice has costs that fall on creators (long-tail income) and on firms that license (investors punish higher costs). Because model builders cannot be inspected from outside, the practical remedy is transparency (lists a third party can check, covering the whole chain of models) backed by audits and red teaming; certification shows licensed AI is possible; and time matters because open models cannot be recalled and synthetic data may remove the data bottleneck. Behind the copyright dispute sits a larger bet: that automation of cognitive work is coming, that creatives are first in line, and that people who do not pay for data will not share gains. His cultural forecast is optimistic about human art (concrete works, live music, a humanist movement) and pessimistic about the working creative’s economics. The 16 months since the talk have shown that courts will split by facts (books, headnotes, lyrics), that big payments and licences follow litigation, and that governments move slowly on disclosure.

Three tensions to keep open

  1. Input or output? He folds training and use together [23:50]; courts have treated acquisition, training and output separately, with different results.
  2. Licence first or licence later? Fairly Trained certifies firms that licensed first. The music settlements convert firms that licensed later. Are they the same outcome?
  3. Disclosure of what? A summary (EU), a checkable list (his bar), or the data itself (what firms fear)? The policy gap sits here.

Reading path

  • Start: The US Copyright Office’s AI report (Part 3, generative AI training, pre-publication May 2025; I did not read it for this page). Fairly Trained’s website for the standard. The Bartz and Kadrey summaries linked in Part 1.
  • Then: The EU Commission’s training-data summary template and the UK March 2026 report (Part 1 links). Carlini et al. on extraction. Gebru et al., “Datasheets for Datasets”. Autor, “Why Are There Still So Many Jobs?“.
  • Deeper: Authors Guild v. Google (2d Cir. 2015) and Warhol v. Goldsmith (2023) for the shape of transformative use. Baumol and Bowen, Performing Arts. Benjamin’s “Work of Art” essay. Brynjolfsson, “The Turing Trap”. Lessig, Remix. Newton-Rex’s own book The Intelligence Heist (not read).

Q&A

(Add questions and answers here only if worth keeping.)

Graph View

Backlinks