Talks

Platforms, AI and Culture Production — Thomas Poell

61 min read

Thomas Poell · 2025 · 1:09:32

Summary

Poell asks one question: when platforms and now generative AI enter the cultural industries, what changes, for whom, and where? He answers in two steps.

Step 1 — the book (platforms). A platform is a data infrastructure that facilitates, aggregates, monetises and governs interactions between end users, content providers and service providers. That gives three institutional dimensions: infrastructure, market, governance. Platformization is the spread of these into other sectors, and the way cultural workers and audiences reorganise their practices around them. The early-2000s story (platforms as democratising) is naive but explains their pull. Platform markets are hierarchical because of network effects, which produce winner-take-all dynamics, and the winners are often the already-big media firms. Because platforms keep changing algorithms, APIs, SDKs and terms, small producers live in constant, exhausting adaptation, which is precarity.

Step 2 — the new project (global perspectives on AI and creative work). Almost all this theory comes from North America and north-west Europe. His team (case-study specialists in Chile, Brazil, Nigeria, India, Japan) treats each case as theory building, to “multiply our frames of reference”. He uses Japanese anime and manga to show three things. (a) Gen AI enters particular stages of production, not the whole thing, and hits some jobs (background colourists, voice actors, translators) more than others. (b) Precarity is “situated”: in rural China, Ghana and Brazil it is a long-standing norm, and platforms can even create “unlikely creative workers”. (c) Creativity should be asked “where is it”, not “what is it”: it is networked and distributed across people and technology. Governance differs by region: Japan and China promote AI in culture; the EU restricts.

Close and Q&A. He ends by pointing to a special issue of the International Journal of Cultural Studies with 20 papers. In the Q&A he says big copyright holders will win licensing deals while individual creators lose, AI search summaries threaten the click economy, “human touch” is the wrong frame because production is hybrid, gen AI will fade into the background like CGI, the cause is the deep-learning paradigm plus huge investment, not “evil technocrats”, and the EU’s DSA, DMA and AI Act are “absolutely essential”, though he expects “a war” with US tech and the US government and “quite dark” years.

Stance to note. He is a critical-institutional scholar, not a technology booster or a doomer. He keeps saying “it is not either/or”. He calls the early democratisation story “naive”, yet says we must still grasp its pull. He is open that this is early-stage research: “we’re very much at the starting phase”.

Map — talking points

#Point
1Host intro; Poell’s plan: two projects, the book, then the new global AI project
2The book’s scope: institutional dimensions of platforms linked to cultural practice. Built on Anglo-American and European research
3Definition of a platform (data infrastructure that facilitates, aggregates, monetises, governs). Three dimensions. Apple App Store as example
4Platformization: spread into sectors, and cultural practices reorganising around platforms. Distinct from digitisation. Netflix, Disney, NYT are media firms, not platforms
5The 2004–2010/11 democratisation and participatory-culture literature: naive, but explains the pull of platforms
6Platform markets are hierarchical: network effects, winner-take-all, big media firms dominate, creators professionalise
7Concentration of power, and precarity from constant changes in algorithms, SDKs, APIs, terms
8Move to project 2: de-Westernising, a task for scholars in Europe and the US too; global perspectives (Chile, Brazil, Nigeria, India, Japan)
9Japan anime and manga case: URVAN (2021), the Midjourney manga (2023), voice and translation, the 95% AI series (2025), government support, artists’ petition
10Concept 1: labour precarity. Break production into stages. Precarity is situated (rural China, Ghana, Brazil)
11Concept 2: creativity as networked and distributed (actor-network view). Context matters (Japan’s AI ecosystem)
12Concept 3: governance. Japan and China promote; EU restricts; platforms in China resemble media companies
13Pointer to International Journal of Cultural Studies special issue (20 papers)
14Q&A: IP and consent in East Asia. Big copyright holders win deals; individual creators lose
15Host adds the Microsoft publisher deal; Le Monde; YouTube Content ID as precedent; synthetic data
16Q&A: AI summaries cut clicks, raising precarity. News outlets can fight back, find niche, build subscribers
17Q&A (Hana): “human touch”. Hybrid, distributed, 95% AI plus human quality control
18Q&A (host): maybe novelty fades and the human touch wins. Poell: CGI analogy; gen AI will be woven in
19Q&A (student): why creative jobs? Deep-learning paradigm, compute, investment; not just “technocrats”
20Q&A (host): EU AI Act, DSA, DMA. “Essential.” Expect a “war” with US tech; “quite dark” years
21Close

Putting it together

Mental model. Think of cultural production as a chain of stages (idea, script, make, finish, translate, distribute, market). Platforms came in at the distribution and market end: they became open infrastructures, then hierarchical markets, then rule-setters that producers must keep adapting to. Network effects concentrate power in a few platforms and, inside them, in firms that already have resources; the many small producers carry the risk, which is precarity. Gen AI now enters at the making end, but stage by stage, taking first the tasks that are routine and well supplied with data (backgrounds, colouring, translation, voices). It is fed by cultural work done by others, which is why copyright and consent matter, and why big holders and big AI firms bargain while individuals do not. Creativity in this world is not “in” the person or the model but in the network between them, so the useful questions are where, at which stage, for whom, and on whose terms. And because almost all of this theory was built from US and north-west European cases, Poell argues we should multiply our reference points (Japan, China, Brazil, Nigeria, India, Chile) and check whether concepts like precarity and governance still hold. Governments differ in whether they promote or restrain, and the EU’s restraint is under pressure.

Open tensions in the talk.

  1. General mechanism vs situated difference. Network effects look general; precarity is “situated”. Which parts of the platform story are universal?
  2. Distributed creativity vs who is owed. If creativity is networked, how do credit, pay and consent follow it?
  3. “Not an evil technocrats story” vs concentration of capital. A structural account can hide who decided what. Is “the paradigm” an excuse?
  4. Defending European values while provincialising Europe vs.
  5. Big winners predicted, but also niche hope. He says big holders win yet advises small news outlets to find niches. Both can be true; the second is a survival strategy, not a reversal.
  6. Early-stage evidence. Several claims (70–80% concern, 95% AI) lack checked sources; the global cases are mostly promised, not yet shown.

Reading path

  • Start: Poell, Nieborg and Duffy, Platforms and Cultural Production (2021), the first two chapters; then van Dijck, Poell and de Waal, The Platform Society (2018). Gillespie, “The politics of ‘platforms’” (2010), as a short opener.
  • Then: Nieborg and Poell (2018) on the contingent cultural commodity; Hesmondhalgh, The Cultural Industries; Duffy, (Not) Getting Paid to Do What You Love; Chakrabarty, Provincializing Europe (introduction); Curran and Park, De-Westernizing Media Studies; the special issue Poell points to (International Journal of Cultural Studies; title not found, ask him or use the UvA project page).
  • Deeper: Srnicek, Platform Capitalism; Bradford, Digital Empires; Crawford, Atlas of AI; Gray and Suri, Ghost Work; Acemoglu and Restrepo on tasks; Latour, Reassembling the Social; Becker, Art Worlds; Lin and de Kloet on Kuaishou (check it); Condry, The Soul of Anime; Whittaker, “The steep cost of capture”.
  • Related page: the Citton page on attention, gesture and machines (citton-attention-gesture-machine) approaches the same AI-and-craft question from philosophy of technics; Poell’s is institutional and economic.

Part 1 — Before you read

Key terms

  • Platform (Poell et al.’s definition). “Data or sometimes … computational infrastructures that facilitate, aggregate, monetize and govern interactions between end users, content and service providers”. Watch out: the word has at least four senses in the literature (see table).
    TraditionWhat “platform” means
    Business studiesA multi-sided market that matches groups (buyers/sellers, users/advertisers) and profits from the matching
    Software and computingA programmable base that third parties build on (an OS, an API)
    Political economyA capitalist business model built on data extraction (Srnicek’s “platform capitalism”)
    Gillespie (2010), discourseA rhetorical word that platform firms use to sound neutral and open (“raised stage”, “opportunity”)
    Poell’s definition is a deliberate merge of the first three, as he says: “we build on a number of different traditions”.
  • Institutional dimensions. Poell’s frame: infrastructure, market, governance. A platform is all three at once.
  • Platformization. Two parts: (1) the penetration of platform infrastructures, markets and governance into a sector; (2) the reorganisation of cultural practices (labour, creativity, democracy) around platforms. It is not the same as digitisation.
  • Platform vs media company. A platform is “directly economically and infrastructurally accessible to third parties”. YouTube lets anyone upload. Netflix does not. So Netflix, Disney and the NYT are media or entertainment firms even though they use data and algorithms.
  • Network effects. More users make a platform more valuable to content providers and advertisers, and more providers make it more valuable to users. Each side feeds the other.
  • Winner-take-all. The result of network effects: “only one or two or maybe three platforms” dominate a sector.
  • Precarity. Insecure, unstable work. Poell uses it both as the book’s finding and as a concept to rethink globally.
  • Situated. Poell’s word for a phenomenon whose meaning depends on its place, workers and economy: precarity “derives its meaning from … complex articulation of workers’ experiences and also positions”.
  • Provincialising / de-Westernising. Treating Western-derived theory as one regional view among many. From Chakrabarty (2000) and the older “de-Westernising media studies” movement (Curran and Park, 2000). Poell calls it work that builds on decades of postcolonial, decolonial theory.
  • Networked (distributed) creativity. Creativity is not a trait inside a person or a machine. It is spread across people, tools, sets, scripts, and so on. Poell calls it “an actor network perspective”.
  • Deep-learning paradigm. Poell’s term for the shift to large neural networks trained at scale on cultural and social data. It explains why creative work is hit.

People and works named (a reading list)

Person / workWhy it comes upStart with
David Nieborg, Brooke Erin Duffy (co-authors)The bookPlatforms and Cultural Production (Polity, 2021)
José van Dijck, Martijn de WaalPlatform SocietyThe Platform Society (OUP, 2018)
Host’s intro: Sage Handbook of Social Media; “social media, materialities and protest”Poell’s earlier editing workTitles as spoken; not checked against catalogues
Business-studies work on platformsSource of “network effects”See idea 1
2004–2010 “democratisation” booksThe naive frame he criticises. He shows them on a slide but does not name them in the audioSee idea 2 for likely candidates (my inference)
“Unlikely creative workers” (colleagues)Platforms as opening for marginalised peopleSee idea 3
International Journal of Cultural Studies special issue (20 papers)His project’s frameworkShown by QR code; I could not find the title or DOI online

Poell’s own frame

Poell’s line of work, from The Platform Society (2018) to the book (2021), is institutional: study platforms as infrastructures, markets and governance frameworks, and watch how sectors and practices reorganise. His new project adds a global and decolonial question: whose cases built the theory? A UvA page for the project says it aims to “provincialize the US” and multiply frames of reference by joining platform studies with postcolonial and decolonial theory (UvA profile; the July 2025 UvA workshop he organised is at IAS UvA). He has given versions of this talk in Hong Kong, Brazil and elsewhere (Unicamp event note). I could not find this particular event (host, institution, date); the date above is inferred from the talk.

The current debate (2025–26)

All items below were checked by search in October 2026 unless marked.

  • Japan: anime, manga and AI. Twins Hinahima (broadcast from 29 March 2025, by Frontier Works and KaKa Creation) claimed over 95% of cuts used AI, with people doing final edits. Critics said the figure is misleading, and fan reaction was largely negative (Inquirer). The 2023 manga Cyberpunk: Peach John was made with Midjourney by an author who said he had “absolutely zero” drawing talent (AFP via Ahram). After OpenAI’s Sora 2 launch (late Sept 2025), Japan’s government sent a formal request to OpenAI, and the industry body CODA, speaking for major publishers and studios, demanded that members’ works not be used for training without prior permission and objected to opt-out (Automaton). This is what the artists’ petition grew into. No resolution found.
  • Copyright and licensing. Anthropic’s $1.5 billion settlement with authors (Bartz v. Anthropic) got final approval on 20 July 2026, the largest US copyright class-action settlement (Bloomberg Law). It covered pirated downloads, not training as such. Licensing deals continue: Le Monde–OpenAI (March 2024, terms undisclosed; TechXplore) and Microsoft–HarperCollins (opt-in per author; eMarketer). The NYT suit against OpenAI and Microsoft was filed in Dec 2023 and seeks “billions” (Forbes AU); I could not confirm its 2026 status. Poell’s prediction (big holders get deals; the rest lose) fits the pattern so far, but the settlement shows courts can also pay out to individual authors.
  • EU regulation under pressure. In Nov 2025 the Commission proposed a “digital omnibus” that delays the AI Act’s high-risk rules; a provisional deal pushes them from Aug 2026 to 2 December 2027, and critics say Europe is caving to Big Tech (To Vima). Poell’s warning of “a war” over the EU laws looks, in hindsight, partly borne out. Specific US-government pressure was only thinly sourced in what I found.
  • Platform studies itself. The “global perspectives” push is ongoing (CAIS Germany, CUHK Hong Kong, Unicamp Brazil events in 2025). I did not find a major published critique of it; treat that as a gap in my search, not proof.

The ideas

Poell spoke for about 39 minutes to a mixed audience. He named few authors. The talk rests on seven families of ideas. Here they are laid out, including thinkers he does not name. Everything in this part is background, not Poell’s claim.

 PLATFORM STUDIES ──────────┐                  ┌──── CULTURAL-INDUSTRIES STUDIES
 business · software ·      │                  │     Adorno · Hirsch · Bourdieu
 political economy          ▼                  ▼     Hesmondhalgh · Caves
 Gillespie · Srnicek ┌────────────────────────────────┐ Neff · Duffy
 van Dijck · Rochet  │  POELL: platformization of     │
 & Tirole · Gawer    │  cultural production, then     │
                     │  gen AI, seen from many places │
 PRECARITY & LABOUR ►│                                │◄ DECOLONIAL & SOUTHERN THEORY
 Standing · Gill &   └────────────────────────────────┘  Chakrabarty · Connell · Curran
 Pratt · Gray & Suri      ▲            ▲           ▲      & Park · Couldry & Mejias
                          │            │           │
          DISTRIBUTED CREATIVITY   GEN AI: HOW IT   GOVERNANCE & REGULATION
          Latour · Becker ·        WORKS & WHO PAYS Bradford · DSA/DMA/AI Act
          Csikszentmihalyi · Boden Bender · Crawford Japan soft law · China state
                                   Sutton · Villalobos

Where each school of thought is covered:

  • Platform studies — three roots, one definition: idea 1
  • Cultural-industries studies — gatekeepers, fields, and creative work: idea 2
  • Precarity and platform labour: idea 5
  • Decolonial, postcolonial and “Southern” theory: idea 3
  • Distributed creativity — from the lone genius to the network: idea 5
  • Generative AI — how it works, who pays, what comes next: idea 6
  • Governance — three regulatory models: idea 7
  • Technology, tasks and labour — why “break it into stages” works: idea 4
  • A shorter thread: network effects and “winner-take-all”: idea 2

1. What a platform is, and what platformization is

In short. Poell defines a platform as a computational infrastructure that facilitates, aggregates, monetizes and governs interactions. It has three dimensions: infrastructure, market and governance, and the Apple App Store shows all three. Platformization is the spread of these into cultural industries, with producers and consumers reorganizing around platforms. Netflix is a media company, not a platform, because third parties cannot upload to it freely.

🎙 In the talk. Poell’s definition: “data or sometimes we call them computational infrastructures that facilitate, aggregate, monetize and govern interactions between end users, content and service providers”. Three “institutional dimensions”: the infrastructures … the platforms as markets … and then finally … platforms as governance frameworks. The Apple App Store shows all three. It is a market, with about 2 million apps. It runs on a data infrastructure that lets developers upload. And Apple “closely governs” it by vetting apps, supplying SDKs “with all sorts of standards”, and publishing developer agreements.

Platformization is “the penetration of digital platforms’ economic, infrastructural and governmental extensions into the cultural industries”. It also means that “cultural producers but also consumers … reorient themselves, reorganize themselves around platforms”. He separates it from digitisation. Netflix, Disney and the NYT “collect loads of data” and use “very sophisticated algorithms” but are “media or entertainment companies” because they produce and license content and are not “directly economically and infrastructurally accessible to third parties”. “Video producers, small or big, cannot freely upload their clips to Netflix in the way they can for example to YouTube”.

📚 Deeper theory

  • The test is openness, and it matters analytically. Two-sided market theory (Roots) says a platform’s value comes from third-party participation. Openness is what sets off network effects and the later dynamics (concentration, dependence). So the line he draws is not decoration; it picks out the cases where his later argument applies.
  • Why include “governance” in the definition? In business studies governance is a design variable (a rule set to keep quality). In political terms it is power: Apple decides what can exist. Poell’s definition keeps both readings. See Gillespie (2018) in idea 7.
  • Three sources, three blind spots. Business studies sees markets but not power; political economy sees power but not code; software studies sees code but not markets. The merge is the point, but it is also a risk: a definition that tries to hold all three is broad enough to include many things.
  • Platform vs infrastructure. Plantin et al. (2018) argue the line blurs: Facebook is a platform for apps and also infrastructure for news. The same actor can sit on both sides of Poell’s test.
  • The test is not stable. Netflix hosts third-party licensed content and some user profiles; Spotify commissions podcasts; TikTok curates and pays creators. As firms move, the label shifts. Poell notes this for China.

📚 Roots: Platform studies — three roots, one definition

The problem. “Platform” is a slippery word. Is it a market, a piece of software, a business model, or a marketing term? Poell’s answer is to join the first three.

The point. The definition and the three dimensions are the output of this chain. His distinction between platform and media company is a test of third-party access: can others upload and build? That test lets him study YouTube and the App Store but exclude Netflix.

The chain

  1. Business studies: two-sided markets and network effects.
    • Katz and Shapiro (1985) gave the economics of network externalities: a product’s value rises with the number of users.
    • Rochet and Tirole (2003, “Platform competition in two-sided markets”) formalised markets with two groups that need each other (cardholders and merchants). Pricing is set to keep both sides on board.
    • Gawer and Cusumano, Platform Leadership (2002) and Gawer (2009) studied “innovation platforms” (Intel, Microsoft) that third parties build upon. This is the source of the infrastructure plus third-party complement idea.
    • Parker, Van Alstyne and Choudary, Platform Revolution (2016) is the practitioner version. This is the tradition Poell means by “business studies” and by “network effects”.
  2. Software studies and infrastructure studies.
    • Tarleton Gillespie, “The politics of ‘platforms’” (2010) showed that firms like YouTube chose the word platform for its multiple meanings (computing base, raised stage, political podium, opportunity) to look open and neutral. This is where “the naive reading” of platforms partly comes.
    • Anne Helmond, “The platformization of the web” (2015) gave the word platformization: platforms extend into the rest of the web through plug-ins, APIs and “Like” buttons.
    • Plantin, Lagoze, Edwards and Sandvig (2018), “Infrastructure studies meet platform studies” argued the two fields describe converging things: platforms are becoming infrastructures and infrastructures are becoming platforms.
    • Taina Bucher, If … Then (2018) on algorithmic life; Bucher and Helmond (2018) on affordances of platforms. Poell’s remarks on SDKs, APIs and documentation shaping what developers can do sit here.
  3. Critical political economy.
    • Nick Srnicek, Platform Capitalism (2016): platforms are a business model for extracting and monetising data. Lean platforms, cloud platforms, product platforms.
    • Langley and Leyshon (2017): “platform capitalism” and its “intermediation” role.
    • Shoshana Zuboff, The Age of Surveillance Capitalism (2019): behavioural data as raw material.
    • Earlier lineage: Dallas Smythe’s “audience commodity” (1977), Vincent Mosco and Christian Fuchs on digital labour.
  4. Platformization as a joint concept. Nieborg and Poell (2018), “The platformization of cultural production” (New Media & Society) defined the “contingent cultural commodity”: content that is continuously updated and tied to a platform’s data and rules. Poell, Nieborg and van Dijck (2019, Internet Policy Review) defined platformisation as the penetration of infrastructures, economic processes and governmental frameworks. Poell restates this almost word for word. van Dijck, Poell and de Waal (2018) carried it into health, education, transport and news, and added public values as the test.

A note on the test. It also explains the odd line that Chinese platforms now “resemble media companies much more strongly than our definition of what a platform is. By his own test they drift out of the category by curating and commissioning. That is a hint that the 2021 definition is under strain, which he says he is happy to rethink”.

Where others push back.

  • “Platform” is too broad or too soft. Some scholars (e.g. in infrastructure studies) prefer infrastructure or ecosystem; some economists say “platform” just means intermediary and is old (newspapers, shopping malls).
  • Platform exceptionalism. Others say platform power is a case of older monopoly and rent-seeking, not new in kind.
  • Technological determinism worry. “Platform” can seem to explain too much of cultural change that has other causes.

What he leaves out. Hardware and cloud layers beneath platforms (AWS, Azure, chip makers). Gen AI makes these layers matter more, since a few cloud firms also fund and host model makers. This was not part of the talk. (Not his claim; it is a gap.)

Check yourself

  • Name the three institutional dimensions and show them in one platform.
  • Why is Netflix not a platform under his definition, and what would change that?
  • How does platformization differ from digitisation?

2. Winner-take-all markets and precarity — the book’s argument

In short. Early research on platforms stressed democratization, and Poell calls that naive. He still keeps part of it, because it explains why platforms draw producers. But platform markets are very hierarchical. Network effects lead to winner-take-all dynamics, where one to three platforms dominate a sector, and inside platforms engagement rewards the already visible.

🎙 In the talk. He starts from the 2004–2010/11 research on “democratization, emancipation and participatory culture”. He calls it “naive” and says they “criticized that very extensively in The Platform Society”, but “we also still want to hold on to part of that analysis” because it explains “the opportunities that platforms offer to culture producers … to find audiences and generate revenue”. Otherwise “it becomes very difficult to understand the drawing power of platforms”. Platforms “appear to facilitate a decentralization of power” for individual producers.

Then the complication: platform markets are “in no way neutral … very very very hierarchical”. The key concept is network effects. “If you succeed … to attract a lot of end users you become very attractive … for content and service providers as well as for advertisers”, and it works the other way too. The result is “winner take all dynamics where only one or two or maybe three platforms tend to dominate an entire sector”: YouTube long-form, TikTok short-form, the Apple and Google app stores, Facebook social networking, Google search. News organisations “might not like Facebook or Instagram very much … but they need to work with these platforms”.

Inside platforms there is a second winner-take-all. Platforms “reward actors that generate a lot of user engagement … making them more visible … which leads to more engagement”. The winners are firms “with lots of resources … talent agencies … partner programs, data tools … brand recognition”: “Walt Disney, Bloomberg, the News Corp, the New York Times”. Individual creators who win “start to work with talent agencies … become more sophisticated in data analytics … become brands in their own right”.

The final claim is precarity: platforms “constantly adapt the way in which third parties … can work”: “changing their algorithms, changing their software development kits, changing access to their data through the APIs, changing their terms of service”. Producers must adapt, “which is very exhausting and puts them in a really precarious position”. He calls the argument “not too optimistic … a realistic assessment of the current power relations”.

📚 Deeper theory

  • Network effects come in two kinds. Direct (more friends on the network) and indirect (more apps draw more users who draw more apps). Poell describes both. Multi-homing (using several platforms) and switching costs decide whether winner-take-all holds. Creators multi-home more than users, but they are tied by audiences and revenue share.
  • Why visibility compounds. Ranking systems use engagement as a signal, so early engagement is rewarded: a preferential attachment process. It produces power-law distributions in followers and views. Merton’s Matthew effect (Roots) is the sociological name.
  • Why big firms win again. The platform’s “open” door lowers entry costs, but staying visible costs money: ad spend, agencies, data tools, rights management. Bourdieu’s field theory (Roots) predicts that capital converts across fields: prestige and money from print and TV buy position on YouTube.
  • Precarity by design: the “contingent cultural commodity”. Nieborg and Poell (2018) argued that platform-bound content is contingent: it is continuously updated, data-driven and dependent on a platform’s changing rules (a game app that needs patches; a news post that depends on a ranking tweak). A producer’s asset can lose value when the platform changes a rule, with no negotiation. Examples: Facebook’s 2018 News Feed change cutting publisher reach; YouTube “adpocalypse” demonetisation; Twitter/X API pricing in 2023. From memory; not in the talk.
  • Dependence and the “platform tax”. App stores take a percentage; creator funds pay by formula; algorithmic opacity means workers guess at rules (Bishop on “algorithmic gossip”; Cotter on “playing the visibility game”). This is the micro-level of “constant adaptation”.
  • Opportunity is real. Poell does not drop the opening side. In India, Brazil and Nigeria, platforms let creators reach audiences without traditional gatekeepers. See idea 5 and idea 3.

📚 Roots: Cultural-industries studies — gatekeepers, fields, and creative work

The problem. What is special about cultural goods, and who decides what gets made and seen? Platforms are the latest answer to the old gatekeeping question.

The point. The book is a bridge: it takes “cultural industries” questions (who gets seen, who gets paid) and answers them with platform concepts. Poell’s judgement that early optimism was “naive” yet explains platforms’ “drawing power” is the field’s standard position now. His finding that big firms (Disney, News Corp, Bloomberg, NYT) are also the platform winners is the Bourdieu point: capital transfers across media.

The chain

  1. Adorno and Horkheimer, “The Culture Industry” (1944). Mass-produced culture standardises taste. A pessimistic view; the term “culture industry” starts the field.
  2. Paul Hirsch (1972), “Processing fads and fashions.” Cultural industries face unpredictable demand, so they use gatekeepers, many releases and “contact men” to cope. This is the model platforms upset: platforms are open at the front (anyone uploads) yet curate by algorithm.
  3. Pierre Bourdieu, The Field of Cultural Production (1993 English). Culture is a field with its own capital (economic and symbolic). Producers compete for position. This frames Poell’s point that the already well-resourced win.
  4. Richard Caves, Creative Industries (2000): the economics of art-business (“nobody knows” about demand, “motley crew” production). David Hesmondhalgh, The Cultural Industries (2002 onward): the best critical overview, which also argues that platforms and “creative industries” policy change the story.
  5. Creative-industries policy. Richard Florida’s “creative class” (2002) and UK policy from the late 1990s sold creative work as growth and glamour. Critics (Angela McRobbie, Rosalind Gill and Andy Pratt) showed the cost: self-exploitation, unpaid trials, insecurity.
  6. Platform-era creator studies. Brooke Erin Duffy, (Not) Getting Paid to Do What You Love (2017): “aspirational labour” of social-media creators. Stuart Cunningham and David Craig, Social Media Entertainment (2019). Burgess and Green, YouTube (2009).
  7. The “democratisation” books. The 2004–2010 literature Poell shows on a slide is, as I infer (he does not name it in the audio), the family of Henry Jenkins, Convergence Culture (2006), Yochai Benkler, The Wealth of Networks (2006), Clay Shirky, Here Comes Everybody (2008), and Burgess and Green. They said that open tools would shift power from big media to participants. (My inference; not stated in the talk.)

📚 Roots: A shorter thread: network effects and “winner-take-all”

  • W. Brian Arthur (1989): increasing returns and lock-in. The same product can win because of early luck, not quality.
  • Sherwin Rosen (1981), “The economics of superstars”: in markets where the best serves everyone at low cost, rewards pile up at the top. Digital media is an extreme case.
  • Chris Anderson’s “Long Tail” (2006) promised the opposite (niche markets thrive). Empirically, platforms mix both: a long tail of tiny earners and a fat head of winners. Poell’s data point sits on the head side.
  • Matthew effect (Merton, 1968): advantage compounds. Visibility → engagement → visibility, as Poell says.
  • Antitrust angle: Lina Khan, “Amazon’s Antitrust Paradox” (2017) argues that platform dominance needs new law. The EU’s DMA names “gatekeepers”. Poell’s concentration claim feeds that debate.

Where others push back.

  • Antitrust realists say network effects are overstated: Myspace, Vine and Google+ fell; dominance can be contested (TikTok displaced Instagram’s lead in short video). Poell’s own list includes TikTok as a winner of a new segment, which supports this: winners are segment-specific, not permanent.
  • Creator-economy optimists point to a growing number of full-time creators and new payment tools (subscriptions, tips), which reduce dependence on one platform.
  • Critics from the left say “precarity” is too gentle a word for exploitation: creator labour is unpaid or underpaid input to platform profit (Fuchs, Terranova’s “free labour”, 2000).

What he leaves out. The audience side: what winner-take-all does to taste and diversity. He raises diversity only in the EU answer. Also, he does not give numbers; the book does.

Check yourself

  • Explain in two sentences how network effects produce winner-take-all, and why the same logic applies inside a platform.
  • Why does he still credit the democratisation literature?
  • Which three things do platforms change that cause “constant adaptation”?

3. Provincialising theory — “global perspectives”

In short. The book’s framework rests on research in the Anglo-American world and Europe. Poell says the new project uses case studies to build theory, and that this should not fall only on scholars from Asia, Africa and Latin America. Scholars in Europe and the US should also stop assuming their theory is universal. The aim is to multiply frames of reference and make places like Chile, Brazil, Nigeria, India and Japan key reference points.

🎙 In the talk. The book’s framework “very much builds on research on the Anglo-American world and … Europe” and so “carries resonances of those institutions, cultural practices, histories”. His notions (platform, capitalism, governance, precarity, creativity, democracy) are very much tied to those regions. The new project develops “case studies as an opportunity for theory building” and says: “this shouldn’t just be the responsibility which so far it has mostly been” of “Asian, African and Latin American scholars”. It should also be taken up by “scholars who are working in Europe and the US”, because their work “often implicitly … assumes the universality of their theory”. He adds that scholars of other regions “have to make that effort … explain how their work is actually quite specific”, and “that work needs to be done by everyone”.

“Global perspectives” is “an effort really to multiply our frames of reference”. The team has specialists “in and speak the languages in Chile, Brazil, Nigeria, India and Japan”, which “have very vibrant cultural industries” and are integrating gen AI, “but are often not key reference points in our theoretical debates”. “The effort is really to develop them as key reference points”.

📚 Deeper theory

  • Three different demands hide in one phrase. (1) Coverage: add more places. (2) Reflexivity: say where your theory comes from. (3) Reconstruction: let other places change the concepts. Poell asks for all three but his stated method is mostly (1) and (2), then (3) through case studies. See Roots.
  • Why language matters. He stresses that team members “speak the languages”. Platform studies in English reads English-language sources; local debates (Japanese industry press, Brazilian labour studies, Nigerian “Nollywood” distribution, Indian regional-language creators) sit in other languages. Theory travels poorly if only English texts are read.
  • Why these five countries? Large cultural industries; different state–platform relations (Japan’s industrial policy; Brazil’s platform-labour scholarship; Nigeria’s informal film economy; India’s mobile-first creators; Chile’s regional lead). Poell’s reason is “large … vibrant” and “not key reference points”. The comparative design is not explained in the talk. Not in the talk: my reading.
  • Case study vs comparison. Case studies can build theory (Yin; “extended case method”, Burawoy) but do not by themselves show which differences come from culture, state, market size or timing. The risk is “everything is context-specific” with no general claim. His answer is to use situated cases to revise concepts like precarity.
  • “Provincialising” is not “relativising”. Chakrabarty argues Western categories remain usable, as long as they are shown to be local. Poell keeps his platform definition while asking it to be tested.

📚 Roots: Decolonial, postcolonial and “Southern” theory

The problem. Social theory is written mostly from a few countries, but speaks as if for everyone. What is lost, and how do you fix it without just adding “case studies” to a Western frame?

The chain

  1. Edward Said, Orientalism (1978). The West’s knowledge of the “East” served power.
  2. Dipesh Chakrabarty, Provincializing Europe (2000). European concepts (history, modernity, the citizen) are both indispensable and inadequate for the rest of the world. The task is to treat them as local. UvA’s wording “provincialize the US” for Poell’s project echoes this (UvA profile).
  3. Aníbal Quijano (coloniality of power, 1990s), Walter Mignolo, Boaventura de Sousa Santos (Epistemologies of the South, 2014). Colonialism continues in how knowledge is ranked.
  4. Raewyn Connell, Southern Theory (2007). Sociology’s “classics” come from a few Northern societies; Southern thinkers are cases, not theorists.
  5. James Curran and Myung-Jin Park (eds), De-Westernizing Media Studies (2000). The media-studies version. Poell’s “dewesternize” draws on this.
  6. Data and AI colonialism. Couldry and Mejias, The Costs of Connection (2019); Payal Arora, The Next Billion Users (2019) (counters the view that users in India, Brazil or Indonesia just copy the West); Paola Ricaurte (2019) on data epistemologies; Abeba Birhane (2020) on “algorithmic colonisation” of Africa.

What Poell adds. Two points. (a) The burden has fallen on Asian, African and Latin American scholars; scholars in Europe and the US should also do it. (b) Europe and US work often “implicitly … assumes the universality of their theory” and does not say how specific its observations are. A tension worth seeing: the method he chooses is case studies as theory building. Critics of “de-Westernising” (including some in the Global South) warn that adding cases can leave the frame in place. Poell says he wants to “rethink” notions, not just extend them.

Where others push back.

  • Who speaks? A Dutch-based team leading “global” research can still centre Northern funding and journals. Some Southern scholars ask for authorship and agenda-setting power, not inclusion. Poell partly answers by noting the earlier work is by scholars from those regions.
  • Exoticism risk. Choosing “non-Western” cases can treat them as exceptions to a Western norm.
  • Universalism defended. Some argue that platform logics (network effects, data extraction) are indeed global, so differences are of degree. Poell’s examples (precarity as norm; state promotion) support difference; his network-effects argument supports similarity. Both fit together only if the general mechanism is separated from local outcomes. He does not spell this out.

What he leaves out. Africa and Latin America appear mainly in a sentence about precarity (rural China, Ghana, Brazil); no case is given from Chile, Brazil, Nigeria or India. He says the research is at “the starting phase”. So the talk offers a programme with one worked case (Japan).

Check yourself

  • What is the difference between “adding cases” and “rethinking concepts”?
  • Why does he say Europeans and Americans must also do this work?
  • Which of his own book concepts gets tested first, and with what result?

4. Gen AI in anime and manga — the stage-by-stage case

In short. Japan’s anime and manga industry has been quick to adopt generative AI. Examples include a studio background tool, a manga made through Midjourney, AI voices and translation, and a series he says was 95% produced with AI. Gen AI enters at particular stages, such as colouring, while key animators keep their status. Poell says you must break the production process into stages to see what changes.

🎙 In the talk. Japan’s anime and manga industry “has been quite aggressive especially last couple of years in integrating generative AI”. Examples he gives:

  • 2021, Toei Animation: “one of Japan’s oldest animation studios used an AI-powered background image production tool … to produce this experimental movie” (the film he calls “Orvan”; URVAN, see warnings). It “builds on a longer tradition of integrating digital technologies” such as 3D modelling.
  • 2023 manga: “entirely generated through Midjourney”, and the author … admitted to have zero drawing talent … produced through prompts only.
  • Voice actors and translation: gen AI to simulate voice actors’ voices and to … translate anime into English and Chinese.
  • 2025 series: two high-school girls “trying to make it big on TikTok”, “95% produced through GenAI with comic artists coming in to do quality control and editing”.

Pattern: “gen AI comes in particular moments of the creation process but not everywhere … background coloring, character coloring”; “the key animation artists … are the creative heart of the industry and still have a very important status”. The Japanese government stimulates this, and the industry has “labor shortages”. But “anime manga artists are very concerned” about training on their content; he shows a petition.

Poell: “it’s quite important to analytically break down the culture production process in these various stages”: “greenlighting phase, content scripting, video editing, visual effects, audio enhancement, background coloring, translation, circulation, marketing”. He says to avoid “overall statements about the impact of genai”.

📚 Deeper theory

  • Why anime fits. Anime production is already divided and outsourced (key animation, in-between frames, backgrounds, colour finishing, often across studios and countries). Tasks that are routine and visual (in-betweens, backgrounds) are the easiest for image models. See the task approach (Roots). Key animation carries a studio’s style and status, and is the hardest to replace.
  • The “Midjourney manga” case. Cyberpunk: Peach John (Shinchosha, March 2023) took the author six weeks for 100+ pages against about a year for a skilled artist (AFP via Ahram). It shows entry (a writer becomes a comic maker) and displacement (the work an artist would be paid for). That is the same ambivalence Poell finds in platforms.
  • URVAN and Scenify. Toei’s PEROs team used Preferred Networks’ Scenify to turn photographs into base anime-style backgrounds, and artists then added the cyberpunk finish. Pre-processing time fell to about one sixth (Toei). This is a clean example of human-in-the-loop at a single stage, which is Poell’s pattern.
  • How “95% AI” should be read. Producers’ own materials say that over 95% of cuts used AI and that people finished by hand; critics say the number hides heavy human work and tool-mixing (Inquirer). A cut count is not a labour or value count. Treat any “X% AI” claim as a marketing figure until the unit is defined.
  • Why the petition matters. Training on a style needs examples of that style. A model that can produce “Ghibli-like” or “manga-like” output uses works from that tradition. After this talk, the Sora 2 episode (late 2025) showed the pattern at scale: Japanese rights-holders and the government demanded prior permission rather than after-the-fact opt-out (Automaton).
  • State and industry together. In Japan, “creative AI” sits in industrial policy (Cool Japan, labour-shortage framing). That differs from the US (courts) and the EU (regulation). See idea 7.

📚 Roots: Technology, tasks and labour — why “break it into stages” works

The problem. Poell insists that gen AI does “not necessarily affect all job categories in one sector equally”. Labour economics has a method for this.

The point. The anime case works because it has clear, separable stages (key animation, backgrounds, colouring, voice, translation). Gen AI has entered the bottom stages first, and the “creative heart” (key animators) is for now protected. That is the task-approach prediction.

  1. Autor, Levy and Murnane (2003): the task approach. Jobs are bundles of tasks; technology automates some tasks and not others. Acemoglu and Restrepo (2018–19) build a model: automation displaces labour from tasks, and new tasks reinstate it. What happens depends on which wins. Poell’s list of production stages is a task decomposition of cultural work.
  2. Polanyi’s paradox (Autor, 2014): we know more than we can tell, so tacit tasks resist rules. Gen AI weakens this barrier because it learns from examples rather than rules. That is why creative tasks, once thought safe, are exposed. It speaks to the student’s question.
  3. “GPTs are GPTs” (Eloundou et al., 2023): language-model exposure is highest for white-collar and creative-adjacent tasks. From memory.
  4. Skill-biased and “winner-take-most” effects. Brynjolfsson’s “Turing trap” (2022): AI built to imitate humans pushes workers toward substitution and concentrates gains in capital.
  5. The anime industry specifically. Anime has long had low pay and heavy outsourcing. Ian Condry, The Soul of Anime (2013) and Marc Steinberg, Anime’s Media Mix (2012) explain its production and “media mix” logic. That helps read why labour shortage plus state support makes AI attractive there. From memory.

Where others push back.

  • “It’s a tool, as always.” Some animators say AI background tools are no different from earlier digital tools; others say learning from living artists’ styles without consent is different in kind.
  • Quality and craft. Veteran artists argue that the “human touch” is in the line and timing, which gen AI misses (see dive 5).
  • Employment in the stages he names. Voice acting unions (in Japan and elsewhere) fight synthetic voice; translators warn that machine translation of anime lowers pay and quality. Poell lists these as at-risk.

What he leaves out. Audiences: do fans accept AI-heavy anime? (Reception of Twins Hinahima was negative, per Inquirer.) Pay data and the actual number of jobs. Also, the foreign market side: AI dubbing opens English and Chinese markets, which may shift power from licensors to platforms. He mentions the markets but not this effect.

Check yourself

  • Why does gen AI enter backgrounds and colouring first?
  • What does “95% of cuts” actually measure?
  • Which stage is “protected”, and why does that matter to the “stage by stage” claim?

5. Precarity, situated — and creativity, networked

In short. Poell says precarity is situated. Platforms may undermine job security in the US and Europe, but in places like rural China, Ghana and Brazil precarity is long-standing, and platforms can open doors for unlikely creative workers. On creativity, he asks where it is, not what it is: it is spread across people, tools and networks. So the human-or-AI question is a mix, not an either-or.

🎙 In the talk (precarity). “The use of Genai is obviously putting the jobs of background colorists, voice actors, translators to some extent also comic artists at risk. But at the same time … the position of these really well-known anime manga artists is insured”. The US and Europe research says platforms “undermine job security”, an argument also made in the book. But “researchers working on … rural China, Ghana, Brazil have pointed out that precarity in their region is not necessarily a deviation from the norm, but rather a constant and long-standing feature”. There “platforms do not necessarily intensify precarity but in certain cases enable marginalized individuals to become what some of our colleagues have called unlikely creative workers”. Hence: “precarity is really a situated phenomena”. Then the caution: “we cannot assume that one region … is the blueprint for other regions”.

🎙 In the talk (creativity). On the claim that “creativity is now taken over by generative AI models”: “the question is not necessarily what creativity is … but rather where it is”, something that’s quite distributed, networked. Seeing creativity as “a unique human cognitive trait really obscures the deeply material and technological and network nature of creative processes”. Film always involved “cameras, scripting, sets”. It is “an actor network perspective”. Context shapes outcomes: Japan has “a very high level of digital technology integration and a very rich ecosystem of AI startups, industry shows, expositions, workshops”, so creativity’s transformation “takes a particular shape” there.

🎙 In the Q&A. To Hana on the “human touch”: “it’s not an either/or question … it’s really often a combination”. The models are “trained on decades of human content” and “synthesize that creativity”; the integration is “in particular aspects”; the 95% series involves comic artists to … make sure that there is this human touch … quality control and … certain things are made more authentic. To the host’s hope that novelty will fade and the human touch will return: “I do think … creative workers will always be involved … but we shouldn’t think about it as one or the other”. CGI “in the 1980s 1990s … was a sort of a gimmick”, then “everywhere today … woven through, integrated in particular moments”. Gen AI will lose its gimmick aspect but “shift the way in which work is being done, what kind of work is being made available, who is profiting”.

📚 Deeper theory

  • Precarity as a variable. A useful split: job insecurity (can I be dropped?), income insecurity (is pay steady?), representation insecurity (do I have a voice?). A worker in a formal-sector studio and a freelance creator in an informal economy experience each differently. Poell’s point is that gen AI changes the mix in a given place. See Roots. (This breakdown is mine.)
  • “Unlikely creative workers”. The idea is that a platform can lower entry barriers for people outside the formal creative sector (rural users in China, low-income workers elsewhere). The same platform that creates a paid micro-career also gives it little security. The same ambivalence will appear for gen AI: a person with no drawing skill can make a manga (dive 4), but may be paid little because everyone can.
  • Distributed creativity is a method choice. ANT (Roots) asks us to trace who and what makes a difference. It is not a claim that all contributions are equal. Critics note it can flatten power: a Midjourney prompt and a key animator’s line are both “actants”. Poell’s practical use is narrower: it replaces the question “is AI creative?” with “what changes when AI is added at stage X?”
  • “Human touch” has two meanings. (a) A quality claim (AI work lacks timing, intent). (b) A value claim (people prefer works made by people). They can come apart: audiences may reject obviously AI work (b) even if the quality is fine, or accept it if (a) is met. Studies of “AI disclosure” find lower valuation of art labelled as AI-made. †
  • The CGI analogy: strong and limited. It supports “gimmick, then infrastructure”. But CGI mostly changed how effects were made; it did not generate a script or an image from a prompt trained on others’ work. The scope of task replacement and the training-data issue make gen AI different. Poell himself ends with distribution questions, “who is profiting”, which CGI also raised, but the data question is new.
  • Homogenisation. Doshi and Hauser (2024) (Roots) suggest outputs converge. Steyerl’s “mean images” idea (not in this talk) makes a similar point about averaged outputs. If creativity is networked and the network contains one dominant model, variety may shrink.

📚 Roots: Precarity and platform labour

The problem. The word precarity describes insecure work and insecure life. Is it a new condition produced by platforms, or an old one?

The point. Poell makes two moves. In the book: platforms increase precarity because of constant technical change. In the new project: precarity is situated, so the same technology can mean loss in one place and entry in another. This is a direct challenge to the idea that the Western experience of platform work is the template.

The chain

  1. Pierre Bourdieu (1998), “Job insecurity is everywhere now.” Early diagnosis of the spread of precarious work in Europe.
  2. Guy Standing, The Precariat (2011). A new class: unstable work, no occupational identity, few rights.
  3. Judith Butler, Precarious Life (2004) and Isabell Lorey, State of Insecurity (2015): precarity as a political and existential condition, not just a labour-market one.
  4. Gill and Pratt (2008) and Neilson and Rossiter (2008): precarity in cultural and creative work. Creative workers accept low pay for passion; project-based jobs; the “network sociality” of contacts. Gina Neff’s Venture Labor (2012) calls it “entrepreneurial” risk-taking.
  5. Platform and gig labour. Alex Rosenblat; Nick Srnicek; Mary Gray and Siddharth Suri, Ghost Work (2019), on hidden click-workers. Rafael Grohmann and the Brazilian platform-labour school.
  6. Precarity as the norm in the Global South. The informal economy was named by Keith Hart (1973); today the majority of workers worldwide have informal jobs (ILO). So when Poell says that in rural China, Ghana, Brazil … precarity … is not necessarily a deviation from the norm, but rather a constant and long-standing feature, he is echoing this literature. The “unlikely creative workers phrase is attributed by him only to colleagues”. A candidate is Lin and de Kloet (2019), “Platformization of the unlikely creative class: Kuaishou and Chinese digital cultural production” (Social Media + Society). †, and not named in the talk.*

📚 Roots: Distributed creativity — from the lone genius to the network

The problem. If a model generates an image, who is creative? Poell sidesteps this: “the question is not necessarily what creativity is … but rather where it is”.

The point. Poell uses distributed creativity for a practical reason, not just a philosophical one: it turns “AI versus human” into “which stage, which tool, which person”. That supports his insistence on stage-by-stage analysis and his answer to Hana.

The chain

  1. Romantic genius (18th–19th c.): creativity as a rare inner gift. This is the view that the “human touch” talk still assumes.
  2. Howard Becker, Art Worlds (1982). Art is collective action: support staff, makers of materials, distributors, critics. “Artists” are one node. This is a sociological form of Poell’s claim about film: cameras, scripting, sets.
  3. Mihaly Csikszentmihalyi (1988), the systems model of creativity. Creativity is a product of person, domain and field together; the field decides what counts.
  4. Bruno Latour and actor-network theory (ANT). Reassembling the Social (2005): non-humans (cameras, software) are actants; they make a difference to what happens. Poell calls his stance very much an actor network perspective. Antoine Hennion’s work on mediators in music is a cultural-sociology cousin.
  5. Margaret Boden, The Creative Mind (1990). Three kinds of creativity: combinational, exploratory, transformational. A classic framework used to ask whether AI is creative. A different question from Poell’s.
  6. Human–AI co-creation studies. Doshi and Hauser (2024, Science Advances): AI ideas raised individual stories’ rated creativity but made stories more similar to each other. † This shows distributed is not automatically good: the network can narrow output.
  7. Media-industries history. Special effects, sound, colour, CGI all arrived as gimmicks, then became background. Poell uses CGI as his example.

What it hides. Distributed credit can also distribute blame and pay. If creativity is “everywhere”, who is owed? This is where copyright (idea 7) and precarity (Roots) pull back in.

Where others push back.

  • Creativity does not live only in networks. Phenomenologists and many artists hold that intention and lived experience matter. The “where” question can be answered “in the person’s purpose”. See Dreyfus-style skill arguments on the Citton page.
  • Precarity is not “just situated”. Some labour scholars warn that “situated” can soften a general trend: gen AI raises the supply of content and lowers prices everywhere.
  • Optimists on human touch (the host’s view) vs pessimists (a flood of cheap content drives out niche human work regardless of preference).

What he leaves out. Evidence on pay (rates by stage), and who benefits among the “unlikely creative workers” once gen AI arrives. He says that how precarity plays out under gen AI “is really not yet clear”. Also, he doesn’t discuss consumers’ taste or willingness to pay for the “human” label, other than the host’s remark.

Check yourself

  • Why does Poell say precarity is “situated”? Give his example.
  • What does he change by asking “where is creativity” instead of “what is creativity”?
  • Where does the CGI analogy hold, and where might it fail?

6. Copyright, consent, licensing — and the click economy

In short. Poell expects large copyright holders to win, because they can make deals with AI firms. Much other content will still be used without a licence, and those owners will lose. Journalists and artists who make the content are often not consulted or paid. He compares this to YouTube’s copyright system, which paid uploaders but not remixers, and notes that synthetic content cannot be traced to an original.

🎙 In the talk. To the question on IP and consent in East Asia, Poell says “we’re very much at the starting phase of this research”, but in Japan “huge concerns around copyrights”; the petition; surveys “going up to the 70 80% of people working in the industry who are concerned”. His reasoning: “if you would like to develop a model that does a good job in … manga style … then you really need that content. … those concerns are very legitimate”. His forecast: “especially large media companies, so the large copyright holders, will be the winners in this. They will start making deals … but a lot of other content … will not actually be licensed, will still be used, and those copyright holders will lose out”.

The host adds the Microsoft publisher deal (see warning). Poell: “these large publishers … are able to make these deals and then those that actually produce the content … the journalists, artists … are not consulted and … not compensated either”. He links it to YouTube’s copyright identification system, which compensated those who had uploaded content, “but not necessarily those that were starting to remix that content”. That has been “an ongoing concern in the creator economy for … a decade”. He adds that models now make synthetic content that “can’t really be retraced back to any original”, which is “important for the training of new” models “because simply there isn’t enough data”.

On search summaries (host,): AI summaries mean you do not click through, “this leads to a higher level of precarity, doesn’t it?” Poell: “Absolutely.” The platform model connected “billions of end users and millions of content producers” and took people to the producer’s site, where it could earn “traffic and subscribers … or advertisement”. That is “being shaken up”. News organisations are “protesting”, and the way forward is “to prohibit these tech companies to use their content to train their models”. News is “always new”, so AI firms “would really need a license”. Strategy for outlets: “find their niche”, local and original content, a subscriber base “rather than … platform advertising”. “The challenge is really to find a unique niche and to protect your content”.

📚 Deeper theory

  • Why the winners are the big holders. Licensing needs a counterparty with a catalogue, lawyers and standing to sue. A collective of freelancers has none of these. This is the transaction-cost problem. Collective management bodies (music’s PROs, press-publishers’ rights in the EU’s 2019 Article 15) are the standard answer; gen AI is testing them. Poell’s prediction is the platform lesson (resources decide) applied to licensing.
  • YouTube Content ID as precedent. Content ID (2007) matches uploads against rights-holders’ reference files and lets them block, track or monetise. It paid registered owners. It did not pay remixers or most small creators. The structure repeats: pay those who can register and match.
  • Consent vs compensation vs control. Three different demands. Consent (opt-in before training), compensation (pay after), control (stop or remove). CODA’s 2025 demand for prior permission is about consent [Part 1]. The Anthropic settlement is compensation for piracy. The EU opt-out regime is a weak form of control. Mixing them causes confusion in debates.
  • Does training copy? Three legal views: (a) training is a transformative use or an analysis exception (US fair use arguments; Japan Art. 30-4); (b) training copies works without a licence and infringes; (c) infringement depends on outputs (memorisation, close style imitation). Courts are split by facts; no global rule. See Roots.
  • The click economy and “zero-click” search. Search used to send traffic to publishers in return for indexing: a bargain. AI answers keep the user on the platform. Publishers have reported falling referral traffic since AI overviews launched. † The bargain breaks, so deals and suits follow. This is Poell’s “restructuring of the online economy”.
  • Why news has bargaining power. Fresh data cannot be got from older training sets. Retrieval systems and live search need current articles. That is a real lever, as Poell says, and why news deals (Le Monde, Axel Springer, FT, News Corp) came first. Local news, though, has little leverage with AI firms, since its audiences are small.
  • Synthetic data and “model collapse”. If models train on model output, errors and sameness can accumulate (Shumailov et al.). It also weakens traceability: a lawsuit needs a link from output to original work; synthetic layers cut it. That makes enforcement harder, as Poell notes.

📚 Roots: Generative AI — how it works, who pays, what comes next

The problem. Poell’s remarks on gen AI are brief but depend on facts about how models are built and fed. Four ideas matter.

The point. Poell’s claim that big holders win and others lose is a platform-markets prediction applied to AI licensing. His point that news is always new, so AI firms “need a licence”, is the idea that fresh data has bargaining power.

  1. Why the “deep-learning paradigm” took over. Neural networks (perceptron, 1958; backpropagation, 1986) lost out to symbolic AI for decades. Three things changed them: huge labelled and unlabelled data, parallel hardware (GPUs), and architectures that scale (the 2017 transformer). Richard Sutton’s “The Bitter Lesson” (2019): general methods that use more compute and data beat hand-built knowledge. Kaplan et al. (2020): loss falls smoothly with model, data and compute size (“scaling laws”). Poell’s account of why creative work is hit: deep learning models … trained on a very large scale … in parallel … increase in compute availability. He is giving the standard history.
  2. Training data is the raw material, and it is cultural. Bender, Gebru et al. (2021), “Stochastic Parrots” pointed at what goes into web-scale data. Kate Crawford, Atlas of AI (2021): AI as extraction of land, labour and data. This is Poell’s argument: to get a model that “does a good job in a particular manga style … you really need that content”. Style is learned from specific people’s work.
  3. Running out of data; synthetic data. Villalobos et al. (Epoch AI, 2022/2024) estimated that the stock of public human text could be used up within a few years to about a decade. Shumailov et al. (2023/2024, “model collapse”): training repeatedly on model output degrades models. This is the “complicated” situation Poell flags: new models train on synthetic content that “can’t really be retraced back to original content”.
  4. Invisible labour behind the model. Gray and Suri (2019); reporting on Kenyan data-labelling workers (TIME, 2023); Muldoon et al., Feeding the Machine (2024). Poell does not mention this, but it belongs to his own “where is creativity” question: the network includes labellers.

Copyright and AI law, in short (from memory unless linked).

  • Japan: Article 30-4 of the Copyright Act (2018 amendment) allows use of works for “information analysis” where the aim is not to enjoy the work, so training is broadly permitted, with limits where it harms rights-holders. CODA now argues for prior permission [see Part 1]. Japan’s AI Promotion Act (May 2025) is soft-law, pro-innovation. (After this talk.)
  • EU: The text-and-data-mining exception in the 2019 DSM Directive lets rights-holders opt out. The AI Act (2024) requires general-purpose model providers to have a copyright policy and publish a training-data summary.
  • US: Case law is unsettled. In 2025 judges in Bartz v. Anthropic and Kadrey v. Meta found training on lawfully obtained books likely fair use in the facts before them, but piracy remained actionable. This produced the $1.5 billion settlement (Bloomberg Law). Details from memory; not re-checked.
  • China: Interim rules on generative AI services (Aug 2023) set content and training-data duties, within a state-led approach.

Where others push back.

  • AI firms: training is analysis, not copying for expression; blocking it would stifle innovation; opt-outs and licences are enough.
  • Creators’ groups: licences negotiated by publishers do not reach the authors; the 2024–25 author protests followed that. (See the opt-in detail in the HarperCollins deal; the host’s version was not accurate.)
  • Economists: blanket licensing could entrench incumbents (big AI firms can pay; start-ups cannot). So Poell’s “big holders win” has a twin: big AI firms win too.

What he leaves out. The individual remedy options (collective licensing, statutory pay, “data trusts”). The Global South angle of licensing (where few rights-holders can negotiate), which is awkward for his own project. He says this is “something we will see unfold over the next couple of years”.

Check yourself

  • Why does Poell predict big holders will win and small ones lose? What does YouTube’s Content ID show?
  • What is the difference between consent, compensation and control?
  • Why does “news is always new” give news outlets bargaining power?

7. Governance and the cause — states, “technocrats”, and the EU “war”

In short. Governments differ. Japan’s state works with industry to promote AI in culture, while the EU is more restrictive, with the AI Act. Poell says the harm to creative work is not due to evil technocrats but to the deep learning paradigm, driven by large-scale training and huge investment. He says states must step in to defend public values, and expects a hard fight with the large US tech firms.

🎙 In the talk (governance). In Japan “the state in close association with industry actors is actually very much promoting the integration of AI in cultural production”. In contrast, “the European Union … has been much more restrictive in terms of its approach to platforms and now more recently also to AI, the AI Act is a good example”. East Asia (Japan, and China “an even stronger example”) shows “very active intervention and governance but also promotion”. In China “platforms … have come to really resemble media companies much more strongly than our definition of what a platform is”. So “it’s really quite important to multiply our frames of reference” to understand how governance works.

🎙 In the Q&A (cause). A student asks if creative jobs are hit because of technocrats behind AI platforms. Poell: “I don’t think … particular bad or evil actors are to be blamed. … It is really the result of what people … call the deep learning paradigm”, enabled by “very large scale” training “in parallel” and an increase in compute. That paradigm “trains on loads of cultural content data” and then “starts replicating that”, so it affects “those aspects of the world where that data is coming from”, which are culture and also academia (students using models for essays). There are billions and billions of dollars being invested … but it is something that is more fundamental … restructuring entire economies than just a couple of evil technocrats.

🎙 In the Q&A (EU). On the AI Act and the DSA: “not perfect … but … absolutely essential that states intervene … in ways that correspond with basic human rights, public values”. He names equality, diversity, inclusion … because they’re so under threat at the moment in the United States. Otherwise large tech firms “are going to undermine these values”, and “we’ll see winner take all dynamics again”. “I think we have to start preparing for a war in that regard”: “huge pressure from the large US tech companies to make these laws either go away or at least not being enforced, and the US government is going to assist them … Europe is going to have to confront that and that’s not going to be easy … the coming years which are going to be quite dark”.

📚 Deeper theory

  • Structure vs agency in explaining AI. Poell gives a structural account: a technical paradigm plus capital flows. The “technocrat” view is agential: a few firms and founders steer. Both can hold: the paradigm made the products possible, but the choices to scrape, to release image generators and to price below the cost of human work are firm decisions. Scholars like Meredith Whittaker (“The steep cost of capture”, 2021) stress that deep learning’s dependence on data and compute concentrated in a few firms is itself a power structure. Poell’s “billions of dollars being invested” line gestures at that.
  • Why creative work was “surprisingly” hit. Earlier predictions (Frey and Osborne, 2013) expected routine and manual work to go first and creative work last. Generative models reversed it because cultural data is plentiful, digital and tolerant of error (an image can be “good enough”). A surgeon’s task has little data and high stakes. See idea 4.
  • Bradford’s three models (Roots). Poell’s East Asia / EU contrast adds a fourth feature: promotion. The EU regulates but also hopes for AI “sovereignty”; Japan promotes and also now pressures OpenAI (Part 1). So “restrict vs promote” is not a clean split.
  • Public values. In The Platform Society, van Dijck, Poell and de Waal argue that societies must defend values like privacy, fairness, accuracy, democratic control, and that platform ecosystems tend to shift these. Poell’s answer in the Q&A is that frame in plain words.
  • What happened next. The EU moved in Nov 2025 to delay and simplify parts of the AI Act; a provisional deal now sets high-risk rules from December 2027 (To Vima). Supporters call it competitiveness; critics call it capitulation. Poell’s forecast of a “war” is contested in cause but not in direction. The facts after early 2025 are not in the talk.
  • Political stance, to note honestly. His closing is normative: he defends a “liberal social democratic” tradition. This is a stated position, not a finding. A reader from another tradition (market-liberal, or Southern and sceptical of European “universal values”) could weigh it differently. Interesting tension: his own project asks us to doubt European universals, yet his EU answer defends European values. He frames it as speaking “from a European perspective”, which fits his reflexivity rule.

📚 Roots: Governance — three regulatory models

The problem. “Governance” in the book means platforms governing users (rules, APIs). In the closing section it means states governing platforms and AI. The two meet.

  1. Platform governance. Platforms govern through terms of service, content rules, moderation, ranking. Gillespie, Custodians of the Internet (2018) is the standard text. Poell and van Dijck call platforms “governance frameworks”; Apple’s app review and SDK are a clean case.
  2. Three models of state response. Anu Bradford, Digital Empires (2023): a US market-driven model, a Chinese state-driven model, and an EU rights-driven model. This fits Poell’s contrast: the EU “much more restrictive”, Japan and China very active intervention but also promotion.
  3. The EU toolset. DSA and DMA (2022) for platforms; GDPR (2016); AI Act (2024). Bradford’s earlier The Brussels Effect (2020): EU rules spread globally because firms comply once. Poell’s argument for the EU laws rests on public values (equality, diversity, inclusion). This is the van Dijck–Poell–de Waal frame.
  4. East Asia. Japan uses “soft law” and industrial policy; Cool Japan promotes content exports; the state backs AI as a fix for labour shortage. China combines promotion with censorship and licensing; its platforms (Douyin, Kuaishou, Bilibili) mix user content with commissioning. This is why they look like “media companies”.
  5. Politics of the “war”. The 2025 US administration opposed EU digital rules as trade barriers, and the EU then moved to simplify them (see Part 1). Whether this was caused by US pressure or by EU competitiveness worries is disputed. Poell’s forecast is his own and made before the omnibus; the later facts are mine.

Where others push back.

  • Innovation-first view: EU rules slow Europe and entrench US and Chinese leaders (the Draghi report, 2024, urged simplification).
  • Rights-first critics: the AI Act has loopholes (national security, “high-risk” lists), and the copyright duties are thin; weakening it makes the problem worse.
  • Global South view: EU rules may become de facto global (Brussels effect) without Southern input. That is the project’s own concern, applied to law.

What he leaves out. Concrete enforcement; the role of courts; India, Brazil or African regulators (despite his case list); labour law and collective bargaining (e.g. the 2023 Hollywood writers’ and actors’ strikes won AI clauses). He also does not discuss open-source models, which change who can build and who can regulate.

Check yourself

  • How does Poell explain why creative work is hit? Is that a structural or an agential explanation?
  • How do Japan and the EU differ in governance, in his account? Why does China complicate the platform definition?
  • What does he mean by “preparing for a war”?