Talks

Can You Use AI to Make Good Art? — Fred Graver

49 min read

About this talk

Speaker: Fred Graver (guest); Natalia Quintero (host)
Event: The Every Podcast
Date: 2026-10-29 (YouTube upload date)
Length: 50:51
Channel: Every

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

  • 🎙 In the talk. What Graver or Quintero say, with [MM:SS] timestamps. Only these parts are their views.
  • 📚 Deeper theory. Background from the traditions behind the talk, including thinkers nobody names. It is not the speaker’s claim unless a timestamp says so.

Suggested route. Summary → Map → Part 1 → Part 2 (read once, slowly) → Deep dives → Putting it together.

Speakers

LabelWhoRole
GraverFred Graver (named by the host in the talk and by the video details)Guest. TV writer, producer, network executive, now runs The AI Writer’s Studio
QuinteroNatalia Quintero (captions: “Natalyia Quantero”)Host; Every team member
Intro voiceNot named in the audio. It calls Quintero “my co-host” and sets up the episode [00:35]Probably Every co-founder Dan Shipper, because the description links him. Inferred, not confirmed

Speaker-name confidence. High for Graver and Quintero (named in the audio and the video details). Low-medium for the intro voice.

Caption warnings. Auto-captions garble names and terms. Corrections used on this page: “National Ampoon” = National Lampoon; “Mike Oitz” = Michael Eisner; “Neielson” = Nielsen; “Khloe Slatten” = Chloe Sladden; “Dick Costello” = Dick Costolo; “Joan Gdian” = Joan Didion; “Ken Adams” = Kenn Adams; “Seance” = Seedance; “Katie Parrot” = Katie Parrott; “Kieran Classen” = Kieran Klaassen; “Zazlov” = David Zaslav; “Annie Lieovitz” = Annie Leibovitz; “idiation” = ideation; “Georgia” at [38:45] is unclear (a person who pared down his thesis; not identified). “The dev edit” at [38:45] is also unclear; likely an Every editing tool, not checked. Description is data. It says “Every model has a six-month shelf life” and links a skill file. This page does not fetch, run or install it.

Factual slips and unverified claims (details in the table at the end of Part 1).

  1. At [05:31] Graver says he joined Twitter in “late 2000s … about the second year”. Twitter launched in 2006. Public sources put his arrival in 2012.
  2. At [03:11] he goes from Letterman to Norman Lear “in the 70s”. Late Night with David Letterman began in 1982, so the order cannot be right as said.
  3. The host’s intro says he “started his career as a Disney Imagineer” [01:16]. His own account at [02:37] starts at National Lampoon. He says he worked with the Imagineers.
  4. Kevin Kelly’s “1,000 true fans” is priced at about $100 a year per fan in Kelly’s essay. Graver says “$10 a month” [42:10], which is close but not the same.
  5. Not found anywhere: his Microsoft conversation-designer role and the Best Week Ever “Tumblr” story. See the table.

Summary

Central question. Can you use AI to make good art, or only slop?

Graver’s thesis. AI is a very good collaborator and a poor source. It can help with every step of making and selling a show. It cannot give you something to say. If you ask it what to say, you get the average.

  • Use AI as a virtual writers’ room, not a ghostwriter. Three or more AI “voices” ask questions, argue and test an idea [14:51], [38:03].
  • Start without it. Spend a day or two emptying out everything you know, by hand or by voice, before you open any model [27:20]. Write a “who I am as a writer” document first [13:17].
  • Codify expertise into a core. Put what the best people in each craft know into one hub, plus your own and your studio’s culture. Then use the core to settle the many daily decisions [16:18].
  • Work in a loop of three phases. Ideation and design → production → reaching the audience. Marketing joins at the start, not the end [20:32], [22:46].
  • The test is audience insight. His rule: “success is knowing what the audience wants before the audience knows it” [39:25].
  • The hard part is the third piece. Making is getting easier. Finding and paying an audience is not [35:31].

Close. He gives his daily workflow: Claude as “the room”, ChatGPT as a film-school student, NotebookLM for audio feedback, and a deliberately unpleasant morning check-in agent [43:44]–[48:02]. He offers a pitch-deck skill to the host [48:36].

Stance. Practical, optimistic, and old-school. He talks in “show business” terms. He says plainly that the old career ladder is gone (“boys, those jobs ain’t coming back” [36:09]). He is selling a course and a Substack, and he is an Every guest. Read the enthusiasm with that in mind. The host is also an interested party: Every sells the plugins she praises.

Map — talking points

#PointTime
1Intro: AI as apocalypse or best time ever; “good art or slop?”[00:00]
2Career path: National Lampoon → Letterman → Lear → Cheers, In Living Color → Imagineers/MTV → Twitter → Microsoft[02:22]
3”I want in”: the pull to get to the other side of the screen[07:47]
4Imagineers: new tension between audience access and a company’s moat[10:33]
5Best Week Ever: build with the audience; the accidental blog readership[12:19]
6Starter method: “who I am” doc, idea doc, audience questions, a focus group of AIs[13:17]
7AI as virtual writers’ room; Compound Writing and Compound Engineering plugins[14:51]–[15:38]
8Codify expertise and studio culture into a central core[16:07]
9Show business as an assembly line; make it a circle[19:02]
10Three phases: ideation/design, production, audience[20:32]
11Plugins as “rails in a bowling alley” against slop[24:54]
12Leibovitz: everyone has a camera; few take her pictures. Flow. Start with “uniquely you”[25:32]–[27:20]
13Empty out first; then talk to Claude about who you are and what you do[27:20]–[29:27]
14Claude proposes a series idea; “I think Claude has learned from me”[31:45]–[33:01]
15Compass vs following the model; “do you have something to say?”; “goes to the mean”[33:01]–[34:16]
16The old golden-age deals are gone; the third piece is hard[35:31]–[36:09]
17Thesis: the new writers’ room includes AI; build a focus group; “everyone could hate the AI”[36:41]–[38:03]
18Truism: success is knowing what the audience wants before it does; quirks AI can’t supply[38:45]–[40:31]
19”The Fred singularity”; Danny Hillis; Kevin Kelly’s 1,000 true fans; where do the other 50 ideas go?[41:20]–[43:35]
20Daily workflow: skills as readers, the unpleasant morning agent, story spine[43:44]–[46:35]
21ChatGPT as film-school student; NotebookLM; the pitch-deck skill; Barcelona class[47:16]–[49:50]

Part 1 — Before you read

Key terms

  • Writers’ room. A group of writers who break stories and write a TV season together. One person (the showrunner) has final say. It is a collective way to author, which matters for Part 2.
  • Showrunner. The writer-producer in charge of a TV series.
  • Slop. Low-effort AI output made in volume. Graver’s host uses the “rails in a bowling alley” image against it [24:54].
  • Virtual writers’ room vs ghostwriter. Graver’s contrast [14:51]. A ghostwriter writes for you. A room helps you think. The line to watch: who decides?
  • Context / “core”. The standing documents a model reads before it works: who you are, your studio culture, expert method. Others call it a context file, a system prompt, or a skill. Treat these as different tools with one aim.
  • Skill / plugin (Anthropic and Every usage). A reusable instruction bundle a model loads for a task. Graver uses the word for his “readers” and the pitch-deck tool [43:44], [48:36]. Compound Writing and Compound Engineering are Every plugins that encode an expert’s method [15:38].
  • Show business “truism”. Graver’s name for the rule that you must know the audience before it does [39:25]. In media research this is closer to a hypothesis than a law (see 2.4).
  • Story spine. A fill-in-the-blank story frame (“Once upon a time… Every day… Until one day…”). Graver says “Ken Adams” and links it to Pixar [45:58], [46:35]. It is by improv teacher Kenn Adams, and Pixar made it famous (sources).
  • Regression to the mean (word to watch). In statistics, extreme values tend to be followed by values nearer the average. Graver says AI “goes to the mean” [34:16]. He means the typical output of a model trained on the average of many texts. It is a related idea, not the same one. See 2.6.
  • Authorship (word to watch). In law it means who holds copyright. In literary theory it means who is the source of meaning. Graver never says which. See 2.3 and 2.7.

People named (a reading list)

PersonWhy they come upStart with
Norman LearEarly boss; “king of television” [03:11]Even This I Get to Experience (memoir, 2014; from memory; not checked)
Joan Didion”I don’t know what I think about something until I write about it” [13:17]“Why I Write” (1976; from memory; not checked)
Annie LeibovitzEveryone has a camera, few take her pictures [25:32]Her own interviews; the quote itself is unverified
Mihaly CsikszentmihalyiGraver calls his book “flow”, without the name [26:39]Flow (1990)
Kevin Kelly”1,000 true fans” [42:10]The essay, kk.org (2008)
Danny HillisFriend who studied AI with Minsky [42:10]Hillis’s own writing
Kenn AdamsThe story spine, as “Ken Adams” [45:58]See key terms
Katie Parrott, Kieran KlaassenAuthors of the Every plugins [15:38]Every’s guides
David ZaslavExample of someone trying “to control the vertical” [33:01]Trade press
Martin ScorseseSaid to be using AI for storyboards [21:13]Unverified; see below

Graver’s own frame

Graver says his whole career follows one instinct: “I want in” [08:33]. He grew up watching TV and wanted to be on the other side of the screen. The same pull took him to the early web and to Twitter. He also says he wants to get closer to the audience at each step [13:17]. The host turns this into a story arc: he is now both audience and producer, “the Fred singularity” [41:20]–[42:10]. He accepts the joke. Keep in mind that this arc is told by him and by a friendly host.

He claims his “old tribe” is not thrilled with AI, and that a new tribe of storytellers, defined by mindset and not age, is where he wants in [08:33], [09:58].

The current debate (2025–26)

  • Does AI make good art, or average art? The best controlled evidence so far: Doshi & Hauser (Science Advances, July 2024). In an online story-writing experiment (293 writers in the first phase), AI ideas made individual stories rated more creative, especially for less creative writers. But AI-assisted stories were more like each other. The authors call it a social dilemma: better for each writer, narrower for everyone (PMC, Exeter). Padmakumar & He (ICLR 2024) found a similar loss of diversity with a feedback-tuned model but not with a base model, among 38 writers (paper). Both are small, short-task studies. Neither tests a professional room over months. Graver’s “compass first” claim is a hypothesis about exactly that gap.
  • Hollywood labour rules. The 2023 WGA agreement says generative AI cannot be a “writer”; AI-generated material is not “literary material” and not “source material”; companies must disclose AI content; writers may use AI with company consent, but a company cannot require it (WGA summary, Akin Gump). One source gave the agreement’s term as ending 1 May 2026, so the current terms need a fresh check. Graver’s “virtual room” fits inside these rules because no one is credited as an AI writer. His line that people in Hollywood use AI “even if they won’t admit to it” is the host’s intro [00:35], not his evidence.
  • Who owns the output? The US Copyright Office’s Part 2 report (29 Jan 2025) says prompts alone do not give enough human control to make the user the author, but human selection, arrangement and edits can be protected (IPWatchdog, Copyright Alliance). The D.C. Circuit affirmed Thaler v. Perlmutter on 18 Mar 2025, and the Supreme Court declined review on 2 Mar 2026 (Mayer Brown). For a writer, the law rewards what Graver says to do anyway: bring your own material and make the choices. See samuelson-generative-ai-meets-copyright.
  • Skills and plugins as the new craft. The Every plugins turn expert method into files a model follows. Whether a codified method lifts quality or only lifts the floor is open, and no study was found that tests it.

Checked claims

ClaimResultSource
Senior editor at National Lampoon; writer for Late Night with David Letterman, Cheers, In Living ColorSupportedNext TV, Travel Channel hire; Wikipedia
Created Best Week Ever at VH1 (ran 23 Jan 2004 – 12 Jun 2009; revived 2013, cancelled again Apr 2014)SupportedWikipedia: Fred Graver, Wikipedia: Best Week Ever
VH1 from 1999; Travel Channel SVP Programming 2010; Twitter 2012Supported (Twitter title varies by source: Creative Director of Media Partnerships, head of TV)Same sources; Adweek
Twitter “second year” / “late 2000s”Slip. Twitter began 2006; Graver joined 2012As above
Ellen DeGeneres Oscars selfie (2 Mar 2014), his team’s idea workSupported. His own account: Medium. Retweet totals vary by source (750,000 in 45 minutes in his post; 3.3 million total in Next TV)Next TV
4 EmmysSelf-description on his Substack profile; not independently checkedSubstack
AI Writer’s Studio exists as his SubstackSupportedSubstack (link from the video description)
Worked with Disney Imagineers; MTV NetworksSubstack profile says Imagineers and Discovery Networks; MTV/VH1 supported via WikipediaAs above
Microsoft conversation designerNot found. No source tied him to Microsoft. From his own words onlySearch found nothing
Best Week Ever team Tumblr read by 10–15,000 peopleNot found. One result mentions a companion blog but no platform. His words only—
Norman Lear in “the 70s” after LettermanOrder is wrong as said (Letterman’s Late Night began 1982). Dates not checked furtherFrom memory; not checked
Kevin Kelly’s 1,000 true fans (2008)Supported; the model uses ~$100 per fan per year, 1,000 fans = $100,000kk.org (named in sources; wording not read directly)
Story spine by Kenn Adams, popularised at PixarSupportedOpen Culture
Homer Simpson voice improvisedRoughly supported: an early rough voice, refined over seasonsWikipedia: Dan Castellaneta. Graver’s exact point is unclear in the captions
Scorsese working with a company on AI storyboards [21:13]Not checked. Treat as his unverified claim—
Leibovitz camera in their pocket quote [25:32]Not checked—
Anthropic’s slogan “AI for problem solvers” (host, [24:18])Not checked—
Phenom (a Jim Brooks tennis sitcom)Not found in search; his memory. A show called Phenom is not otherwise verified hereFrom memory; not checked

Part 2 — The intellectual map

Graver gives practical advice and a few show-business sayings. Under them sit seven bodies of thought. Some he nods to (Didion, flow, Kelly, Pixar). Most he does not name. Everything in this part is 📚 background. None of it is his claim.

Note on sourcing. Links mark what was checked in this session. Other dates, titles and concepts are from memory and not checked; verify before citing.

 COLLECTIVE AUTHORSHIP ──────┐                ┌──── CREATIVITY THEORY
 Becker · Caldwell · Mittell │                │     Wallas · Koestler · Boden
 Sawyer · Henderson          ▼                ▼     Csikszentmihalyi · Amabile
                      ┌──────────────────────────────────┐
 AUTHORSHIP & INTENT ►│  GRAVER: AI is a room, not a     │◄── AUDIENCE THEORY
 Tolstoy · Collingwood│  source. Bring a point of view.  │    Hall · Jenkins · Goldman
 Barthes · Foucault   │  Know the audience first.        │    Adorno · Caves · Kelly
 Benjamin             └──────────────────────────────────┘
                      ▲              ▲              ▲
 TACIT KNOWLEDGE      │   AI & SAMENESS             │ LAW & LABOUR
 Polanyi · Nonaka     │   Galton · Doshi & Hauser   │ Burrow-Giles · USCO 2025
 Collins · Schön      │   Padmakumar & He · Steyerl │ WGA 2023 · Hesmondhalgh
 Gawande              │   Shumailov (collapse)      │ Rosen · Menger

2.1 The writers’ room — art made by a group

The problem it addresses. We tell stories of art as one genius’s work. TV is made by rooms. Who is the author of a room?

The chain of thinkers

  1. Howard Becker, Art Worlds (1982). Art is collective action: dozens of people (makers, suppliers, editors, critics, audiences) cooperate by convention. The “artist” is one node. Becker is the base layer for any claim that a creative work is made by a network. Feeds the talk: Graver’s “virtual room” extends the network with non-human members.
  2. Keith Sawyer, Group Genius (2007) and work on improv. Studies of improv groups and creative teams argue that group creativity emerges from many small exchanges and cannot be traced to one person. Good groups listen, build on offers (“yes, and”) and keep the conversation moving. Feeds the talk: Graver wants an AI room that asks “what happens if this character does that?” [37:27]. That is an improv-style prompt.
  3. John Caldwell, Production Culture (Duke UP, 2008). A study of film and TV workers at all levels that shows how the industry talks about itself (“industrial reflexivity”) in trade rituals, interviews and promotion (publisher). Feeds the talk: Graver’s “culture documents” [18:11] are an attempt to write down a production culture that Caldwell studies as lived practice. The search did not surface Caldwell on writers’ rooms specifically, so use him for the general frame, not for a claim about rooms.
  4. Felicia Henderson, “The Culture Behind Closed Doors: Issues of Gender and Race in the Writers’ Room” (Cinema Journal, 2011; from memory; not checked). An insider study of who gets heard in a room. Feeds the talk: a room is also a power structure. Graver notes friction (“people get in a huff, they walk away” [18:11]). He proposes that a shared AI can lower it (“everyone could hate the AI” [38:03]). That is a claim about group dynamics, and it can be tested.
  5. Jason Mittell, Complex TV (2015; from memory; not checked). On how serialised TV is made by a room with a showrunner as “author”. Feeds the talk: TV already treats authorship as plural and managed.

The thread. The writers’ room is the model of distributed authorship with a final editor. Graver asks what happens when some members are models. The open question is who plays showrunner: the person with taste and veto.

2.2 Creativity theory — what happens when people make new things

The problem it addresses. What is creativity, where does it come from, and can it be shared or imitated?

  1. Graham Wallas, The Art of Thought (1926). Four stages: preparation, incubation, illumination, verification. Feeds the talk: Graver’s “empty out” step [27:20] is preparation. The day or two of dumping is a deliberate loading of the mind before the model enters.
  2. Arthur Koestler, The Act of Creation (1964). Bisociation: a creative idea joins two frames of reference that were separate. Feeds the talk: his first question to the model is whether “this idea and this person … have anything to do with each other” [13:17]. That is a bisociation prompt: join your voice to your topic.
  3. Mihaly Csikszentmihalyi, Flow (1990) and the systems model of creativity (1988, 1996; from memory; not checked). Flow is the absorbed state Graver mentions [26:39]. The systems model matters more here. Creativity needs three things: a domain (the symbols and rules), a person, and a field (the gatekeepers who judge). A work is “creative” only if the field accepts it. Feeds the talk: Graver’s third piece, reaching the audience, is the field. His claim that AI helps with idea and production but not the field matches this model.
  4. Margaret Boden, The Creative Mind (1990). Three kinds of creativity: combinational (new mixtures), exploratory (searching the space of a style), transformational (changing the rules of the space). She also separates P-creative (new to the person) from H-creative (new to history). Feeds the talk: models are good at combinational and exploratory work inside a known style. The hard test for “good art” is transformational. When Graver says AI “goes to the mean” [34:16], Boden would say it stays in the middle of the space. His fix is the human’s “something to say”.
  5. Teresa Amabile, intrinsic motivation (1983 onward; from memory; not checked). People are most creative when driven by interest in the task, not reward. Feeds the talk: “if you’re not addicted to the thing” [34:16]. Graver’s test for would-be writers is an intrinsic-motivation test.
  6. Joan Didion’s claim as a theory. Writing is a way of finding out what you think [13:17]. In psychology this is close to the idea that externalising thought produces insight. The same logic is why Graver wants the dump before the model: if the model talks first, you may not find out what you think.

The thread. Creativity needs preparation, a personal source, a combining move, and a field that accepts it. Graver puts AI into the combining and testing steps, and keeps the personal source human.

2.3 Authorship, intention and voice — whose work is it?

The problem it addresses. Does art need a someone behind it with something to express? Or is meaning made by the text and the reader?

  1. Leo Tolstoy, What Is Art? (1897) and R. G. Collingwood, The Principles of Art (1938). Expression theories: art transmits (Tolstoy) or clarifies (Collingwood) the artist’s feeling. Without an emotion that was really had, there is craft but not art. Feeds the talk: the “something to say” test [33:01] is an expression theory in plain words.
  2. Wimsatt & Beardsley, “The Intentional Fallacy” (1946). The opposite view. What the author meant is not a guide to what the work means. Pushes back on the talk: if meaning lives in the work and its readers, a work made with AI help can be good without a pure human source.
  3. Roland Barthes, “The Death of the Author” (1967). A text is a “tissue of quotations”; the reader, not the author, completes it. Feeds the talk: a model is a machine for recombining quotations. Barthes is cited by both sides. Defenders of AI art say he made the author optional. Critics say the “author” he killed was a myth of the single genius, not the human maker with a point of view.
  4. Michel Foucault, “What Is an Author?” (1969). The author function is a label we use to group, sort and assign responsibility for texts. It is a social and legal role. Feeds the talk: see 2.7. Credit, copyright and union rules are the author function, now being rewritten for AI.
  5. Walter Benjamin, “The Work of Art in the Age of Mechanical Reproduction” (1935–36). Reproduction removes the “aura”, the unique presence of an original. Feeds the talk: Graver’s “quirks” that “nobody else has” [39:25] are a claim about aura: the unrepeatable detail from one life. See also the related talk page citton-attention-gesture-machine, where Benjamin appears on the apparatus.
  6. Wayne Booth, The Rhetoric of Fiction (1961). The implied author: the version of the writer you infer from the work. Feeds the talk: the “who I am as a writer” document [13:17] is a way to build an implied author on purpose and keep it stable.
  7. T. S. Eliot, “Tradition and the Individual Talent” (1919). A writer is in a dialogue with the whole past. Originality is a shift in the existing order. Feeds the talk: Graver’s “codify the experts” idea [16:18] is the tradition half. The “uniquely you” idea is the individual half. Eliot says you need both.

The thread. The talk assumes the expression view (a work needs a source with something to say). Barthes, Foucault and the intentional-fallacy school show that the claim is a position, not a fact. It is a good position for working artists. It is not the only one.

2.4 Audience theory — “knowing what the audience wants before it does”

The problem it addresses. Cultural products are hard to predict. Who is the audience and how do you reach it?

  1. William Goldman, Adventures in the Screen Trade (1983; from memory; not checked). “Nobody knows anything” about which film will work. Pushes back on the truism: if nobody knows, Graver’s rule is an aim, not a skill anyone can reliably have.
  2. Richard Caves, Creative Industries (2000; from memory; not checked). Names the “nobody knows” property and the economics around it: high up-front cost, uncertain demand, many near-miss products. Feeds the talk: the marketing loop that tests a “proxy trailer” before the show exists [22:46] is a way to buy information under that uncertainty.
  3. Theodor Adorno & Max Horkheimer, “The Culture Industry” (1944). Mass entertainment standardises and pre-chews, and trains audiences to want what they are sold. Pushes back on the talk: “knowing what the audience wants before it does” is, for Adorno, close to making the want. A system that tests trailers hourly is the culture industry made fast.
  4. Stuart Hall, “Encoding/Decoding” (1973). Audiences do not just receive a message. They decode it in dominant, negotiated or oppositional ways. Feeds the talk: what questions they care about [13:17] is a decoding question. A focus group of AI personas [14:51] is a guess at readings. It cannot show real ones.
  5. Henry Jenkins, Textual Poachers (1992) and Convergence Culture (2006). Fans are active makers who remix and share. Feeds the talk: Graver’s Imagineers story is the worry that “the barbarians were building bridges over the moat” [11:43]. Jenkins is the positive side: fans building with the IP can be an asset. His Best Week Ever story, where the audience read the producers’ own blog and gave signal [12:37], is a small case of this.
  6. Kevin Kelly, “1,000 True Fans” (2008). kk.org. A creator who earns about $100 a year per fan from 1,000 fans can live on it. Feeds the talk: Graver wants the same for television writers [42:10]. Critics of the model note the hard part is finding and keeping the fans, which is the third piece he says is missing.
  7. Pierre Bourdieu, The Field of Cultural Production (1993; essays from 1980s; from memory; not checked). Cultural value is fought over inside a field; “pure” art and commercial art follow different logics. Feeds the talk: Graver treats market insight and artistic value as friendly. Bourdieu treats them as rivals. His own phrase “show business” lands on the commercial side.
  8. A line often attributed to Steve Jobs (“people don’t know what they want until you show them”, from memory; not checked) and to Henry Ford’s “faster horse” (a common misquote; not checked). Both are folk versions of Graver’s truism. Evidence for the attributions is weak. Use them as folklore.

The thread. Graver’s truism joins two ideas that sit in tension. One says the audience can be understood and served. The other says the audience cannot say what it wants, so the maker must lead. The honest middle is a hypothesis loop: guess, show something, read the reaction, revise. His circle of ideation → production → audience [22:46] is that loop.

2.5 Tacit knowledge and codifying expertise

The problem it addresses. A great craftsperson knows more than they can say. Can that knowledge be written down and handed on, to people or to a model?

  1. Michael Polanyi, The Tacit Dimension (1966). “We can know more than we can tell.” Skill and judgment sit in practice; rules only point at them. Feeds the talk: Graver says “a thousand and one decisions” a week need expertise [17:15]. Part of that is codifiable (marketing methods, story frames). Part is not (what makes a joke land).
  2. Hubert Dreyfus & Stuart Dreyfus, five stages of skill (1980s; from memory; not checked). A novice follows rules; an expert acts from trained perception, often without a rule. Feeds the talk: the plugins act as a novice’s rules (“rails in a bowling alley” [24:54]). They lift the floor. By this model they will not make an expert.
  3. Ikujiro Nonaka & Hirotaka Takeuchi, The Knowledge-Creating Company (1995; from memory; not checked). The SECI model: knowledge moves between tacit and explicit by socialisation, externalisation, combination, internalisation. Feeds the talk: Graver describes externalisation (writing his method and culture down), combination (merging experts’ material in a hub), and the user’s internalisation (the person learns by following the steps). It is an old idea in a new tool.
  4. Harry Collins, Tacit and Explicit Knowledge (2010; from memory; not checked). Splits tacit knowledge into kinds: some is only contingently tacit (could be said, nobody has), some is relational, some is collective (learned only by living in a community). Feeds the talk: a studio’s “culture” [18:11] is collective tacit knowledge. Documents may capture part of it. They cannot replace being in the room.
  5. Donald Schön, The Reflective Practitioner (1983; from memory; not checked). Professionals think “in action”, with a back-and-forth with the material. Feeds the talk: Graver’s loop of asking, testing, rejecting a past decision [46:35] is reflection-in-action with a model as the other voice.
  6. Atul Gawande, The Checklist Manifesto (2009; from memory; not checked). Even experts gain from simple steps in complex work. Feeds the talk: “follow directions clearly” is what the host says AI does well [19:02], and checklists exist for the same reason.

The thread. The plugin idea is the oldest move in knowledge management: write the method down. The theory says what survives the write-down (the explicit steps) and what does not (the perception behind them). Graver’s own use gets this right when he says the real skill is “your version of everything” [30:20], which only a person can supply.

2.6 AI and sameness — regression, homogenisation, collapse

The problem it addresses. If models learn from the average of many texts, do their outputs, and the work of people who use them, drift toward the average?

  1. Francis Galton, “Regression towards mediocrity in hereditary stature” (1886; from memory; not checked). The original statistical sense of “regression to the mean”. It describes a statistical effect. Graver uses the phrase loosely, for typical output.
  2. Doshi & Hauser (2024). The evidence above: individual gain, collective sameness (PMC). Feeds the talk: Graver’s advice (bring your own material first) is aimed at the individual-level risk. The collective-level risk (everyone gets the same suggestions) is not in his frame. If every writer starts from the same plugin, that risk grows.
  3. Padmakumar & He (2024). Writing with a feedback-tuned model, not a base model, reduced diversity (arXiv). The authors suggest that tuning to human preference may push models to the safe middle. Feeds the talk: this supports the idea that “the mean” is partly a product of training, not an unavoidable fact.
  4. Hito Steyerl, “Mean Images” (2023; from memory; venue not checked). Statistical images are “mean” in both senses: average and cruel. See citton-attention-gesture-machine, where she is cited. Feeds the talk: Graver treats the mean as a style problem. Steyerl treats it as a politics problem too: whose faces and stories sit at the centre.
  5. Ilia Shumailov and colleagues, “AI models collapse when trained on recursively generated data” (Nature, 2024; from memory; not checked). Models trained on model output lose the tails of the distribution. Feeds the talk: a world of AI-assisted scripts that feed the next model’s training would narrow further. This is a system-level reason to keep human outliers, which Graver calls quirks [39:25].
  6. The taste argument. Models can rank and edit well. Whether they can originate an outlier is open. The best-supported small claim today: they raise the floor for weak writers (Doshi & Hauser) and do less for the ceiling. Graver’s own experience is a mild counterexample. He found a model-made series idea “not a terrible idea” [32:19]. That is a floor result.

The thread. There are two kinds of “mean”: a single output that is generic, and a crowd of outputs that is uniform. Graver’s method defends against the first. The second needs a different defence, such as different seeds, different sources and humans who are odd on purpose.

2.7 Law, credit and labour

The problem it addresses. Who is the author in law? Who gets paid? What happens to the people who used to do the work?

  1. Burrow-Giles Lithographic Co. v. Sarony (US Supreme Court, 1884; from memory; not checked). A photograph can be authored because the photographer made creative choices (pose, lighting, arrangement). The early test for machine-aided authorship. Feeds the talk: the same logic sits behind the 2025 Copyright Office view.
  2. US Copyright Office, Copyright and AI, Part 2 (29 Jan 2025). Prompts alone are not enough; human selection, arrangement and modification can be protected; the protection covers only the human parts (IPWatchdog). Thaler v. Perlmutter: no registration for a work with no human author; D.C. Circuit affirmed 18 Mar 2025; certiorari denied 2 Mar 2026 (Mayer Brown). Feeds the talk: Graver’s “who I am” file, his dump of his own material, and his rewrites give a stronger authorship record than a prompt does. He does not talk law. The link is mine. For the wider legal picture, see samuelson-generative-ai-meets-copyright.
  3. WGA 2023 agreement. AI is not a “writer”; AI output is not “literary material”; disclosure is required; writers may use AI by consent (WGA summary). Feeds the talk: a “virtual writers’ room” has to stay on the tool side of that line.
  4. David Hesmondhalgh & Sarah Baker, Creative Labour (2011; from memory; not checked), Pierre-Michel Menger, Artistic Labor Markets (from memory; not checked), Sherwin Rosen, “The Economics of Superstars” (1981; from memory; not checked). Creative work has low pay for most and huge rewards for a few; careers are insecure; rewards concentrate. Feeds the talk: Graver’s “boys, those jobs ain’t coming back” [36:09] and “where do the other 50 good ideas go?” [42:10] describe the squeeze these authors study. Cheaper production may increase the supply of ideas, which, in superstar markets, pushes more of them below the line.

The thread. Law and labour set the price of Graver’s advice. A writer who uses AI as a room can stay an author. The harder question, who pays for the work, is the “third piece” he calls unsolved.

Part 3 — Deep dives

1. The AI as a virtual writers’ room, not a ghostwriter

🎙 In the talk.

  • His first rule is about what AI is not doing: “none of that is about AI writing for you” [14:51]. He lists the uses. Ask if the idea and the person fit [13:17]. Ask what audience is out there and where they are [13:17]. Build a “focus group of eight people”, including “the guy you really wish you hadn’t invited” [13:17]–[14:51].
  • In the writers’ room picture, he recalls seeing a shared document on a screen capturing the room’s talk [36:41]. He expects AI to join: “you will have AI in there and they will be asking questions like” what happens if a character does something [36:41].
  • He wants more than one voice: “much better if you’ve suddenly got three people with different perspectives” [38:03]. He adds a friction point: “everyone could hate the AI” [38:03].
  • A real room also has odd specialists, such as a tennis coach on a show about a young tennis star [40:31].
  • Quintero’s frame: use AI to do “in different capacities” what you used to need a team for, “instead of working for AI” [33:01].

📚 Deeper theory

  • Distributed authorship has a long record (2.1). Becker’s Art Worlds and Sawyer’s improv studies both say that groups make work no member makes alone. The AI-room claim is a claim about adding members, not about replacing the author.
  • Devil’s advocate and “red team” roles are known to improve group decisions. Several personas with set viewpoints is a way to put them in. This is a design pattern, not a result from Graver’s talk, and the evidence that a single model playing three personas is as good as three people is thin (from memory; not checked).
  • The friction argument. A shared AI target takes heat off people. That fits a general idea in group work: a common outside object lowers interpersonal conflict. It also hides a cost: the room may stop arguing with each other.
  • Persona groups are guesses. A focus group made by a model reflects the model’s training, not real people. Hall’s encoding/decoding work (2.4) warns that real audiences read in ways the maker did not plan.

Where others push back.

  • Writers’ unions put the line at credit and consent (2.7). A room of AI voices is fine if no one hands the pen to it.
  • Critics of persona simulation argue models flatter, agree, and give the middle view. Doshi & Hauser’s collective-sameness finding applies when many people use the same room (2.6).

What they leave out. Who runs the room. In a human room the showrunner picks. Graver says “you have to be pushing it” [38:03] but does not say how a junior writer with no taste can tell a good note from a bad one. He also leaves out cost, privacy of unpublished work, and what it does to junior writers who used to learn in a room.

Check yourself

  1. What is the difference between “ghostwriter” and “virtual writers’ room” in terms of who makes the final decision?
  2. Why might three AI personas be less diverse than three humans, even if each persona is written differently?
  3. Which part of Becker’s idea of art as collective action does the AI room extend, and which part does it leave out?

2. Start with yourself: empty out, then open the model

🎙 In the talk.

  • His starter pack: a “who I am as a writer” document with inspirations, loves and aims; a second document for the work in progress [13:17]. He quotes Didion: “I don’t know what I think about something until I write about it” [13:17].
  • On the order of work: “Spend a day or two dumping everything you know about what you’re going to do onto a page” [27:20]. By hand or by voice. Then: “you have to start with uniquely you, what you know” [27:20].
  • Why: once you open a model, “there’s one chance to kind of be yourself” [27:20]. That is his claim that the first draft of thought should be yours.
  • He ties this to flow: “a book that was very popular for a long time and you don’t hear much about it called flow” [26:39]. If you have not first gone there, “you’re not going to get a good product” [26:39].
  • He compares it to his old pitching habit: write all you know, then edit it down to a two-minute pitch [30:20].
  • Quintero’s summary: “go inward” before using the tools [30:20]. Graver agrees: “I really firmly believe that” [30:20].
  • The compass image is hers [33:01]: with a compass the model works for you; without one “you’re just kind of following Claude in whatever direction it’s taking you in”.

📚 Deeper theory

  • Preparation and incubation (Wallas, 2.2). Dumping is deliberate preparation. It loads the mind with material the model has not touched.
  • Anchoring. In psychology, the first idea someone sees tends to pull later ideas toward it. If the model speaks first, its idea may anchor yours (from memory; not checked as applied to LLMs). Graver’s order of work guards against that, though he gives no evidence.
  • Didion’s claim as method. Externalising thought is how you learn what it is. A model can help, but if it does the externalising you may skip the learning.
  • Expression theory (2.3). The “uniquely you” step is the expression theory turned into a workflow: the source comes first, tools second.
  • Implied author (Booth, 2.3). The “who I am” document is a way to build and hold a consistent voice for a model to serve.
  • Tacit knowledge (2.5). A dump is an attempt to move tacit knowledge into words so a model can use it. Some will stay unsaid.

Where others push back.

  • Some creative people start by talking with a model because the blank page blocks them. A rule to always dump first may be a personal habit made into a rule. No test of his order of work was found.
  • Barthes-style views say the “source” is partly the reader and the culture, not only the writer’s inner store (2.3).

What they leave out. What to do when you have not yet got a voice, which is where most beginners are. He half admits it: “some people just don’t have it” [34:16]. He also leaves out how a “who I am” file can fix a writer into a past self.

Check yourself

  1. Which stage of Wallas’s model is the “empty out” step?
  2. Why might asking the model first be a problem even if its idea is good?
  3. What is the difference between having a compass and having a “who I am” document?

3. Codifying expertise: the core, the culture, the plugins

🎙 In the talk.

  • Quintero sets the idea: plugins are “built by experts in their domain” [15:38]. She defers to Klaassen on engineering and Parrott on writing, so the steps are followed “safely and following best practices” [15:38].
  • Graver’s version has a central core with two inputs [16:18]. First, expertise in making entertainment, much of it published, from David O. Selznick to Francis Coppola to Quentin Tarantino [16:18] (the captions render these names loosely), covering writing, producing, editing, marketing, distribution, business and legal [17:15]. Second, personal and studio culture: each person adds a document about themselves, so the core has a really great petri dish of cultural documents [18:11].
  • Use case: when marketing and writers disagree, ask how others have solved it, to help mediate and reduce friction [17:15]–[18:11]. He says show business often turns personal and people “walk away” [18:11].
  • On form: “People in show business still make things the way Henry Ford made cars” [19:02]. He wants the marketing and business people talking to the writer earlier, before decisions get locked in [19:02].
  • Quintero on why plugins work: AI follows directions, and plugins turn a complex process into clear steps [19:02]. She also describes the three engineering steps: plan, build, review [19:02]–[20:22].
  • Quintero again: plugins are “the rails in a bowling alley” against slop [24:54].

📚 Deeper theory

  • Explicit vs tacit (2.5). Published advice from famous filmmakers is explicit knowledge. The selection (“which advice, when, for this show”) is closer to tacit. A core that holds both opposing views leaves the choice to the user.
  • Nonaka’s SECI (2.5). The core is externalisation and combination. The loop that follows (use, reflect, update) is internalisation. The “culture documents” are an attempt at socialisation by text.
  • Novice rules vs expert perception (Dreyfus, 2.5). Plugins are strongest for those who lack a method and weakest for those with a method of their own. That fits Quintero’s point that she trusts experts in fields not her own [15:38].
  • Mediation by a third party. A neutral reference lowers the heat in a dispute (2.1). It does not decide who is right; it gives both sides a shared frame.
  • A risk: monoculture. If many teams adopt the same codified method, their work will look alike (2.6). Doshi & Hauser’s collective effect would apply at the level of method.
  • Another risk: authority by citation. Quoting Selznick or Tarantino in a core makes advice sound settled. The famous often disagree with each other, and with their own films.

Where others push back.

  • Some craft traditions say you cannot learn the craft from rules (Polanyi, Dreyfus).
  • Some creators say the hub makes teams default to the safe, tested path (Adorno-style critique, 2.4).

What they leave out. Who vets the experts in the core. What the model does when two experts clash and which source it favours. Who owns the codified studio culture, and whether a person leaving takes their “cultural document” with them. Also: how often the core must be updated, a point the video description raises with “six-month shelf life”, which this page does not use as evidence.

Check yourself

  1. Which parts of a studio’s culture can be written down, and which only lived?
  2. What would a Dreyfus-style critic say a novice gains and an expert loses from a plugin?
  3. How could many teams using one core make their output more alike?

4. The three-phase loop and the audience

🎙 In the talk.

  • The three phases [20:32]: “ideation, and we’ll call it architecture or design” which produces a script; production, “everything from how do we prepare to make this thing”; and reaching the audience.
  • Production example: an AI-animated series with writers, an animation studio in India and computer scientists in Zurich [21:13].
  • Marketing already uses AI: studios watch ticket-purchase intent from trailers and “rewrite the trailer almost on an hourly basis” to A/B test [22:46].
  • His proposal: “make it a circle” so marketing joins at the start. “Let’s create a proxy trailer for that” and see if you can sell tickets “even though it doesn’t exist yet” [22:46]. His comment: “which is essentially what show business has always been about” [22:46].
  • Earlier, on audience: Best Week Ever’s producers had a blog and found 10–15,000 readers, “read the audience signal, see what’s going on and it was a complete accident” [12:37]. (The platform and numbers are his words; unverified.)
  • The Imagineers lesson: a new tension between what the audience can make and what a company can control, “the barbarians were building bridges over the moat” [11:43].
  • The thesis: “success is knowing what the audience wants before the audience knows it” [39:25]. AI can tell him “a whole bunch of stuff” about the audience for his senior-citizen series, but not the character quirks [39:25].
  • The weak point: “I believe that that is for the first two pieces of the circle. I think all that is coming into place. It’s the third piece that’s really really hard” [35:31].

📚 Deeper theory

  • The field in the systems model (Csikszentmihalyi, 2.2). A work counts as creative when a field accepts it. Marketing is the field’s front door.
  • “Nobody knows” (Goldman, Caves, 2.4). Testing a proxy trailer is a way to buy information cheaply under that uncertainty. It also changes what gets made: ideas that test well are favoured.
  • The culture industry (Adorno, 2.4). The risk of testing at the start is that the test shapes the work toward what already sells. This is the sharpest critique of the circle.
  • Hall and Jenkins (2.4). Real audiences decode and remix. A trailer test measures click intent, not the meaning people make later.
  • The long tail and true fans (Kelly, 2.4). The 1,000-fans model is the economic answer for a creator who skips the studio. It makes the third piece depend on direct support.
  • Product thinking. The three-phase loop echoes product work (discovery, build, launch). Graver says there is “a fair amount of kind of product design and product management” in entertainment [16:18]. He treats a show as a product with a market test, which some art-minded people reject.

Where others push back.

  • A film made to pass a trailer test may be safe and forgettable.
  • Slow-burn works that found their audience late (many cult shows) would have failed an early proxy test.

What they leave out. Evidence that proxy trailers predict real sales. What to do when the test and the writer disagree. And the ethics of using purchase-intent data from viewers.

Check yourself

  1. In Csikszentmihalyi’s model, which part is “the field”, and where does Graver put AI in relation to it?
  2. What does Adorno’s critique say about testing a proxy trailer before the show exists?
  3. Why does Graver call the third piece the hard one?

5. “Something to say”: the mean, the quirk, and taste

🎙 In the talk.

  • Leibovitz: “Everybody has a camera that’s better than the camera that I used for 30 years in their pocket, but there are very, very few people who are taking pictures like I take” [25:32] (attribution unchecked).
  • The test of a career: “Do you have something to say?” [33:01]. AI helps you say it. “If you ask it to tell you what it is you should be saying”, he says, “it goes to the mean” [34:16].
  • He admits it sounds harsh: “some people just don’t have it” [34:16]. He asks would-be writers about the last thing they wrote, and if it was weeks ago, he suggests other work [34:16].
  • Claude’s series idea: a web series about 60- and 70-somethings running a senior home’s Facebook page [31:45]–[32:19]. His reaction: “that’s not a terrible idea” [32:19]. He adds: “I think Claude has learned from me” [32:19].
  • The limit: AI “wouldn’t tell me … the personal quirks that would make these characters endearing” [39:25]. His example is Homer Simpson’s voice, which was improvised [39:25]. The point he draws: “what is inside of you that nobody else has” [39:25].
  • A room works the same way: each person brings something, from a Harvard background to improv to a tennis coach [40:31].

📚 Deeper theory

  • Expression theory (2.3). The “something to say” test follows Tolstoy and Collingwood. Its opposite (Wimsatt & Beardsley, Barthes) holds that meaning comes from the work and the reader. A fair reading is that Graver is describing what a working maker needs, not a proof about all art.
  • Boden’s three kinds (2.2). A model-made premise is a combinational idea (senior citizens + web series + Facebook). It is a floor-level result. A new form would be transformational.
  • Floor and ceiling (2.6). The Doshi & Hauser result says AI helps weaker writers most. Graver’s “not a terrible idea” fits. His claim that you need your own source for the ceiling is consistent with that, though not proven by it.
  • Aura and the particular (Benjamin, 2.3). The improvised voice is a case of unrepeatable human detail. Models can imitate a quirk once they have seen it. Whether they can start one is the open question.
  • Mean in two senses (Steyerl, 2.6). “Goes to the mean” is a taste complaint. It may also be a justice complaint: whose stories sit in the middle.
  • Taste as selection. Many theorists (Bourdieu, Koestler-style combination) treat taste as the ability to pick among options. This matches Graver’s demand that “you have to be pushing it” [38:03]. If taste is selection, then a model that makes many options raises the premium on selection.

Where others push back.

  • “Do you have something to say?” can be a gate that keeps people out. Many good writers began without a clear voice and found one by writing.
  • A model that “learned from me” [32:19] may be mirroring his style back. That is flattering and also a trap, because it reduces surprise.

What they leave out. Any evidence for how often models make real outliers, or how to tell a quirk from noise. Also the question of who decides that a result “goes to the mean”: the maker, who may be biased toward their own style, or an outside reader.

Check yourself

  1. Is “something to say” a claim about art or about a working career? Which does Graver mean?
  2. What is the difference between Boden’s combinational and transformational creativity, and where does the senior-home series sit?
  3. Why might a model that “has learned from me” be less useful than one that has not?

6. The working stack: Claude, ChatGPT, NotebookLM and an unpleasant check-in

🎙 In the talk.

  • On Claude as the home base: “I’ve lived in Claude for a long time. So, it’s the room I keep going back to” [44:10]. He points it at a folder of his work and asks for three things: does this sound like me, is there an audience on a real platform, and what will I need to reach a finished product [44:10].
  • He wants honesty about cost: “do you really want to go down the hard road of writing all of this” [44:10].
  • The check-in agent: “I’ve asked him to be very unpleasant” [45:19]. If he starts a new idea at night, it asks in the morning what happened to the first, because “you owe your thesis statement to every tomorrow afternoon” [45:58].
  • Frameworks: the story spine, “what Pixar uses in a lot of their stuff … the simplest way to do anything” [45:58]. He asks the model whether his story “is doing the work” [46:35].
  • Memory: “the compound part is it knows what decisions I’ve already made” [46:35]. It can say “You’ve … clearly rejected the decision you made three days ago” [46:35].
  • Second tool: ChatGPT is “the film school student who has read every guide, seen everything, does not know what it means” [47:16]. It is good at “the compulsories” [47:16].
  • Third: NotebookLM, for the two podcast hosts who “love what you’re working on” [48:02].
  • A pitch-deck skill from his Barcelona course that will not “let you do a step until you’ve completed the other” [49:18]. He says “it might not be that good. You can tell me.” [49:18]. He says European film schools have embraced AI more than American ones [48:36].
  • Advice early on to the unsure: ask the model how it can help, and feed it Substack posts about skills and ask “Is this useful?” [27:20]–[28:54].

📚 Deeper theory

  • Commitment devices. An agent that nags is a modern commitment device: you bind your future self to a plan. Behavioural economics has studied this for decades (from memory; not checked). The unusual part is that you design the critic.
  • Sycophancy. Models tend to agree with users. Asking one to be unpleasant is a hand-made fix. Studies of model sycophancy exist (from memory; not checked), and a stored “mean” persona is one workaround.
  • Reflective practice (Schön, 2.5). The loop of ask, test, review, reject is reflection-in-action with a stored history.
  • Division of cognitive labour. Using different models for different roles echoes team design: one reviewer is never the best reviewer of its own work. It also hedges vendor risk.
  • Scaffolding (Wood, Bruner and Ross, 1976; from memory; not checked). A step-gated skill that blocks progress until the last step is done is scaffolding. It teaches by structure. The danger is that learners follow it without understanding.
  • Anthropomorphism. Graver calls the check-in agent “he” and “very unpleasant” [45:19]. People find it easier to take critique from a character. That may help or may hide that the critique is statistical.

Where others push back.

  • A stack of three models and a daily agent may be more machine than a lone writer needs. The routines can become procrastination.
  • A heavy stack raises the data question: what happens to unpublished scripts.

What they leave out. Costs, privacy terms, and failure cases. Also an honest account of when the nagging agent fails. His claim that a model “has learned from me” [32:19] is a feeling, not a test.

Check yourself

  1. Why would a designer ask a model to be “very unpleasant”?
  2. In Schön’s terms, what does the memory of past decisions add?
  3. What does a step-gated skill teach, and what might it fail to teach?

7. The moat, the jobs and the missing third piece

🎙 In the talk.

  • On the old industry: “They flew you first class and it was glorious” [36:09]. “Boys, those jobs ain’t coming back” [36:09].
  • On the Imagineers: the fear in the late 1990s that you could “digitize a copyright” and let the audience run free [10:33].
  • On the new tribe: storytellers who “don’t think like David Ellison”, who think of a blanket and a garage [09:58]. (Ellison is likely the Paramount Skydance chief; from memory; not checked.)
  • On Scorsese and storyboard artists: a joke that “you’re going to have to get another job” [21:13]. The claim that Scorsese works with an AI storyboard company is unchecked.
  • On scarcity: “Fox is only buying one a year, where do the other 50 good ideas go?” [42:10]. He wants a Kevin Kelly-style place to put them [42:10]–[43:35].
  • On skill as a door: “your technical skills count. They will often get you in the door” [03:11]. And: “you’re going to get replaced by somebody who knows how to use AI” [03:11].

📚 Deeper theory

  • Superstar economics (Rosen, 2.7). When production and distribution cost less, the top gets more and the middle gets less. Cheaper tools make more supply, not more pay.
  • Creative labour (Hesmondhalgh & Baker, Menger, 2.7). Creative careers carry high risk and low pay for most. Graver’s “golden age” was a protected corner of that. His new tribe is closer to the normal condition.
  • IP and the moat. The 1990s Imagineers worry was about control of copyright. Today the same worry has moved to training data and output (see samuelson-generative-ai-meets-copyright).
  • Disclosure and credit (WGA, 2.7). The union deal draws a line between a writer who uses AI and AI as a writer. It protects the first and not the second.
  • Direct-support models (Kelly, 2.4). Substack, Patreon and similar tools are the nearest real answer to the third piece. They work best for solo or small-team creators and less well for expensive, team-made TV.
  • A caution on the “gatekeeper” story. Graver’s story of a better world with a group in a garage is the Jenkins-style optimism. Studies of platform creators show most make little money (from memory; not checked).

Where others push back.

  • “Get skills and you will be fine” puts the burden on the worker. Labour scholars note that new tools often drive wages down even for the skilled.
  • A garage-and-blanket model assumes that distribution is open. Platforms still control reach.

What they leave out. Who pays for the “other 50 ideas”, what happens to those who cannot or will not learn the tools, and the lost training ground of rooms and assistant jobs.

Check yourself

  1. What does Rosen’s superstar model predict about pay when making gets cheaper?
  2. Which part of the WGA deal fits Graver’s “room, not ghostwriter” line?
  3. Why is the “third piece” harder for TV than for a solo newsletter?

Putting it together

The mental model in one paragraph. Graver treats AI as a new member of the room, not as the author. A work of art has three jobs: finding what you have to say, making it, and getting it to an audience. AI is strong at the second and useful at the first and third, but it cannot supply the source, the odd personal detail that makes a character real. So the order of work matters: first empty yourself out, then bring in a room of voices, a codified method and a culture file, then loop back to the audience before you build too much. The theory shows what holds up and what does not. His expression view of art is a real tradition, and a useful rule for working makers, but it is a position. The research on homogenisation shows the risk is both individual (generic output) and collective (everyone’s output alike). Law and union rules reward the human choices he recommends. The economics, the “third piece”, is untouched by any tool.

Three tensions to keep open

  1. Room or ghostwriter: who decides? The line is the human veto. It holds only if the person has taste. Whether taste can be built with the same tools is not shown.
  2. Know the audience or lead it? “Before the audience knows it” mixes a service view and a leadership view. Testing a proxy trailer early may favour the safe.
  3. Codify or keep tacit? Plugins lift the floor and might narrow the field. The more teams share a core, the more their work looks alike.

Reading path

  • Start: Doshi & Hauser (2024), the abstract and figures (PMC). Kelly, “1,000 True Fans” (kk.org). Csikszentmihalyi, Flow, the chapters on creativity.
  • Then: Howard Becker, Art Worlds, ch. 1. Margaret Boden, The Creative Mind, the opening chapters. Polanyi, The Tacit Dimension. Barthes, “The Death of the Author” and Foucault, “What Is an Author?” (short). The US Copyright Office Part 2 summary (Copyright Alliance); then samuelson-generative-ai-meets-copyright.
  • Deeper: Sawyer, Group Genius. Caldwell, Production Culture. Nonaka & Takeuchi, The Knowledge-Creating Company. Collins, Tacit and Explicit Knowledge. Hesmondhalgh & Baker, Creative Labour. Padmakumar & He (2024) (arXiv). Steyerl, “Mean Images” (via citton-attention-gesture-machine).

Q&A

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

Graph View

Backlinks