The Writers Aren't Luddites. They're Bargaining Over Who Holds the Models

ChatGPT had been daily life for six months, GPT-4 for two and a half, when Nathan Labenz sat down with three members of the striking Writers Guild of America. Italy had briefly banned the chatbot. EU regulators wanted disclosure of copyrighted training data, prompting Sam Altman to float leaving Europe before walking it back. Labenz had expected the first AI labor fight to break among doctors or lawyers. The writers got there first, with strike demands about how AI may and may not be used in Hollywood — and with online caricatures of them as reactionary Luddites that he did not trust.

A WGA board member declined to talk while negotiations were live, so Labenz went through friends. Trey Kollmer, co-executive producer on CBS’s *Ghosts*, had been described as perhaps the single most knowledgeable Guild member on AI; he cited Neel Nanda’s grokking paper, had gone to the Asilomar AI conference in 2016 and NeurIPS the next year. Sophia Lear, also a co-executive producer on *Ghosts* and a veteran of *New Girl* and *The Unicorn*, had the least hands-on experience and the most radical uncertainty. Garrett Schabb, who had written for *Tosh.0*, *Suits*, and Crooked Media, had used ChatGPT extensively and written himself a personal ethical code for how much creative contribution he would allow.

Labenz’s takeaways, stated up front, are blunt. The writers are not Luddites. They are curious about the technology, they get that it is both an amazing tool and a potential threat, and they are not trying to ban ChatGPT. They want the shows and movies viewers watch to remain fundamentally human creations — and, yes, they want to keep getting paid. Collective bargaining, he thinks, will matter more in the AI era, because the range of possible economic structures is so wide. Wendy’s had just said its AI drive-through outperformed human order-takers; fast-food workers do not even have a table to sit at. Letting AI walk into labor markets with no structural reform, he argues, is more likely to provoke the backlash and overregulation that accelerationists fear. It is no coincidence that Altman is also running UBI experiments.

The right answer, he adds, varies wildly by domain. In medicine, an individual’s right to quality health information should outweigh a doctor’s right to high income, and to their credit the medical establishment has been more positive than he expected. Culture is murkier. Yuval Noah Harari’s warning that language is the operating system of civilization, and that AI is gaining write access — even the seed of new religions, since sacred texts have always claimed non-human origin — makes a certain French-style cultural protectionism look prudent. And everyone may still be thinking too small. Dynamically generated 3D worlds with AI-animated characters could force an entirely new set of questions, far past ChatGPT.

Kollmer has written broadcast sitcoms since about 2010 and is finally on a show that is sticking around. He walks through how television writers get paid, because the AI fight is unintelligible without it. There are two paths. One is development: you pitch an idea, and if a network buys it they are hiring you to write the pilot script. A decade ago each network might buy about seventy comedy pilots, film ten to twelve, and green-light four or five. Those numbers have all shrunk. Writing that pilot is what delivers the “created by” credit, ownership, and a slice of backend profit participation — the most valuable things in the business.

The other path is joining a staff once a show is green-lit. *Ghosts* has fourteen writers; streaming rooms are much smaller, which is itself a strike issue. The room generates season arcs and episode ideas together, breaks the beats, then assigns a “captain” to write the first draft. When “written by” flashes on screen, Kollmer says, it really means that person wrote the first actual draft. Everyone had been brainstorming what happens; after the draft, the whole room usually rewrites it together. He can already see many different steps where a model could try to do some of the work.

The union contract is the Minimum Basic Agreement, the MBA — floors on pay, options, health insurance, and pension. Once a major studio signs, for content the MBA covers they can only use union writers. The last strike was about getting streaming under the contract. This one is pushing for more residuals from international streaming, and even for transparency on subscriber numbers, which platforms do not want to share. The old network world paid for reruns at home and abroad; that money has been hard to translate to streaming.

The writers had been back only three weeks after ChatGPT’s launch before the strike hit. Kollmer did not see anyone using it for work in his room. Among friends, a few were experimenting and having fun; a lot were not using it at all. He had been playing. He fed the models a parked premise — a Google-like company builds a realistic robot and tasks an employee with moving it in, to observe and learn the human world — and asked GPT-3.5, Bing, and GPT-4 to write the cold open of the pilot.

The split is clean. They are very good at structuring a scene and getting out the premise and themes, the big-picture work the scene has to accomplish. They struggle to get specific. For a joke you want unique language, and they go vague. They are not unfunny. In one bit the proud boss says they designed the robot to look exactly like one of their most attractive famous celebrities. The employee cannot place it. “Ryan Reynolds.” “Does it?” “Yes, it looks exactly like Ryan Reynolds.” “I guess so.” Kollmer thought that was really funny, and kind of scary. His verdict: medium-level execution, not the polish, perfection, and specificity you want.

For younger writers who do not yet know how to lay out a scene from a paragraph of outline, he thinks it could be a real ramp: drop in the paragraph, get the bones, then do what they are good at — jokes, character, polish. Once you have enough experience you should already be better than current models at layout. For his own work it has not yet made him better or produced anything usable. He is curious about how the models are improving, almost for its own sake. Where he would actually query them now is not alternate jokes so much as plot mechanics: five ways a couple could get money from a bank, what ID you would need to fake, a catalog of frauds from a book like *Lying for Money*.

Labenz notes the rhyme with early ChatGPT jailbreaks: ask how to hotwire a car and the model refuses; frame it as fiction and the concrete steps appear. Kollmer wonders whether training out those jailbreaks will make the models less useful for creative tasks. There may simply be a class of things — a detailed hot-wiring in an action movie — that a lab will never put in a script, because it looks bad if the model gives that information out at all, fiction or not.

Labenz cites a number he uses often: GPT-4 moved from the bottom 10 percent to the top 10 percent on the bar exam relative to GPT-3.5, yet users prefer it only about 70 to 30. Creative evaluation is in the eye of the beholder. He also hears, somewhat consistently, that earlier versions felt more freewheeling for creative work and that later ones have been pruned toward the normal and the less offensive. Comedy often lives near the line of discomfort. Kollmer’s deeper worry is downstream of ranking. A lot of the best scripts, movies, and shows are pretty polarizing. Depending on how reinforcement learning chooses and penalizes outputs, you get a risky piece beloved by 10 percent of users and hated by 90 — or one that is kind of fine with everyone.

Labs are hiring up the expertise stack, Labenz says, often through partners such as Scale AI, with PhD-level evaluator jobs in chemistry, accounting, and other deep fields. He does not know whether comedy writers are on those lists. Hollywood already has an army of assistants who read scripts and write coverage. He does not want to give anyone ideas, but at some point those people could be appropriated for a massive fine-tuning project on TV and movie scripts. He also expects community-driven customization: Discord servers shaping open-source models into infinite writers for anime or *Fast & Furious* fanfic, in a space totally divorced from typical production. Turning those scripts into finished assets is another question, and text-to-video is coming up hot.

Kollmer lists the four asks. MBA-covered companies should agree that AI cannot write scripts, cannot rewrite scripts, and cannot generate the source material that shows or movies are based on — the articles, short stories, comic books, and novels that are routinely optioned. And MBA-covered material cannot be used to train AI models. Most of the high-quality scripts that have been written and not produced, perhaps an order of magnitude more than the ones that were shot, sit with the studios. Without that clause they could fine-tune on them tomorrow. The Guild, he is careful to add, is in favor of writers using the tools themselves.

The first big fear is the force-multiplier shape of work. A show might have one or two showrunners with the big idea; the models generate first drafts or decent episode scripts; most of the writers become gig-economy cheap labor who come in to polish and punch up. One writer might do three shows. That nightmare, Kollmer stresses, is most realistic in a world of *medium* capability growth: a little better, a little more, not so much better that everyone is screwed anyway — just enough to flatten the middle of the profession.

The second fear can almost be done with current versions: producers on the studio side using a model to cheat the designations. Writing a pilot is what gets you “created by” and the backend. If a model drafts that script, the property quietly moves. He also sees the other side of the wall. At some point models will make production cheaper too. A writer might go off, write a script, and not need studios, actors, directors, or composers. If capabilities jump across the whole stack, studios that sign a no-AI MBA could get their lunch eaten by an outside company, or by fans making shows for themselves. The MBA lasts three years. Promises rot on that clock.

Labenz asks the practical question: is there any way to police this? Detectors are weak, broken by a superficial edit, or they flag human prose. Kollmer’s honest answer is that a producer who has never been writing and is suddenly pumping out content might be obvious now. As people break in and the technology gets better, it is probably really hard to tell. If the models start generating multi-million-dollar movie ideas, the incentive to cheat is enormous and the cost of inference is collapsing — Labenz cites something like a 97 percent price drop in eight months for equivalent models, plus a new high end that costs the same and does more.

Cheap expertise, Labenz thinks, is a big part of the future; he has sat through forty-five-minute medical and legal consults with models that rivaled the humans he actually used. He is surprised lawyers have not taken a stronger stance. On the art side he tries a thesis on Kollmer: the remaining value is curation, a tastemaker role. You will not ask GPT-5 for one hit sitcom idea and get a bona fide hit every time. You will get a lot of pretty good ideas, and the scarce skill will be knowing what resonates, what is worth developing. Kollmer agrees that a differentiated brand — knowing what people will like, or being trusted to recommend — will keep capturing value, whether as a few massively popular curators or as a decentralized Instagram of taste.

There is a catch. Networks until recently bought about 250 ideas a year to put 15 to 20 shows on the air. If making content gets much cheaper, you almost get a TikTok limit: generate so much, see what sticks, iterate. Labenz wonders how long the human-model symbiosis window lasts, or whether it is like chess, where “human plus machine” was the best result for a few years and then the human could not add much. He cites that often because it feels like a cope. Kollmer says the teamwork ending is a very attractive storytelling device. Nature does not always work that way. Nature is full of non-happy endings.

Kollmer’s own attempt to prompt a timeless hit — it is 2044, write the magazine review of the most beloved show of the past twenty years, which premiered in fall 2024 — did not do the best job. Because he set the vibe in the future, all the suggestions were sci-fi. Language models take a lot of cues from the vibe you put out. Labenz takes that as evidence that writers are well positioned to be effective users of these models, if they are willing. On the picket line the signs tend to be dismissive: it is bad at writing, but studios do not care about quality, so they will use it because it is cheaper. Some writers see an existential fight. Others, especially drama and sci-fi people who got into the work because they are fascinated by ideas, are afraid and also excited to play. The Ryan Reynolds spark is the thing that exceeds expectations. The Guild, Kollmer repeats, wants writers free to use the tool; it does not want studios to seed control of culture to an unknown force. Another demand, minimum staffing on streaming rooms, is synergistic with the AI asks: if you keep enough humans on a show, you are less worried the models will destroy the middle of the job.

Lear is a co-executive producer on *Ghosts* and has spent years in 22-episode network rooms that run nine months, from breaking episodes through production and editing — a rhythm somewhat outside the streaming experience driving much of the strike. Asked how important the AI issues are compared with core revenue fights, she says they really do not know. The implications feel too nascent. Her personal judgment is stark: if AI can do a writer’s job as well as or better than humans, she does not think there is anything they can put into contractual language to change that. She believes humans will always have something to offer in scripts, stories, and especially jokes. She also says she could be 100 percent incorrect.

In the room, ChatGPT arrived as a slightly naughty, intriguing toy. A colleague’s brother had asked it for movie prompts; it fleshed out whole stories, pretty good ones. She thinks it is currently strong at log-line-level ideas. There is always time around a table of twelve when nobody has any idea what an episode could be about. “Does the robot have ideas?” is a real question. Lower-level writers already get sent off to generate raw story areas. She can imagine asking a model for fifty. What she will not give up is the alchemy of a comedy room: conversation, feeding off each other, unexpected collisions. What comes out of people talking is always more exciting than someone working alone — if quality is the bar.

If companies do not care about that bar, her belief does not matter. She reaches for a cookie: you can make a slightly better one, and people will still buy the cheaper one. Conditions for writers were already not great leading into the strike. There is a lot of existential fear in the Guild, and AI is part of it. Her own posture is unusually loose. If being a human writer was a phase in a longer history, she is glad she caught it, and she needs to get some more skills she can market. The thing is too large to fret about every day.

Labenz finds it hard to reconcile “studios will feed the audience crap” with how fiercely competitive content is. Lear says writers often feel the writing is the overlooked part of why something is good. She points at the *Lord of the Rings* show that landed as a dud: not enough time to write, the story not thought through, a bet that it would look cool. Audiences registered it as bad without quite knowing why. Money saved in the writing shows up on screen. On the ATM-and-bank-teller story — more branches, more jobs — Labenz doubts the pattern holds when people already watch eight or nine hours of video a day and cannot absorb a hundred times more content. Lear’s loss function is different: if everyone watches something tailored only to them, the sharedness of television goes away. She likes *Succession* in large part because she can talk about it with other people. Her pitch to the machines, if they are listening: go cure cancer. Let us keep writing stories.

Schabb has been in the Guild since 2012, comedy and drama, most recently *Suits*. The strike is in its third week, and the timing, to him, is almost luck. If ChatGPT and the other large language models had landed a year or two earlier, the studios would already have implemented them and the Guild would be arguing from behind. Instead they have a moment to get out ahead of the conversation. The Guild position, as he articulates it, is that AI should not generate literary material of any kind. A studio should not ask a model for a short story about time travel and then hire a writer, at a discount, to adapt that piece of intellectual property. It should not generate drafts for someone to punch up.

The dystopian and, he thinks, realistic view is that studios see a future with as little human contribution as possible. Writers are not opposed to change; they think the studio approach is geared toward profit rather than enriching human stories, and they can see how language models continue an existing trend to its logical end. There are rumors, uncorroborated, that some studios are already feeding public-domain books into proprietary models and asking for feature-length scripts — *The Count of Monte Cristo* as a modern-day, female-driven action thriller. Copyright law, he notes, has not been worked out at all.

His own use is extensive and fenced. During the strike he is outlining a new pilot alone at home, and he treats ChatGPT almost as a writer’s assistant just out of college: not someone you ask for pitches, someone you ask to hear you, recite back with clarity, and hold onto ideas. He talks about characters and dynamics, parks an under-a-thousand-word outline broken into acts and two-or-three-sentence scenes, and says “outline” whenever he wants it spat back. Then he hit the unexpected failure. On the way back out, the model makes subtle changes and omissions. It quietly changes the profession of the main character. It forgets a scene. After nine, ten, twelve iterations he realizes a scene he had relied on five hours earlier is gone, and he has to scroll up thousands of words to find it. Verbatim regurgitation, the simplest-looking job, is unstable.

Labenz offers two prompt habits: pin the structure with odd format tags, XML-style, and replace negation with positive instruction. “Do not change the structure” can work like “don’t think of a white rhino.” Say instead: you must always return exactly this structure and may modify only the parts I point to. Schabb has also, experimentally, tried to coax novel concepts out of the model after spending a lot of time explaining how novel concepts are arrived at. Almost without exception it reverts to cliché. That is why he holds his moral line: he does not ask it for pitches. He does not think he is missing anything novel by keeping it.

Every writer, Schabb says, has other writers’ voices in their head — Jesse Armstrong of *Succession*, Vince Gilligan of *Breaking Bad* — and they write toward those voices, consciously or not. Invoking a famous name in a prompt raises a question nobody has litigated: what is that artist owed when their unique voice is automated? High-profile writers who spent careers getting to that voice are worried. His scarier question is about the people who have not arrived yet. If most written content for TV and film is generated by a model, he is not sure we will see new voices. He thinks we will see an endless whirlpool of regurgitation of what is already out there.

Labenz rejects the popular shrug that humans remix too, so AI doing the same is no big deal. These systems are fundamentally alien. They do understand some things — he pushes back on the claim that they understand nothing — but not in the way we do, and human inspiration from other humans does not imply that AIs should run wild. Schabb would still call it morally okay to filter model ideas through a human prism and never let them see the outside world whole cloth, especially if they were asked for in someone else’s voice. Using those ideas unfiltered feels wrong to him. He also wonders whether models will get amazing at storytelling in *different* ways, so that human stories and AI stories sit side by side, the way games already do — as long as the AI is not asked to masquerade as us.

Fixed stories, he argues, are where cultural power lives. People read the same book, watch the same movie, listen to the same album, and bond over what it means. A former Screen Actors Guild director described studio ambitions for choose-your-own-adventure films: you in the passenger seat next to Vin Diesel, a *Fast & Furious* generated in a moment. Schabb can see the technology getting there. He does not think that form has much cultural power. *The Lord of the Rings* does not change for whoever is reading it; generations recognize the same thing. Stories are an antidote to the loneliness of life. A sugar rush of personalized plots loses that. *Succession*, written in a full room over a year, has depth and texture that shorter, faster shows do not; he can see that gulf growing. Studios, especially public ones, are not incentivized to find the best writing. AI could accelerate a world where shipping content is the point and a three-dimensional story is an afterthought. Writers have a financial fight and, he admits it sounds a little woo-woo, a moral one: the sanctity of really good storytelling.

Labenz asks how universal basic income would change the argument. Schabb is all for decoupling the right to eat from jobs. He cites Aaron Bastani’s *Fully Automated Luxury Communism*: the post-scarcity future is easy to imagine; the hard part is getting through the stretch where technology can automate a lot of work while the people who control it manufacture artificial scarcity. The Guild has to react as if that good world will not exist. If it does, he would happily give up being paid to write stories in exchange for everyone having a basic income and a high standard of living — “give me that over my stupid writing job for sure.” Labenz leaves him with a challenge: the thing in shortest supply is a compelling positive future, and writers are the group who might supply it. Schabb laughs. They are a cynical bunch, more so three weeks into a strike. They will try to accept the challenge.