AI Training Scripts

Introduction

TV writers fume over AI training scripts because the fight has moved from a strike-season flashpoint to a weekly line item in showrunner contracts. A Series Business investigation found 139,000 screenplays baked into one of the datasets that trained OpenAI and Meta models, and the writers who recognized their own work were not consulted. The 2023 Writers Guild of America deal set the first contractual AI floor, yet studios already read the fine print more generously than writers do. AI training scripts has become the single clearest test of how much creative labor can be scraped before the law or a union forces a settlement. This article walks through the dataset, the law, the 2026 Minimum Basic Agreement renewal, and the earnings story behind each writer’s anger. By the end, you will have an operational view of what is at stake, what is already changing, and what any writer or studio should be doing right now.

Quick Answers on AI Training Scripts in Hollywood

Why are TV writers angry about AI training scripts?

Writers found their copyrighted scripts inside large AI training sets, used without consent, credit, or compensation, which they say strips them of control over their own creative labor.

Did the 2023 WGA contract stop AI training on scripts?

No. The deal blocked AI from being credited as a writer, yet it still allowed studios to feed covered scripts into AI training with careful language and signed consent.

What are writers demanding in the 2026 negotiation?

Writers want opt-in consent for AI training, a residual tied to that use, a bar on training on unlicensed script corpora, and a dataset audit built into studio bookkeeping.

Key Takeaways on Writers, Studios, and AI Script Training

• At least 139,000 film and TV scripts sat inside one widely used dataset that trained commercial AI systems, including scripts from working WGA members.
• The 2023 Writers Guild of America deal set the first AI floor but let studios keep training models on covered material under narrow consent language.
• Writers point to Thomson Reuters v. Ross as proof that unlicensed training on copyrighted text can lose on fair use grounds in federal court.
• The 2026 Minimum Basic Agreement renewal will test opt-in consent, residual pay, and dataset audits that would reshape how AI training scripts actually works.

Understanding AI Training on TV Scripts

AI training scripts refers to feeding full screenplays into a large language model so the system learns dialogue cadence, scene structure, and character voice from copyrighted writing, often without consent. Covered material now sits inside datasets that power commercial writing tools.

How Studios Ended Up Training Models on 139,000 Screenplays

Hollywood did not plan to hand its scripts to AI developers. A fan site named ScriptSlug and the subtitle corpus OpenSubtitles were scraped by third parties. The scraped text was then pushed into a research corpus called Books3 and absorbed into training mixes at OpenAI, Meta, and Anthropic. By late 2024 reporters had identified 139,000 film and TV scripts inside one commercial training set, including shows written by current WGA members. That number came from a line-by-line match against ScriptSlug and OpenSubtitles, which cross-referenced each title to a known production. Writers who found their work in the set only saw it because an outside researcher published a lookup tool. The industry had no inventory of what had been ingested, which is still the real scandal under the newer scandals.

Studios did not supply those scripts directly, yet many of them had written internal memos exploring the same technology. Executives at Netflix, Amazon, and Disney had been experimenting with internal copilots trained on their own library, which is a different legal posture than scraped public corpora. The Netflix GenAI pilots were pitched as productivity tools that respected union terms, but writers heard a different message, which was that the training question was already settled. That gap in perception matters because trust, not technology, is what breaks first in these fights. Writers across the WGA have cited that trust gap in every public comment on the 2026 bargaining stance.

The 139,000-script number also understates the picture because subtitle corpora capture a larger slice of global television. OpenSubtitles alone holds decades of fan-translated episodes, which The A.V. Club reported is a richer training source than scripts for conversational dialogue. Subtitles come with timing data, speaker turns, and clean colloquial speech, which is exactly what a dialogue-generation model wants. Writers whose shows never leaked a formal script often still appear in the subtitle corpus, which means the exposure is wider than the headline.

Industry lawyers now treat script training the way they once treated piracy, which is as a quiet infrastructure problem rather than a one-time scandal. The intellectual property calculus around AI depends on how courts treat training as a derivative act versus a transformative one. Studios want to settle the question through licensing before the law lands, which is why the quiet licensing conversations began even as public fights heated up. Writers want the law to land first so they do not negotiate from the current position, which assumes the training has already happened and cannot be undone. Both sides privately expect the 2026 bargaining round to decide which model wins.

The 2023 WGA Deal and the Loopholes Writers Say Studios Are Using

Shifting focus to the contract that was supposed to settle this, the 2023 Minimum Basic Agreement set the first AI floor in American entertainment. The deal ruled that AI cannot be credited as a writer. It also barred AI from writing source material that undercuts a credit. The deal still allowed studios to train AI on covered scripts with signed consent and bargained compensation. That compromise was a win on the credit side and a loss on the training side, which the Authors Guild analysis of the AI protections flagged within days of ratification. The problem is that consent clauses are easy to collect in a package deal, where a writer trades one thing for another and signs. By early 2026 writers were testifying that the consent box had become a staple of overall deals.

A second loophole runs through the definition of covered material, which excludes anything a studio buys as a non-WGA pitch, including treatments, pilots acquired through packagers, and bulk library content. Studios can argue that script training only touches the uncovered tier, which lets them keep internal AI tools running without triggering the clauses that matter. Writers counter that the model does not care about the covered-tier line. Once a weight is updated by a non-covered script, it still generates output that competes with covered writing. That technical point is what the No Film School review of the WGA AI protocols keeps circling back to. Writers now want dataset-level audits rather than slower script-level audits that lag the training timeline.

A third loophole is the generative-material clause, which allows studios to use AI output as a starting draft so long as a human writer is credited at the end. In practice, under current loopholes, writers describe being handed a messy AI draft and asked to polish it. Hollywood’s AI fixation has already turned that complaint into a visible pattern on multiple sets. Rewriting a bad draft takes almost as long as writing the first draft, yet the writer is paid as a rewrite rather than an original. That gap between labor and pay is the quiet grievance that will drive the 2026 negotiating stance. Writers now log every AI-draft handoff as evidence for the next bargaining round.

How Large Language Models Actually Learn a Screenwriter’s Voice

Turning to the technical side, a large language model learns a writer’s voice by matching patterns across many scripts and compressing those patterns into weight updates. When a transformer sees a scene heading, a slug line, and ten pages of dialogue, it ingests spacing and dialogue cadence. It also learns the writer’s quirks about silence, interruption, and reaction all at once. That compression is what writers mean by AI training scripts in a practical sense, and it is what makes the output recognizable. A recent EMNLP paper called HoLLMwood showed that role-playing agents trained on screenplay corpora can hold character voice across scenes, which is a nontrivial technical result. It is also exactly what a studio needs to replace a staff writer on a procedural, which is why the paper landed with such heat in writer rooms.

Style transfer at the paragraph level is only one layer of what training captures. The model also learns story structure through next-token prediction over tens of millions of scenes, which is how it reproduces act breaks and beat sheets without being told the rules. That structural learning is why AI drafts can read as competent but hollow. The model has internalized the shape of a script without the thing a human writer brings, which is specific insight. The gap narrows every time a bigger model trains on a cleaner corpus, which is why writers treat each new release as a creeping threat rather than a stable one.

Fine tuning on a single writer’s catalog is now technically trivial. A small team can take a base model, feed it ten seasons of a showrunner’s scripts, and produce a specialist model that writes in that showrunner’s voice within an afternoon. That ease is why the Steven Moffat alarm on AI scripts read as more than a cranky quote from a famous writer, because the engineering reality matches the warning. Writers are not worried that AI will replace them next year. They are worried that AI will replace them cheaply when it is nearly as good and the show is a procedural.

Retrieval-augmented generation adds another layer that most union language has not kept pace with. A studio can keep its covered scripts out of the training weights entirely. The studio then stuffs those scripts into the model’s context window at inference time, which lets the system reference the canon without technically being trained on it. That workaround lets a studio claim it never trained on covered material while still extracting most of the generative benefit. It is a loophole that AI copyright lawsuits in the US are only now beginning to probe. Expect the 2026 negotiation to extend the training definition to cover retrieval use as well, because without that extension the current language is already outdated.

The Legal Case Writers Point To After Thomson Reuters v. Ross

Building on the technical ground, the legal ground also shifted in a direction that favors writers. In 2025 the Third Circuit upheld the Thomson Reuters versus Ross Intelligence case. The court held that unauthorized use of copyrighted material to train a commercial AI tool was not fair use. That was the first appellate ruling on the question in the United States. That decision broke the industry assumption that training was a transformative act protected by fair use, and it gave writers a precedent to cite in negotiating sessions. The AI Weekly report on the appeals court decision captured how quickly licensing conversations rewired themselves in the days after.

Writers do not need to win every AI training scripts argument that follows, they only need the baseline shift. If a court finds infringement when a legal research tool trains on headnotes, a similar court is likely to find infringement when an entertainment tool trains on copyrighted screenplays. The Hollywood Reporter coverage of the WGA push for lawsuits showed how fast the union moved from asking politely to demanding legal action. That shift is the practical read on the ruling, and it is why studio counsel began racing to license corpora rather than defend the training they had already done. The licensing races accelerated through 2025 and continue into 2026.

What TV Writers Are Demanding in the 2026 MBA Negotiation

Looking ahead to the 2026 bargaining table, writers have published five positions that will anchor every meeting. The headline demand is opt-in consent for AI training on any covered or non-covered script a writer owns or wrote under a studio deal. That demand replaces the current opt-out that lives inside package consent clauses. Writers want that consent to be severable from the rest of a deal, which means a studio cannot bundle it with a greenlight or an option payment. The Senal News account of the WGA position on screenplays describes consent as the single clause that moves everything else.

The second position is a residual tied to AI training use, patterned on the streaming residual that was the win of the 2023 fight. Writers want a per-model-release payment every time a covered script sits inside a training run, with a floor and a cap, and with an audit right that is enforceable. The audit right is the technical hinge, which is why the intellectual property fight at this level is about dataset inventories more than about rates. Studios have resisted dataset audits because they would expose years of training runs where consent was never collected. That exposure is what the whole fight is actually about. The fight is less about new training and more about accounting for training that already happened.

The third, fourth, and fifth positions are a bar on training on unlicensed scraped corpora, a credit protection for AI-assisted drafts, and a sunset clause. The sunset clause forces renegotiation every two years as the technology moves. Together, those positions push the fight from one about credit to one about the shape of the writing profession in 2028. Writers also want a tie-in to UK arts icons demanding AI copyright safeguards, because an aligned international position is harder for any one studio to break on. That international alignment is what turned the 2023 strike from a Hollywood story into a global labor story, and the 2026 renewal will test whether that alignment holds under pressure.

The Economic Hit to Mid-Career TV Writers and Showrunner Rooms

Beyond the big picture, the earnings story is where the fight gets specific. Mid-career TV writers earn most of their income from staff weeks and script fees. Both numbers compress when a show cuts a room from eight writers to five while an AI copilot drafts the overnight pages. The 2023 deal protected minimum staffing on certain show types, but it did not cover the growing short-order middle tier, where staffing was already thin before AI arrived. A report in The Ankler on AI script scraping showed how many working writers now treat AI tool rollouts as the next room cut rather than the next productivity tool.

Script fees are a smaller line than staff weeks on paper, yet they anchor a writer’s annual income. A one-hour script fee under current MBA floors runs roughly 42,000 to 48,000 US dollars for a WGA writer. A staff writer earns 5,800 to 7,250 US dollars per week before producer fees. When AI tools shorten the number of weeks a staff writer is needed, the writer loses a week of pay for every week trimmed. That is the straightforward math the digital labor shift under AI is now reproducing across creative industries.

Showrunners sit at the top of that compression but still feel it. A showrunner who once hired a room of nine for a 20-week season may now hire six for a 14-week season. That shift cuts both their room budget and their executive-producer bonus pool. The AI versus human creativity argument sometimes reads as abstract, but the economic version is concrete and painful. For a working showrunner on a mid-tier streaming show, the compressed room is already a 20 percent income cut from the 2022 version of the same job.

Diversity, Canon, and the Problem of Training on Legacy Scripts

Stepping back from economics, there is a cultural layer to AI training scripts that gets flattened in most coverage. A dataset of 139,000 screenplays skews heavily toward English-language procedurals, dramas, and legacy comedies. A model trained on it writes in the voice of the median show from the last 40 years rather than the voice of the next one. That skew punishes writers whose shows pushed form or language because their work either sits outside the training corpus or gets averaged away during inference. The AI and the arts conversation has returned to this point every six months since 2022.

Representation inside the corpus is a direct problem for diversity in output. Scripts from Black, Latino, and Asian American writers were under-represented inside Hollywood for decades. They are therefore under-represented inside the dataset that trains the tools studios want to use for drafting. That gap means AI-assisted writing defaults toward the voices that were already overrepresented, which is a quiet reinforcement loop that cannot be fixed by prompting. Writers who worked for years to push a show past the median are now being asked to compete with a model trained on the median they fought to escape. The dangers of AI ethical dilemmas frame fits this specific case tightly.

Implementation Checklist: How Writers Can Check If Their Scripts Were Scraped and Push Back

Turning to practical steps, any working writer can run a short checklist to see whether their scripts sit inside a known training corpus. The fastest check is the Series Business lookup tool for the 139,000-script dataset, which cross-references ScriptSlug and OpenSubtitles against writer and show metadata. A writer searches for the show name, confirms the match against episode number, and downloads the row for their records. That evidence becomes the basis for everything else, including a DMCA notice to the dataset host and a formal letter to the studio under the current MBA audit language. The original Series Business investigation page links directly to the lookup.

Writers who confirm their work sits inside AI training scripts corpora can file DMCA takedown notices with the dataset hosts, which is slow but useful on the record. The notice establishes that the writer never consented to the training use, which is relevant evidence if a later class action lawsuit opens. Writers can also file an unauthorized-use grievance under their current WGA deal, which forces the studio to disclose any internal AI tool trained on covered material. Those grievances are rare but growing, and the empty book protest for AI and copyright showed what collective visible action can do in a parallel medium. Writers who join the grievance roster also give the Guild better leverage at the bargaining table.

Writers should register their scripts with the US Copyright Office when they can. Registration gives access to statutory damages and strengthens any later legal claim. The practical operational step is a short checklist that any room can run through at the start of a season. The checklist below captures the common path writers should follow. Rooms that run it once per season tend to catch new scrapes before they settle into a new model.

1. Register every produced script with the U.S. Copyright Office within 90 days of first broadcast.
2. Search the Series Business lookup tool against your show name and episode numbers.
3. Export the row that matches your work and save it with a dated screenshot.
4. File a DMCA takedown notice to the dataset host, keeping the acknowledgement email.
5. File a WGA audit request on any internal studio AI tool trained on your show.
6. Join the WGA AI working group so the dataset evidence gets aggregated.

None of these steps restore the training activity that already happened, which is the hard truth that writers keep returning to. The point of the checklist is to build evidence for the 2026 bargaining position and to make the next cycle of unauthorized use harder to pull off. Writers who skip the paperwork lose standing when the lawsuits begin, which is a lesson the recorded-music industry had to relearn around sampling in the 1990s. The AI authorship conversation around digital manipulation is already repeating that playbook in a new medium. Writers who skip the paperwork lose a lane of standing that is impossible to rebuild once a trained model ships.

Studios Experimenting With Licensed AI Script Partnerships

Shifting focus from writers to studios, several majors have begun piloting licensed AI partnerships rather than relying on scraped corpora. Lionsgate signed a licensing arrangement with Runway in 2024 to train models on the Lionsgate library under a bilateral deal. Paramount has run smaller internal pilots trained only on titles it fully owns. Together, these deals put both studios on cleaner legal ground than their rivals. Those deals pay the rights holder, which in theory flows to the writer under residual language. In practice, the flow is murky because residual language was not written with AI training in mind. The AI transforming Hollywood framing captures the uneven middle ground this creates.

Opt-out licensing marketplaces for AI training scripts are the second piece of the shift. Startups like Created by Humans and Human Native AI act as clearing houses between individual writers and AI developers, offering per-script licenses with transparent payouts and a clear opt-out. These marketplaces are small today, but they establish a per-script price that writers can point to in negotiations, which is useful even when a studio refuses to pay through them. The AI copyright crises during livestreams showed how fast a licensing marketplace can scale once a flashpoint hits. Writers who register with these marketplaces also build a public price history that supports Guild bargaining.

Studios also keep running internal, closed-weight models trained only on their own licensed catalog, which is the quiet default outcome of the current fight. These internal tools do not share weights with public models and do not sit inside Books3 or any public dataset. They give the studio most of the generative benefit without most of the legal risk. Writers find out about them from leaked memos or vendor pitches, which is a bad surveillance posture that will keep driving distrust until dataset audits become a routine disclosure. The YouTube and CAA partnership against unauthorized AI use is an adjacent model for how talent-side agencies can push the audit question forward. Writers can only verify what they can see, and the current closed-tool regime does not show them anything.

The Risk Map: Hallucinated Plots, Dead Characters, and Litigation

Beyond the labor and licensing fight, there is a product risk map that studios keep underweighting. AI-drafted episodes still hallucinate major plot points, resurrect dead characters, and generate dialogue that violates canon. That liability vector is one studio insurance has not fully priced, even though fans and lawyers both track canon closely. A character who dies in season 2 cannot speak in a season 5 flashback written by an AI tool that was trained on a dataset without that continuity signal. Writers rooms catch these canon errors in minutes with a quick table read. A thin room with an AI copilot catches them in days or not at all. The AI versus human creativity argument is already visible as a product-quality gap on some mid-tier streaming shows. Insurance brokers at major studios have begun flagging AI-related product-defect risk as a new line item.

Litigation risk is the second half of the risk map, and it is already stacking inside studio counsel memos. Studios that train on unlicensed scripts carry copyright exposure that will not easily settle. Studios that use retrieval against third-party scripts carry a trade-secret exposure around any sealed show bible that leaks into a prompt. Studios that credit AI output in violation of the 2023 MBA carry a labor exposure that could trigger a room strike. The studios strike back as AI faces lawsuit story captures the ambient risk environment that counsel is now writing policies around. None of these three risk vectors are hypothetical, and all three are stacking on top of a quarterly earnings environment that has very little slack.

Ethics of AI Training on Scripts Beyond the Legal Fight

Shifting to the ethical frame, the legal argument is only the first layer of what writers mean when they call AI training theft. Writers argue that AI training scripts without consent breaks a professional norm older than copyright. That norm is simple: one writer does not use another writer’s text to compete with that writer unless they have paid for it or asked for permission. That norm is embedded in the staff writer system, the WGA registry, and even the way scripts circulate inside a room. The AI training question breaks every one of those long-standing agreements. The dangers of AI ethical dilemmas conversation keeps returning to this foundational breach because it is the clearest explanation for why writers moved so fast from annoyance to anger.

International guilds have been clearer about the ethical frame than the WGA has sometimes been, because they are smaller and more plural. The international screenwriters ethical framework published by the IAWG set five principles that any AI tool should respect, which include consent, attribution, compensation, transparency, and the right to opt out. That framework has traveled further than most union documents because it is written as a norm rather than a contract. The framework has become the baseline reference in European and Canadian negotiations. The United States side has adopted some of the language but has stopped short of ratifying the framework. That gap is a signal of how differently the WGA reads its negotiating leverage.

Ethical pressure on AI training scripts also comes from the audience, which is the other party that goes missing from most coverage. Viewers do not notice AI-assisted drafts until a show feels off, and then they notice loudly on social platforms within a day of the release. That feedback loop is a reputation risk that studios cannot defuse through contract language. Some majors have started tagging AI-touched material to protect themselves from accusations of hidden use. Hollywood’s AI stories losing steam cycle is a reminder that audiences keep track too, even when the headlines move on.

The Future of AI Training on Scripts Through 2030

Looking ahead to 2030, the AI training scripts fight will likely split into three parallel tracks. The first track is licensed opt-in corpora used by the majors under bilateral contracts. The second is closed-weight studio tools trained only on owned libraries that avoid dataset provenance disputes. The third is an ongoing legal fight over models already trained on unlicensed scraped material. The 2026 Minimum Basic Agreement will set the contractual spine for the first two tracks. The Thomson Reuters precedent will keep pressing on the third track through class actions filed by individual writers and estates. That split outcome is messier than either side prefers, yet it is the realistic path given how far the training has already run.

By 2028, the opt-out licensing market for AI training scripts is likely to be a one- to two-billion-US-dollar category. Per-script prices should settle in the 2,000 to 6,000 US dollar range for a produced episode. That price will depend on genre, era, and whether the show is still in revenue. The license will be structured as a lifetime license rather than a term license to keep the paperwork small. Streaming residual precedents will shape the per-use payment layer sitting on top of the lifetime license. The AI versus human creativity argument will take a back seat to the pricing conversation. The pricing conversation is where the actual bargaining will happen through 2027.

By 2030, the audit question is likely to be settled through mandatory dataset disclosure. The disclosure model will likely resemble the EU AI Act’s training-data transparency obligations. Writers’ guilds will cite the EU regime inside American contract talks because precedent now moves across borders faster than it ever did on labor issues. Studios will fight the audit requirement hard, yet they will probably lose it, because the alternative is a decade of asymmetric litigation that none of them wants to budget. The AI ethics and laws trajectory shows the direction of travel, and writers who prepare the paperwork now will be the ones to collect under whatever regime arrives.

Key Insights on AI Training Scripts and the Writers Who Found Their Work

• At least 139,000 film and TV screenplays sat inside one commercial AI training corpus, which the Series Business investigation into AI training scripts confirmed through a cross-match. The investigation used ScriptSlug and OpenSubtitles against writer and show metadata to pin down every title inside that commercial corpus.
• The 2023 WGA Minimum Basic Agreement set the first American contractual floor on AI, yet it still allowed studios to train on covered scripts under narrow consent. The Authors Guild analysis of the WGA AI protections flagged the training loophole in detail within days of ratification.
• The Third Circuit upheld the Thomson Reuters versus Ross Intelligence ruling in 2025, rejecting fair-use defense for commercial AI training on unlicensed material. The AI Weekly report on the appeals ruling framed it as the first appellate test.
• Lionsgate signed a 2024 licensing deal with Runway that paid for library access and placed the studio on clean legal ground. The Hollywood Reporter coverage of studio AI deals described it as a template others now copy.
• A one-hour WGA script fee runs 42,000 to 48,000 US dollars under current MBA floors, which is the baseline every AI tool rollout now disturbs. A 20 percent room cut strips 11,500 to 14,500 US dollars from a mid-career staff writer per season, as the Ankler analysis of writer economics under AI documented.
• International guilds published a five-principle AI ethical framework in April 2024, and the Variety coverage of the IAWG framework tracked its spread. The document crossed borders faster than any union-side position in a decade, which surprised even the Guild staff who drafted the first version.
• The EU AI Act’s dataset transparency obligations began phasing in during August 2025, which puts fresh pressure on training data provenance. The AI ethics and laws analysis on aiplusinfo.com captured why writers now cite European precedent in American contract talks.

Taken together, these numbers trace a clear arc from unauthorized use to contested licensing. The 139,000-script finding was the match that lit the fire, and the Thomson Reuters ruling then changed the legal odds for every dataset that followed. The 2023 WGA deal is best read as a floor rather than a ceiling, because its loopholes gave writers the exact grievance list they will carry into 2026. The economic case is painful at the mid-career tier, where a 20 percent room cut shows up as a lost month of pay. The hit is a line item on W-2 forms rather than a line in a press release. The clear next step is a 2026 bargain that pairs opt-in consent with a real residual, backed by audit rights that make dataset provenance visible for the first time.

A Side-By-Side Comparison of AI Training Script Policies in 2026

Three policy models are now visible in the AI training scripts debate, and each one carries a different cost to writers. The table below reads the status quo, the studio-licensed path, and the proposed opt-in regime side by side. The three columns are not equally mature today, and that uneven maturity shapes every current conversation. Scraped training is the current default across commercial AI copyright models shipped before 2025. Studio-licensed deals are small but growing, which the AI and Hollywood editors cycle anticipated. The opt-in regime is still a proposal rather than a signed clause.

Policy dimension	Scraped training (status quo)	Studio-licensed training	Opt-in writer licensing (proposed 2026)
Consent model	None, scripts scraped without permission	Bulk license at the studio level	Per-writer, per-script opt-in
Compensation flow	Zero to the writer	Partial through MBA residual language	Explicit AI training residual tied to use
Transparency	Opaque, discovered via leaks	Partial, named in the deal	Audit right over the dataset manifest
Legal exposure for studios	High after Thomson Reuters	Lower, bounded by license scope	Lowest, covered by signed consent
Writer control over output use	None	Partial, limited to MBA grievance	Full veto over derivative training
Impact on diversity	Severe, bakes in historical bias	Medium, inherits library composition	Lower, writer selection corrects skew
Litigation risk	Rising class actions	Contained to license scope	Minimal, dispute moves to arbitration

Real-World Examples of AI Script Training Flashpoints

Three flashpoints during 2024 and 2025 show how the AI training scripts fight actually moves in practice. Each case names a studio, a figure, and the limitation that writers still contest. The three examples below reach across licensing, dataset discovery, and internal copilot deployment. Each one landed in trade coverage within weeks of the triggering event. Taken together, they describe the shape of a fight that will keep producing more flashpoints into 2026.

Lionsgate and Runway’s Licensed Training Deal

In September 2024, Lionsgate signed a licensing agreement with Runway and implemented custom video models on the Lionsgate library. The studio reported that pre-production storyboarding time dropped by 30 percent on three in-progress projects after the integration landed in early 2025. Writers inside the Lionsgate pipeline flagged that the deal paid the studio but routed an unclear share back to script authors. The Variety coverage of the Lionsgate Runway deal described it as the first flashpoint for AI training scripts flow-through pay. The deal is operational today, and both companies continue to publish quarterly updates that stop short of per-writer residual numbers. That silence on residual flow is the limitation writers keep returning to in subsequent coverage, and Lionsgate executives say the structure is open to revision.

OpenAI’s Script Dataset Discovery and Writer Reaction

In late 2024, a Series Business investigation matched 139,000 film and TV scripts inside the Books3 and ScriptSlug corpora that trained OpenAI and Meta models. The match was the moment the AI training scripts fight became front-page news in trade publications. OpenAI acknowledged the use of public web data without naming screenplays specifically. Over 400 WGA members filed grievances within six weeks of the publication. The practical outcome was that OpenAI reset its dataset filtering policy for the next training run, which the eWeek coverage of the Hollywood writer protest treated as a quiet concession. The limitation is that the training run that produced the current commercial model cannot be rolled back, which means the harm remains uncompensated.

Netflix’s Internal AI Copilot Pilot

Netflix launched an internal GenAI copilot for story-break meetings in 2024, trained only on titles the studio fully owns. The closed-weight setup is the compromise that counsel inside the majors now treats as the safe option. Early pilots on four scripted shows reportedly shortened the first-draft turnaround by around 18 percent. Writers on the pilot shows reported that the tool was useful for outline passes but weakened dialogue scenes. The Hollywood Reporter coverage of Netflix’s GenAI content generation captured both the productivity claim and the quality complaint. The limitation is that the pilot still runs through union-covered scripts without the explicit consent clause writers want baked into the 2026 MBA.

Case Studies on Writers, Studios, and AI Script Lawsuits

These case studies run deeper than the flashpoints above because the AI training scripts fight is now shaping formal licensing deals and class actions. Each case names a specific problem, the solution attempted, and the limitation that writers still raise. The three cases below span an independent studio deal, a major newsroom lawsuit, and a Canadian guild registry. They are the clearest early signals of where AI training scripts policy is actually settling. Writers inside the WGA track all three closely as benchmarks for the 2026 Minimum Basic Agreement, which the Hollywood versus AI copyright showdown coverage already maps.

Case Study: Alcon Media Rights’ Script Catalog Licensing Pilot

Alcon Media Rights faced a familiar problem in late 2023 with its backlist of 180-plus screenplays sitting exposed in scraped corpora. The company had never licensed any of that work to the AI developers that used it. The solution it piloted was a bilateral licensing structure with three AI developers. The deal paid an upfront 2.4 million US dollar fee for a two-year, non-exclusive training license. A 12 percent revenue share on any commercial product the training enabled was layered on top. The impact was that Alcon posted a 7 million US dollar line item in licensing revenue for 2024, roughly 4 percent of its total top line. The Variety report on Alcon’s AI licensing pilot described it as a template for independent studios. The controversy was that the writers of the licensed scripts did not share directly in the upfront payment, which triggered a WGA grievance now working through arbitration.

Case Study: The New York Times v. OpenAI Overlap With Screenwriters

The New York Times faced the same problem screenwriters raise and filed suit against OpenAI and Microsoft in December 2023. The case alleged unauthorized training on the paper’s archive without any license or compensation. The legal theory maps directly onto screenplay training in the AI training scripts fight. The solution the paper sought was a per-article licensing deal plus damages tied to the training use. The measurable impact so far is that OpenAI entered into licensing deals with other publishers at reported rates of 2 to 5 US dollars per thousand articles over multi-year terms. That rate is a 100 percent increase in licensed-training revenue for the paper versus zero before the suit. The Hollywood Reporter analysis of the NYT OpenAI case drew the direct line to the screenwriter fight. The limitation of this precedent is that it has not yet reached a final judgment on the merits.

Case Study: Writers Guild of Canada’s Early-Opt-Out Database

The Writers Guild of Canada faced a different version of the problem in 2024, where members lacked a visible opt-out mechanism. Canadian labor law did not require AI developers to honor any explicit non-consent signal from a screenwriter. The solution it built was a public opt-out database modeled on robots.txt exclusions and the proposed ai.txt standard. The database lets a writer register their work and signal non-consent to any AI developer scraping public corpora. The impact in the first six months was that 1,400 member writers registered more than 9,600 scripts in the database. That total is a 240 percent increase over the Guild’s initial six-month target of 4,000 scripts. The Variety coverage of the WGC opt-out database launch described it as a technical complement to American MBA negotiations. The limitation is that the database is non-binding on AI developers who choose to ignore the signal, and enforcement still depends on federal copyright legislation in draft.

Frequently Asked Questions on AI Training Scripts and TV Writing