The writers’ strike showed how to negotiate AI before using it
The 2023 dispute showed how to turn AI into labor rules: define the work, protect credit and pay, control data and require evidence.
Hollywood’s writers’ strike, which began on May 2, 2023, brought a new question into labor bargaining: what happens when a tool does more than assist writing and can alter who receives the assignment, credit and compensation? The answer could not come from deciding whether ChatGPT wrote good dialogue. Technical risks had to become enforceable terms.
The method travels beyond film. When a company introduces AI, a promise of “human oversight” is not enough. A useful agreement defines protected work, separates assistance from substitution, preserves attribution, states which data may feed the system and creates a way to inspect and challenge its use.
What was disputed on May 11
The Writers Guild of America strike announcement said that six weeks of bargaining with Netflix, Amazon, Apple, Disney, Warner Bros. Discovery, NBCUniversal, Paramount and Sony had failed to produce an agreement. The WGA argued that company responses were inadequate for a streaming economy that had reduced the continuity and compensation of writing work.
AI was one front in the conflict, not its entire cause. The WGA proposal and AMPTP response chart dated May 1 also covered minimum pay, residuals, employment duration, room size and pensions. Treating the strike as a purely technological revolt would erase the economic structure that made the technology consequential.
On AI, the union proposed regulating its use on projects covered by the agreement: it could not write or rewrite literary material, its output could not qualify as source material, and covered material could not be used to train it. The chart recorded that the AMPTP rejected the proposal and offered annual meetings on technological advances.
The difference concerned contractual force. A meeting enables discussion; a clause enables a claim of noncompliance. “We will discuss technology every year” also says nothing about what happens tomorrow if a company hands a writer an automated draft, reduces the fee or adds a script to a training set.
First define the protected unit
“AI will not replace people” is too broad to enforce. An agreement needs nouns and verbs: script, treatment, story, rewrite, adaptation; generate, deliver, revise, train. It must also identify who is covered and the production stage at which each obligation begins.
The distinction between literary material and source material mattered because those concepts organize credit and compensation. If a company could label machine-generated text as a “source” and hire a person to adapt it, the technical label could devalue human labor even when that person solved the structure, characters and every usable scene.
The practical test is to trace a real assignment. Who received the first instruction? What material were they given? Which parts did they change? How was the result classified? If the contract cannot answer that sequence, an abstract debate over whether a machine “creates” will hide the labor question: who performs recognized work and at which rate.
Labor credit and copyright are not the same
Two legal systems can use similar words while answering different questions. Screen credit follows professional and contractual rules that affect pay and careers. Copyright determines which expression may receive protection and who may claim it. A finding that an automated output cannot be registered by itself does not decide what a person must earn for transforming it.
The United States Copyright Office guidance, effective March 16, 2023, required human authorship for registration. It set out a case-by-case inquiry: distinguish whether technology was an assisting instrument or determined the traditional expressive elements. Applicants also had to disclose AI-generated content included in a submission.
The guidance recognized that sufficiently creative human selection, arrangement or modification might be protected, while protection would not extend to the automatic material by itself. It did not decide the use of protected works in model training; it announced a later study. Presenting it as a complete answer to Hollywood’s conflict would give the document claims it did not make.
The transferable skill is keeping four columns separate: legal authorship, professional credit, compensation and permission over data. A person may deserve contractual pay for revising material even when part of the result receives no copyright; a company may own rights in a script without that settling every training use of the text.
Assistance, substitution and bargaining power
OpenAI had introduced ChatGPT on November 30, 2022 as a research release. Its own documentation warned that it could produce plausible but incorrect answers, change with small variations in phrasing and guess intent when a question was ambiguous. Those limitations mattered, but they did not eliminate the labor problem.
A tool need not finish a film to change a labor market. It only needs to produce cheap drafts that redefine the assignment: instead of hiring someone to write, a company hires them to “polish.” Final quality may still depend entirely on the person while the administrative category reduces time, pay or credit.
Classify use by effect, not interface. Correcting spelling under the author’s control is assistance. Suggesting variants can remain assistance if the person retains choice and responsibility. Generating a treatment that the company imposes as the starting point changes the assignment. Using accumulated works to build a tool that produces substitutes raises a separate question about data and bargaining power.
Individual consent does not always solve a collective asymmetry. If refusing the tool means losing the job, clicking “accept” does not prove a free choice. A collective agreement creates a common floor so one person’s decision does not lower everyone else’s conditions.
Five clauses that make a promise auditable
First, a functional definition that covers present and future systems without depending on a brand. Second, permitted and prohibited uses by stage, with a duty not to reduce credit or compensation. Third, disclosure of the origin of material supplied to the professional, including any automatically generated portion.
Fourth, data rules stating which works may be used for training, fine-tuning or evaluation, on what basis and for how long. Fifth, an enforcement mechanism that preserves records, enables audits and provides a remedy. Without evidence and consequence, a prohibition can remain a declaration.
Periodic review and a rule that operates between reviews are both necessary. Technology changes, so an annual table can help update definitions and risks. But it does not replace the contractual floor in force while the next meeting approaches. Adaptability and enforceability are not alternatives.
How to read any AI labor agreement
Apply a short test to any workplace policy. Replace “responsible AI” with the concrete action: supplying a draft, assessing performance, training on documents or assigning a shift. Identify who authorizes it, who must be informed, which record remains, how it affects pay and credit, and how an error can be challenged.
Then find the gaps. “With a human in the loop” does not say whether that person has time, information or power to refuse. “Authorized data” does not identify who granted permission. “No replacement” does not prevent reducing a full role to a worse-paid correction task. Every reassuring adjective needs an observable condition.
On May 11, 2023, the outcome of the negotiation was unknown. The durable lesson was already visible: bargaining over AI is not mainly a contest over whether a machine writes like a person. It is a way to protect the relationships the tool may reorder. Definition, attribution, compensation, data and verification turn a promise about the future into a rule for work today.
This article was produced with artificial intelligence under human editorial oversight.
