Union AI Clauses Compared: What WGA, SAG-AFTRA and CWA Contracts Actually Say

We read the AI terms in 10 union contracts reached between late 2023 and mid 2026, and they split cleanly in two. All 5 entertainment agreements give an individual worker a consent right or a pay protection tied to AI. None of the 5 Communications Workers of America (CWA) tech and service contracts does, on the union's own published summaries. Union AI clauses are spreading into software work, but the version arriving there is notice and bargaining, not consent and pay.

The short answer. Union AI clauses come in four kinds: bans, consent requirements, payment triggers and notice-and-bargain duties. Hollywood contracts use all four. CWA contracts covering software, telecoms and digital agency staff mostly use the last, with one outright ban on AI surveillance. If you run a delivery team, notice and bargaining is the term to plan for first.

Four kinds of union AI clause, and why the type matters

Every AI term we found in a ratified union contract falls into one of four types. The type decides who holds the power once a dispute starts, so it matters more than the wording.

A ban or restriction stops the employer from doing something at all, or only under a stated test. A consent clause hands a veto to the individual worker, not the union. A payment trigger means AI use creates a fee, or at least cannot be used to cut one. A notice-and-bargain clause requires the employer to tell the union and negotiate over the effects, but leaves the decision with the employer.

The last type is the weakest for the worker and the cheapest for the employer. It is also, by a wide margin, the most common outside entertainment. My view is that most commentary on union AI deals blurs these four together. That is why the spread to tech looks bigger than it is.

WGA AI contract terms: AI is not a writer

The Writers Guild of America made the first enforceable AI deal in the US entertainment industry. Its 2023 Minimum Basic Agreement (MBA), the master contract with the studios, ran from 25 September 2023 to 1 May 2026. Members ratified it with 99% voting yes, ending a 148-day strike.

The 2023 rules: status, consent and disclosure

The core move was definitional. In the Guild's own words, "no written material produced by traditional AI or GAI can be considered literary material". That one line does most of the work, because the MBA's pay and credit system attaches to literary material.

It means a studio cannot hand a writer an AI draft and pay a rewrite fee. AI material given to a writer "doesn't count as assigned material (for compensation purposes) or source material (for determining credits)," according to the WGA's guide to the AI provisions. The writer is treated as the first writer and paid that way.

Consent runs in both directions. A writer may use AI if the company consents and its policies allow it. But "the company can't require the writer to use AI software (e.g., ChatGPT) when performing writing services". The company must also tell the writer about AI in what it hands over. It must disclose if any materials given to them "have been generated by AI or incorporate AI-generated material".

Training was left open. The Guild only "reserves the right to assert that exploitation of writers' material to train AI is prohibited by MBA or other law". That is a placeholder for a future fight, not a protection.

The 2026 MBA adds a training-licence notice

The successor deal runs from 2 May 2026 to 1 May 2030, and members ratified it in April 2026 with 90.38% voting yes. It keeps every 2023 protection and adds one new term on training.

Per the WGA's summary of the 2026 deal, companies "must give the Guild written notice if they license writers' work to train a commercial GAI system in order to create outputs". The Guild can then request discussion "including any remuneration for writers".

Read that closely. It is a notice duty and a right to talk. There is no consent requirement and no rate. I think writers got less here than the headlines implied, and the fight over training money has been pushed to 2030. Courts are moving faster, as the tally of AI copyright cases over training data shows.

SAG-AFTRA AI provisions: consent and pay for a likeness

SAG-AFTRA's problem was different. A writer's work is text. A performer's work is a face, a voice and a body, and a digital replica can copy all three. So the union bargained over likeness, not authorship.

The 2023 TV/Theatrical contract, ratified in December 2023 by 78.33% of voting members, built the framework. For a replica made with the performer's participation, the deal requires "explicit signed/initialed consent" both to create it and to use it, according to a Loeb & Loeb summary of the AI terms. Compensation terms are set out in detail for each use.

For a replica built without the performer, consent and bargaining are required unless the use matches what was scripted or performed. If the performer is dead, consent can come from the estate, or from the union if no estate can be found. First Amendment uses such as parody, docudrama and biography are exempt.

Synthetic performers, meaning AI-generated characters that copy no specific actor, got the lightest touch. The studios owed "Notice to Union and an opportunity to bargain in good faith over appropriate consideration, if any" when a synthetic replaced a role a human would have filled.

The 2026 deal adds a value test for synthetics

The 2026 contract tightened the synthetic rule. It was ratified in June 2026 by more than 90% of votes on roughly 19% turnout, according to the Associated Press. Producers may use an AI performer only if it brings "significant additional value" over a live actor or a digital capture of one.

Producers using a synthetic must notify and bargain with the union and show that value. According to TheWrap's May 2026 report, the union can arbitrate for damages "not necessarily be limited to" what a human performer would have earned. TheWrap also reported that the parties did not agree a definition of "significant additional value". That gap is where the first arbitration will land.

The video game agreement priced real-time replicas

The most detailed SAG-AFTRA AI terms are not in the film contract. They are in the Interactive Media Agreement, ratified in July 2025 by 95.04% of voters after a 320-day strike that TheWrap described as spurred solely by AI.

Companies must tell performers all the ways AI is planned to be used on their "voice, likeness and movements" and get written consent. Performers can suspend that consent during a strike. Minimum rates apply to replica use, and a replica used for real-time generation, such as a chatbot in a game, earns at least 7.5x that minimum. Developers must also send consenting performers a report on how the replica was used and how pay was calculated.

That usage report is the clause I would watch most closely. It is the first union term I found that forces an employer to produce a per-person audit trail of AI use. It is also the one most easily copied into software contracts.

Union AI clauses compared side by side

The table quotes the operative language for each contract, sorted by clause type. Where a union publishes only a press summary, the quote comes from that summary and says so.

WGA quotes come from the Guild's own guide and deal summary. SAG-AFTRA and CWA quotes come from law firm, press and union summaries because the full contract text was not available to us. "None found" means absent from the published summary, not proven absent from the contract.

CWA AI bargaining: notice, not consent

CWA represents telecoms, media, video game and digital agency workers. It is the union most likely to show up at a software or services company. Its AI terms are the ones most readers here should study.

ZeniMax set the template: principles plus a duty to bargain

In December 2023, 376 members of ZeniMax Workers United-CWA won a tentative agreement with the Microsoft subsidiary on AI. ZeniMax committed to six principles: AI at work must be fair, reliable and safe, private and secure, inclusive, transparent and accountable.

The enforceable part is narrower. Per the CWA release, ZeniMax agreed "to provide notice to the union in cases where AI implementation may impact the work of union members and to bargain those impacts upon request". The first full ZeniMax contract, reached in May 2025, folded that agreement in. CWA says other Microsoft units have contract articles requiring notice "when the implementation of AI technology impacts work performed by bargaining-unit employees".

Notice what is missing. No tester can refuse an AI tool. No AI deployment triggers a payment. The union gets a seat at the table after the employer has decided, which is the standard shape of effects bargaining in US labour law.

ActBlue, Frontier and the hotline deals widen the pattern

ActBlue staff in CWA Local 1400 ratified their first contract in July 2024 with 95% support. It includes what the union called "a first-of-its-kind prohibition on any kind of surveillance of staff using artificial intelligence". That is the only outright ban in the CWA set.

At Frontier Communications, a 3-year contract ratified in November 2025 includes "a structure for addressing the implementation of artificial intelligence tools". At the National Abortion Federation's hotline, workers ratified a 3-year deal in July 2026 after a 24-hour strike. They won "protections against unilateral AI implementation". MissionWired is a digital fundraising agency with about 136 workers in the unit. Its May 2026 first contract lists "guardrails and protections" on AI use.

None of the public summaries says what those structures or guardrails require. That vagueness is itself a finding. When a union can quote a clause it usually does, so I read the silence as a sign the terms are procedural.

Why union AI clauses split between output and process

The split is not an accident of bargaining skill. It follows from what each group of workers sells.

A screenwriter or an actor sells a specific, identifiable work product. A script has an author and a replica has a face. Both slot into a residuals and credit system the unions have policed for decades. Consent and pay clauses bolt onto machinery that already exists.

A QA tester, a fundraising analyst or a telecoms technician sells hours of labour, not a named asset. There is nothing for AI to copy that the worker owns. The only lever left is control over how work changes, which is why those deals focus on notice and bargaining.

There is a second reason. Entertainment unions struck for 148, 118 and 320 days to win their terms. The CWA units are mostly small and new, bargaining first contracts where wages and job security come first. I would expect consent-style terms in tech only where workers sell something closer to a likeness. Think voice talent, creative staff, or people whose output trains a model.

That is not a theory about the future. The NABET-CWA unit at radio producer Snap Judgment already negotiated "protections from transfers of their work to artificial intelligence models," per CWA's March 2026 roundup. Creative workers inside tech-adjacent employers get output terms. Everyone else gets process terms.

Where this comparison is weakest

The table looks precise. The evidence under it is uneven, and you should weigh it that way.

Summaries are not contracts

Only the WGA publishes plain-language AI terms close to the contract text. SAG-AFTRA's own documents blocked automated access during our research, so its terms here come from a law firm summary, the AP and TheWrap. CWA publishes press releases, not clauses. A 0 in the CWA consent column means "not in the summary". The full ZeniMax or Frontier contracts could contain more, and I would revise the count if they did.

The sample is also chosen, not random. It covers the contracts with the most public AI language. CWA says 58 NewsGuild-CWA contracts carry AI terms, and newsroom contracts may well contain consent or byline rules that would weaken the clean split shown here.

Enforceable is not the same as enforced

No public arbitration award has yet tested the 2026 "significant additional value" rule, and the parties never defined it. Effects bargaining in tech has the same problem. A duty to bargain "upon request" is only as strong as a union's ability to strike, and first-contract units rarely strike over a single AI rollout.

The strongest counter-case is simple. A notice-and-bargain clause may achieve more in practice than a consent clause. It covers every AI deployment across a workplace instead of one replica at a time. On that reading, the CWA approach is wider rather than weaker, and my framing overstates the gap.

Which union AI terms will reach software and service teams

If your team is organised, or might be, these are the terms most likely to appear in a first contract. The ranking is my judgement from the pattern above, not a measured forecast.

The surveillance ban deserves more attention than it gets. Employers are adding AI scoring to productivity tools faster than the law is setting limits, a gap covered in where the legal line on AI employee monitoring sits. A contract ban is far easier to enforce than a statute, and it costs a union very little to ask for.

Notice terms also overlap with state law. Several states now require disclosure around AI-driven job cuts, as the analysis of Connecticut's AI layoff disclosure rules shows. A union notice clause and a statutory notice duty can land on the same rollout. Draft one internal process that satisfies both.

Finally, honest communication is cheaper than any clause. Workers organise when AI changes arrive without warning or explanation. Some companies announced AI cuts and then walked them back. Their record, in the review of AI layoff statements, is a good guide to what not to say.

Frequently asked questions

How to prepare before a union asks

Start with an inventory, not a policy. List every AI tool that changes how your team's work is assigned, measured or reviewed. Write one line per tool on whose work it affects. That list is exactly what a notice clause would require you to produce, and you will want it either way.

Then take one tool off the list: any AI feature that scores individual workers. Ask whether you would defend it in front of your team. If the answer is no, switch it off now, before it becomes the first item in somebody else's bargaining demands.

Related reading