How does a studio, network or streamer insure AI indemnity?

Why does AI risk land on the studio or network?

Because claims about content name the people who published it. A copyright owner can sue anyone who reproduces, distributes or publicly performs the work, and the company whose name is on the title is usually the one most worth suing. Commissioning the work from a vendor does not move that exposure. It gives you a contract claim against the vendor.

So AI indemnity is a chain, not a clause. The tool vendor indemnifies its own customer, if anyone. The production vendor indemnifies you, up to its cap. You indemnify the platform or licensee you deliver to. Each promise is only worth what stands behind it, and the last link is your own program.

The tool’s promise rarely reaches a studio. It runs to the account holder, which is usually the vendor rather than you, and most exclude edited output and anything made from material the customer supplied. Production work is almost all editing and combining. Which tools indemnify at all, and on which plans, is on our AI indemnity guide.

What do the 2026 guild agreements require about AI?

All three major guilds signed new four-year agreements in 2026, and each keeps or adds AI terms. SAG-AFTRA’s TV/Theatrical Contracts, ratified on June 4, 2026, build on the 2023 consent and compensation rules for digital replicas. Producers now need an articulable business reason to scan a performer, may not use a performer’s replica during a strike where consent would be required, and may use a synthetic performer in a role a human would otherwise fill only where it brings significant additional value.

Three of the new terms land squarely on a buyer. A producer stays responsible for the replica protections even after handing a replica to someone else, unless the union approves the new owner. Replicas have to be kept secure, with access limited to people who need it. And from pictures starting principal photography on July 1, 2027, dubbing a performance into another language with a replica needs the performer’s consent. If a synthetic is used in breach, the union can arbitrate for damages that are not capped at what a human performer would have been paid.

The WGA’s 2026 agreement keeps the 2023 rules: AI is not a writer, AI-generated material is not literary or source material, and a writer must be told if material handed to them was generated by AI. It adds written notice to the Guild if a company licenses writers’ work to train a commercial generative AI system. The DGA’s agreement reaffirms that directors’ work is performed by a person and gives the director control over AI-generated footage.

For insurance, the guild terms matter in two ways. A claim that a replica was used without the consent or payment the agreement requires is usually a contract and labor matter, which media liability policies are not built to pay. And the consent records, the business reason for each scan and the notices you gave are what an underwriter will ask to see before it covers AI likeness and voice. Read the agreements themselves, not this summary, before relying on any term.

Which laws make voice and likeness claims easier to bring?

California is the one most productions meet first. Since 2025, a contract term letting a digital replica stand in for work a person would have done in person is unenforceable unless it describes the intended uses reasonably specifically or the person had a lawyer or a union. A replica used under an unenforceable term is a likeness claim waiting to happen. A replica of a deceased performer in an audiovisual work needs the estate’s consent, with liability of at least $10,000.

New York’s right of publicity statute already reached voice. In 2025 a federal court let voice actors’ claims proceed against a company that sold clones of their voices. New York also requires a label on advertisements that use a synthetic performer, from June 9, 2026, though films, television and streaming are generally exempt, so the exposure sits in the trailer and marketing rather than the show.

Outside the United States, the EU AI Act requires deepfakes to be disclosed from August 2, 2026, with a lighter duty for evidently creative or fictional work. Federal legislation, the NO FAKES Act, would create a national digital replica right and was advanced by the Senate Judiciary Committee in June 2026, but it is not law.

Will your existing media, E&O and general liability respond?

General liability was never the answer for content. The standard ISO form excludes personal and advertising injury for a business that broadcasts, publishes or telecasts, and since January 2026 ISO has offered generative AI exclusions as well. If one is attached to your general liability, check whether your umbrella follows it.

Media liability is where the claims live, and every media policy now sits in one of three places on AI: affirmative, silent or excluded. Some insurance companies have added affirmative wording; CFC did for its Media policy from July 30, 2026. Others have added exclusions at renewal. Ask for the AI wording by form number on the primary and on every excess layer, because an exclusion on one layer and not another leaves a hole in the middle of the tower.

Production E&O has a quieter problem. The application asks about title reports, copyright reports, script clearance, releases and music, and many applications were written before generative AI. The application is a set of representations. If a production used AI and the form did not ask, say so anyway, describe how it was cleared, and get the insurance company’s answer in writing, rather than learning about the question when a claim arrives.

How do large buyers structure AI coverage?

Usually in two layers of responsibility. The company buys an annual media liability program for everything it produces, acquires and distributes, with a retention it can carry and excess layers above it. Each production, especially one bought from an outside producer, carries its own E&O naming the company as additional insured, kept in force for a set period after delivery. The AI question has to be answered on both.

On the program, read four things for AI. The AI wording, on every layer. Any AI sublimit or separate AI retention, which would mean the headline limit is not the AI limit. The trigger and the start date for past work, so titles already delivered with AI are not outside the policy. And whether defense costs sit inside the limit, because an infringement case can spend a large part of a limit before damages are discussed.

Content claims are not the only AI exposure a studio has. A company that runs AI systems of its own, such as a dubbing pipeline, a recommendation engine or an AI agent that handles talent or customer data, has a different risk: a system that makes an error or produces harmful output. Standalone AI liability policies exist for that. Armilla, a Lloyd’s coverholder, offers one with limits up to $25 million. It sits beside the media program, not inside it.

What do underwriters want to see before they write AI content?

Evidence that AI use is governed rather than incidental. On the quotes we requested for AI studios in 2026, underwriters asked for a written description of how the studio confirms that AI-generated content does not infringe anyone’s rights, including the human review steps, and for the clearance process for names and likenesses. A studio or network will be asked the same, at a larger scale.

Expect questions about which AI tools are approved, on which plans, and who approves AI use on a production. Expect to show a disclosure log for a sample title: the tool, the inputs, the date, the human edits and any licensed material behind each AI output. Expect questions about replica consents and guild compliance, about the indemnity and insurance terms in your vendor agreements, and about where your content is distributed.

Chain of title has a new wrinkle. Purely AI-generated material has no human author and cannot be registered for copyright, a rule the Supreme Court left in place in March 2026. A chain of title that shows the human authorship in each title, and records where AI was used, answers both the underwriter and your own distribution partners.

What should a studio require from AI tool vendors and production vendors?

From an AI tool vendor on an enterprise agreement: an IP indemnity that covers the models you actually use, including video and voice, sitting outside the liability cap; a promise not to train on your material; and an answer on what insurance stands behind the vendor’s indemnity. Most creator plans offer no indemnity, so a vendor working for you on one is relying on its own insurance instead.

From a production vendor, post house or AI studio: media liability with affirmative AI wording, or written confirmation that no AI exclusion is attached, at a limit that matches the indemnity cap you agree. You and your affiliates as additional insured on that policy by endorsement. A start date for past work that reaches the first day of the engagement, and the policy kept in force, or an extended reporting period bought, for a set period after delivery. And a disclosure log for every deliverable.

The vendor will push back on three points, and each is reasonable. A knowledge qualifier on the originality warranty, because nobody can know what a third-party model was trained on. Carve-outs for material you supplied, your edits and your instructions. And a cap tied to their policy limit. Accept those and ask for the insurance to match, and the indemnity you receive is one an insurance company stands behind.