SAG-AFTRA’s AI Rules Just Took Effect (And They Don’t Cover Where Most Body Swaps Actually Happen)

Six weeks ago, on July 1, the AI enforcement provisions in SAG-AFTRA’s 2026 TV/Theatrical Agreement moved into their active phase. Members had ratified the contract in early June. Since then, a studio working under that agreement cannot generate a digital replica of a performer and drop it into a scene without going through a specific process first.

It is a real change, and it took two strikes and three years to get. It also covers a narrower slice of production than most coverage of it suggests.

What actually changed on July 1

The core of it is consent, defined tightly enough to be enforceable.

A producer who wants to create and use a digital replica has to get the performer’s agreement in writing, separately signed or initialed rather than buried in a stack of paperwork. The performer has to be given at least 48 hours notice before the services are needed. The employer has to describe what the replica will be used for with reasonable specificity, not just ask for a blanket release.

Two additions in the 2026 round matter more than they sound. The agreement extends protection to what it calls no-scan replicas, meaning a likeness assembled from existing footage rather than from a dedicated body scan session. Under the older language, an actor who had never sat for a scan could argue they were less protected. That argument is now unnecessary.

The other change closed a gap from 2023: producers could previously use a replica of a performer’s voice for foreign-language dubbing without asking. That exception is gone.

Who it does not cover

Union agreements bind signatories. That is the whole mechanism, and it is also the limit.

A SAG-AFTRA agreement governs productions that have signed it. It does not govern a non-union commercial shoot, a brand’s in-house social team, a YouTube channel, a corporate training video, or somebody making content in a bedroom in another country. Those are not loopholes in the contract. They are simply outside its jurisdiction, which is a different problem and a much larger one by volume.

Federal law has not filled the space either. The NO FAKES Act of 2026 has been introduced but not passed. State-level rules in California, Illinois and New York are moving, and law firms advising production clients are already drafting non-union digital replica releases in anticipation. Last week the EU AI Act’s transparency obligations came into force, requiring disclosure of AI-generated content, which touches labeling rather than consent.

So the protections are real, and they apply to a fraction of the video being made.

Meanwhile, on the tools side

The reason this gap matters now rather than in five years is that the technical barrier has collapsed.

Replacing a person’s entire body in a video clip, tracked across frames rather than pasted on, used to require a compositor and a schedule. iMideo, an AI video editing tool, runs an ai body swap video feature that takes a reference image and a source clip and replaces the body throughout. It accepts MP4 or MOV files up to 60MB, handles clips between 5 and 120 seconds, and outputs at 360p, 540p or 720p depending on whether you are testing an idea or delivering to a client. Most jobs finish in a few minutes.

Note the ceiling: 720p is the top output. That is adequate for social platforms and client review, and it is nowhere near a theatrical deliverable. The tools that concern SAG-AFTRA at the studio level are not these tools. But the volume of video that shapes what audiences see is not theatrical either.

Where the responsibility actually sits right now

Read the terms on almost any of these tools and you find the same structure. iMideo’s documentation says results can be used for personal and commercial projects, then adds that you must have the rights to the reference image and source video you upload, and that you are responsible for what you create.

That is not evasion. It is accurate. The tool cannot know whether the person in your reference image agreed to be there. No automated system can verify consent, because consent is a fact about a conversation that happened somewhere else.

Which means that for everything outside a union agreement, the enforcement mechanism is you. Not a contract, not a regulator, not the platform.

Consider how this plays out in practice. A brand hires a small agency to make six social spots. The agency licenses a stock clip of a person walking through an office, swaps in a body that matches the brand’s casting, and delivers. Somewhere in that chain, four separate parties each assumed one of the others had handled permissions. The stock license covered the footage. It did not cover replacing the performer in it, and nobody involved was under a union agreement at any point.

That scenario is not hypothetical or unusual. It is the ordinary shape of small-scale commercial production, and it is exactly the shape that current rules do not reach.

What a production should do now, union or not

The practical answer is to adopt the union standard voluntarily, because it is well designed and because the alternative is inventing your own and getting it wrong.

Get consent in writing, as a separate document rather than a clause inside a general release. Describe the specific use, not a category. If the replica will appear in an ad, say which campaign. Give people notice rather than asking on the day. Keep the paperwork somewhere you can find it in two years, when someone asks.

For anything running publicly, check whether disclosure is required in your market. The list is growing, and after last week the EU is on it.

None of that is expensive. It costs a document and a conversation, and it puts you on the right side of a line that is currently being drawn in several jurisdictions at once.

The mistake worth naming is the one that looks like diligence. A general release signed at the start of a shoot, with a broad clause about likeness and derivative works, feels like coverage. It is the thing union negotiators spent two rounds specifically writing out of the agreement, because a performer signing a stack of documents on set is not meaningfully consenting to a use nobody described to them. If your paperwork works the way the 2023 contract worked, you are behind by two revisions and both of them exist for a reason.

The second one is treating the reference image as the only asset with rights attached. Both files matter. The clip you are modifying has a performer in it too, and their agreement to appear in that footage was not an agreement to have someone else’s body put on their performance.

And if you are the one being swapped

Everything above is written from the production side, which is where the decisions get made. The other side has fewer options and should know what they are.

Union members have a route: the agreement is enforceable, and the union processes claims. That is the entire value of the thing, and it only exists because two strikes paid for it.

Outside that, the leverage is almost entirely in the contract you sign before the shoot. Read what the release says about likeness, derivative works and modification, and ask what happens to the footage after delivery. A producer who intends to behave well will answer that question without difficulty. One who hesitates has told you something.

After the fact, the options thin out fast. State likeness statutes exist in some places, platform reporting exists everywhere and works unevenly, and both are slow. None of it restores a video that has already circulated for a week.

The asymmetry is real and worth naming: the person whose body appears in the clip usually has the least say in what happens to it.

The part worth watching

The interesting question is not whether the union rules work. Inside their jurisdiction they probably will.

It is whether the standard they set becomes the industry default outside that jurisdiction, the way union safety practices eventually shaped non-union sets. The law firms already drafting non-union replica releases suggest that process has started. Whether it finishes before the technology gets another order of magnitude cheaper is the actual open question, and nobody in this business has a good record of predicting that timeline.