AI Video Contract Terms for 2026: Scope, Rights, Disclosure, and Who Pays the Penalty

Why AI video contract terms are now the compliance layer

On June 9, 2026, New York's synthetic performer law took effect, amending section 396-b of the general business law. It requires any person who produces or creates an advertisement to conspicuously disclose the use of a synthetic performer where that person has actual knowledge of it, backed by civil penalties of 1,000 dollars for a first violation and 5,000 dollars for each subsequent one. The statute runs a few paragraphs. Its operational consequence is much larger: AI video contract terms, not your prompt library, now decide who carries the exposure.

Read the liability map and the reason is obvious. The duty attaches to the party that produces or creates the ad, while media owners who merely disseminate it are carved out. So when a brand, an agency, and a render vendor all touch one AI-generated spot, the question of who produced it and who held actual knowledge gets answered by the statement of work. If the SOW is silent, it gets answered later, by a regulator reading a chain of emails.

None of this is new legal theory. It is the contracting habit that AI video work skipped during the experimental years. Teams that already run {{link}} have internal approvals to point at, but approvals only bind your own staff. A contract is the only instrument that reaches the vendor who rendered the shot and the agency that trafficked it. Most brands discover the difference during the first escalation, when the vendor's position turns out to be that they delivered exactly what the brief asked for, and the brief never mentioned a synthetic performer at all.

Teams that already run an AI video governance playbook have internal approvals to point at, but approvals only bind your own staff.

Scope: describe the AI work in the SOW, not on the kickoff call

Most AI video disputes begin as scope ambiguity. A brief says AI-assisted and the delivered cut contains a fully synthetic on-camera spokesperson. Write the SOW so the categories are explicit: which shots are generated end to end, which are live plates with generated elements, which use a real performer's likeness, and which are product or environment renders with no human figure at all. That last distinction matters, because assets depicting animals, objects, or clearly non-human entities sit outside the New York definition entirely.

Then name the artifacts. A deliverable list for AI video should include the model and version used per shot, the job or seed identifiers, the source of any reference imagery, and a flag on every asset containing a human-appearing figure. Vendors resist this until they price it, at which point it becomes a line item rather than an argument. Put it in the master agreement so you are not renegotiating it campaign by campaign. Ask for the manifest as a delivery condition rather than a reporting nicety, so an incomplete one blocks sign-off instead of surfacing weeks later. It is the only artifact that makes the disclosure decision auditable rather than merely remembered.

Scope language also decides what a change order looks like. Regenerating a synthetic spokesperson because legal wants a different disclosure treatment is not the same unit of work as a color pass, and {{link}} vary enough that the gap can be a factor of ten. Define revision rounds against asset categories rather than a flat count of rounds.

Regenerating a synthetic spokesperson because legal wants a different disclosure treatment is not the same unit of work as a color pass, and AI video pricing models vary enough that the gap can be a factor of ten.

Synthetic performers and digital replicas need separate clauses

These two things get conflated constantly, and they answer to different rules. A synthetic performer, under the New York definition, is a digitally created asset intended to create the impression of an audiovisual performance by a human who is not recognizable as any identifiable natural performer. A digital replica is close to the opposite: a highly realistic representation that is readily identifiable as a specific person's voice or likeness. One triggers a disclosure duty. The other triggers consent, compensation, and contract-enforceability rules.

California codified the replica side in Labor Code section 927. A provision allowing the creation and use of someone's digital replica in place of work they would otherwise perform is unenforceable for new performances fixed on or after January 1, 2025, unless it carries a reasonably specific description of the intended uses, or the individual was represented by counsel who negotiated the replica licensing with commercial terms stated clearly, or by a union whose collective agreement expressly addresses digital replicas. All media, all uses, in perpetuity does not survive that test.

Practically, that means two annexes rather than one. The replica annex carries the specific-use description, the term, the territories, the compensation model, and the revocation triggers. The synthetic annex carries a representation that no identifiable person's likeness was used, plus the vendor's warranty on reference and training material. Anyone building this from scratch should start from a {{link}} checklist and layer the jurisdiction-specific language on top.

Anyone building this from scratch should start from a rights-safe AI video checklist and layer the jurisdiction-specific language on top.

Disclosure: name the party, the wording, and the placement

New York requires the disclosure to be conspicuous and then declines to define conspicuous. It does not specify content, placement, duration, or type treatment. That silence is a contracting problem, not a loophole. Decide in writing who drafts the disclosure copy, who approves it, who burns it into the master versus adding it at trafficking, and what happens when a platform format strips overlays. Assign it to a named role, not to the parties. Then write down the fallback: if a placement cannot carry the agreed treatment, either the placement is dropped or an approved alternative wording applies, and one named person makes that call.

The carve-outs belong in the contract too, because they are where mistakes happen. The New York duty does not reach audio-only advertisements, nor cases where AI is used solely to translate a human performer's language, nor advertising for expressive works where the synthetic performer's use matches its use in the underlying work. Each exception is narrow enough that someone has to apply it per asset, and the SOW should say who that is.

Multi-market work compounds the problem. The European Commission's own summary is that the AI Act's transparency rules come into effect in August 2026, requiring generated content to be identifiable and deep fakes to be clearly and visibly labelled. One global master with per-market disclosure variants is cheaper than five regional recuts, but only if the {{link}} process runs before the edit locks rather than after.

One global master with per-market disclosure variants is cheaper than five regional recuts, but only if the synthetic performer clearance process runs before the edit locks rather than after.

Indemnity, warranties, and the penalty math

Run the arithmetic before negotiating the cap. One thousand dollars for a first violation and five thousand for each subsequent one sounds survivable until you ask what counts as a violation. A campaign shipping forty cuts of one synthetic spokesperson across six placements is not obviously a single violation. The New York statute creates no private right of action and enforcement runs through civil penalties, so the realistic downside is a regulator's arithmetic plus the cost of pulling and relabelling live media.

So the indemnity has to be specific. Ask the vendor to warrant that they will identify every synthetic performer and every digital replica in the delivery, that reference material was lawfully sourced, and that they will notify you when a model or pipeline change alters those answers mid-flight. Tie indemnity to breach of those warranties rather than to a general compliance-with-law clause, which in practice indemnifies nobody.

Then check whether anything stands behind the promise. A render shop with a thin balance sheet cannot absorb a recall of paid media, and standard policies were not drafted for generated content. The {{link}} is wide enough that a signed indemnity from an undercapitalized supplier is a comfort document rather than a remedy. Require evidence of coverage, or accept that the brand is self-insuring and price the work accordingly.

The AI liability insurance gap is wide enough that a signed indemnity from an undercapitalized supplier is a comfort document rather than a remedy.

Records: what you need to prove twelve months later

Disclosure duties turn on knowledge, and knowledge has to be provable after the people involved have moved on. The contract should require a delivery manifest per asset: model and version, generation date, whether a human-appearing figure is present, whether an identifiable likeness was licensed and under which annex, and the disclosure text that shipped with it. Store it somewhere finance and legal can reach, not in an editor's project folder.

Set a retention period longer than the campaign. Exposure escalates on subsequent violations, which makes your own record of the first correction genuinely useful evidence of process. Give yourself audit rights against the vendor's logs, with a defined notice period and a defined scope, and require the logs to survive termination. Scope the audit narrowly enough that the vendor can price it and broadly enough that it actually reaches the per-shot records. Those same logs are what let you answer a regulator in a week rather than a quarter. Vendors accept all of this far more readily during onboarding than after a dispute.

None of this needs a new legal department. It needs four documents: a master agreement carrying the warranties and audit rights, a scope annex that classifies every asset type, a replica annex with specific-use language, and a manifest template the vendor fills in on delivery. Teams that put those in place before the next flight will spend the rest of 2026 shipping. Teams that do not will spend it reconstructing who knew what, from a chat thread, under a deadline.