AI Video Labeling Rules in 2026: A Working Briefing for Marketers

AI content labeling rules became enforceable in Europe on 2 August 2026. California's provenance requirements became operative the same day. New York's rule on synthetic performers in advertising started on 9 June 2026. The UAE began requiring advertiser permits on 1 February 2026.

None of this bans AI video. All of it says the same thing in different words: if it looks real, say it is not.

This is a working briefing for the people who actually ship the content: what changed, what applies where, which videos carry real risk, and what it means for agencies and the companies that hire them.

Start here: does this specific video need a label?

Most of what follows is context. This is the part you will actually use, so it goes first. Answer for one video you are about to publish.

The rules, market by market

European Union

The transparency obligations under Article 50 of the AI Act apply from 2 August 2026. Three things matter more than the rest.

It follows the audience. If your content reaches people in the EU, the rules apply regardless of where your company sits. A Karachi agency posting for a Dubai client whose customers include Europeans is inside scope.

Two separate roles. The tool maker embeds hidden, machine-readable marks. The publisher, meaning the business posting the content, tells the audience. You cannot discharge your duty by pointing at the tool's hidden mark. The Commission is explicit: the disclosure must be perceptible without special software.

Realism is the trigger, not deception. Content that resembles real people, places, objects or events must be disclosed even where nobody intended to mislead and even where no real individual is depicted. Purely cartoon or obviously stylized work generally falls outside. Artistic, satirical and fictional work gets a lighter touch: disclose, but not in a way that wrecks the piece.

Minor AI assistance for ordinary editing is not caught. Color correction is not the same as generating a scene. Content published before 2 August 2026 does not need retroactive labeling. Tool makers already on the market have until 2 December 2026 for the hidden-marking part, but publisher disclosure duties had no such grace period.

United States

There is no single federal labeling law. Three layers matter instead.

Truth in advertising, nationwide. The rule against deceptive advertising applies to AI content exactly as it does to anything else. Two consequences follow, and they are where the real money is. A testimonial from a person who does not exist is prohibited outright, carrying civil penalties above $50,000 per violation under the consumer reviews rule in force since October 2024. The figure is adjusted for inflation annually, so check the current number. Endorsement rules apply to synthetic endorsers with no carve-out.

New York, since 9 June 2026. Any advertisement reaching New York audiences that features a synthetic performer requires conspicuous disclosure. Read the trigger carefully: it covers performers not recognizable as any identifiable real person. "It is not a real person" is the reason the rule applies, not a defense. Audio-only ads are excluded, as is using AI purely to translate a real performer. There is no private right of action, so enforcement runs through state authorities.

California, operative 2 August 2026. Large AI providers, meaning over one million monthly users, must offer a free public detection tool, offer a visible disclosure option, and embed hidden provenance data. This lands on tool makers rather than on you. It matters anyway, because it means the tools you use will increasingly mark their output by default. From 1 January 2027, large platforms must not knowingly strip provenance data from uploads.

There is a federal push to override state AI laws, but nothing has actually displaced them yet. The practical advice from firms tracking this is to keep complying with state rules until something concrete changes.

United Arab Emirates

The UAE has no AI-specific labeling law. What governs AI video is the general media framework, which is unusually strict and easy to trip over.

Since 1 February 2026, anyone creating promotional content from inside the UAE needs an advertiser permit from the UAE Media Council. This covers individuals, businesses, influencers and visitors, and it applies whether the promotion is paid or unpaid. Advertising must be clearly identifiable as advertising. Content standards cover misinformation, respect for religious and national values, defamation and privacy, and AI content is judged against those standards like anything else. Penalties reach AED 1 million, with content removal and license suspension alongside.

What this means practically. The UAE will not fine you for failing to write "AI-generated." It will act if AI content misleads, breaches content standards, or promotes without a permit. Given how broadly misinformation is treated, realistic AI footage that could pass as a real recording is the risky category.

Pakistan

Pakistan has no dedicated AI labeling law. What exists is thinner than the headlines suggest.

The National AI Policy sets direction but is a roadmap rather than an enforceable rulebook. The Prevention of Electronic Crimes Act remains the main law for online content; it predates generative AI and names neither deepfakes nor synthetic media. Its 2025 amendment added an offense around disseminating false information, carrying up to three years imprisonment or a fine up to PKR 2 million, and deepfake cases have been registered under its defamation and false-information provisions.

Punjab has drafted a Performers' Digital Identity and Artificial Intelligence Protection Act. As of writing it is a draft, not law. If passed it would require written and specific consent before creating an AI replica of a performer's voice or face, require AI-produced content to carry a clear disclaimer, protect the digital identity of deceased artists for 25 years, establish a provincial Digital Rights Registry, and carry prison terms alongside heavy fines.

What this means practically. Two things. The absence of a Pakistani labeling rule does not help you if your content reaches Europe, because the EU duty follows the audience. And platform rules apply in Pakistan in full, enforced automatically at upload.

The platform rules bite faster than any regulator

Regulators investigate. Platforms enforce at upload, automatically, everywhere, today.

Three details worth knowing. YouTube's separate rule on mass-produced repetitive content is a monetization rule that operates independently, so a video can be correctly labeled and still fail it. TikTok reads Content Credentials in uploaded files and labels automatically when it finds them, which is how it reached more than three billion labeled videos by July 2026, up from 1.3 billion the previous November. And Meta has shifted toward labeling manipulated media rather than removing it, though removal still applies when content breaks another rule.

Which videos actually attract penalties

Most AI video carries no meaningful risk. A narrow set does, and the pattern is identical across every jurisdiction and platform: realistic content that could be mistaken for a record of something that actually happened.

Highest risk. A real, identifiable person appearing to say or do something they did not. Altered footage of a real event presented as unaltered. A synthetic person giving a testimonial or review. AI-generated before-and-after results for weight loss, skin, hair or home improvement. Product demonstrations showing performance the product does not deliver. Anyone's voice or face used without permission. Political and election content. Health, medical and financial claims. And deliberately removing provenance marks before publishing, which is now a liability in itself rather than a clever workaround.

Lower risk. Clearly stylized or animated visuals. AI used for scripts, captions, planning and research. Color correction, upscaling, background cleanup and noise removal. Product imagery that is plainly a render and does not claim to be a photograph.

Is an invented AI person a "deepfake"?

This question comes up constantly and the answer is less comfortable than people expect.

Under the European approach, two things must be true together: the content resembles a person, place, object or event that exists or could plausibly exist, and it would falsely appear authentic to someone seeing it. A photorealistic invented person can plausibly exist. So a realistic AI human presented as though it were real footage generally does fall inside, even though no such person exists. The test is whether the viewer would take it as real, not whether a specific individual was copied.

Under New York's rule it clearly counts, because that law was written specifically for invented performers not recognizable as any real individual.

What does not count: obviously animated or stylized characters, illustrations, and anything the audience plainly is not meant to read as real. The practical rule is simple. If the AI person is photorealistic and appears in commercial content, label it. Labeling is free. The alternative is not.

Provenance: Content Credentials and invisible watermarks

A machine-readable mark is a signal inside the file saying "AI made this," readable by software but not visible to a viewer. It is not the logo in the corner, not the filename, not the caption. There are two kinds, and the difference explains everything else.

Content Credentials, the C2PA standard. A cryptographically signed record attached to the file listing what created it, when, and what happened to it afterward. Think of a nutrition label that travels with the product. Rich in detail, and fragile: re-encoding or re-saving can strip it. If someone alters the record the signature breaks, so you can delete it but you cannot quietly forge it.

Invisible watermarks, such as SynthID. A pattern woven into the pixels or audio itself. Carries almost no information, essentially just "this came from our system," but durable: it survives screenshots, cropping, compression and format changes because it lives in the content rather than attached to it.

They cover each other's weaknesses. Metadata says a lot and is easily lost. The watermark says little and is hard to remove.

What survives a real editing pipeline

Take a common chain: an image from ChatGPT, animated in Google Flow, edited in CapCut, posted to social.

Both tools embed credentials and watermarks at generation. Then CapCut joins clips, trims, crops, adds transitions, a logo and color grading. Every one of those operations requires decoding and re-encoding, which rebuilds the file from scratch. Credentials live in the file container, and a rebuilt container generally does not carry them forward. Assume the Content Credentials are gone after a consumer editor. Canva behaves the same way. Adobe's apps preserve and extend the chain; most consumer editors do not.

The invisible watermark is a different matter. It sits in the pixels, so trimming, cropping, transitions, overlays and color adjustments do not remove it. Your video still carries a detectable AI signal even with every trace of metadata gone. Then platforms transcode aggressively on upload, which is a final filter on anything that survived.

The net result: metadata almost certainly lost, invisible watermark almost certainly intact, and your disclosure obligation entirely unaffected by either. That last point is the one to internalize. Provenance is a technical chain that will often break. Disclosure is a decision you control completely. Do not confuse the two, and do not use tools that strip marks: it does not remove the pixel watermark anyway, and it converts an oversight into evident intent.

What this means for agencies and companies

This is where the operational cost actually lands, and the two sides of the table face different problems.

If you are an agency

Disclosure becomes a deliverable, not a favor. Build the label into the edit at production time as on-screen text in the first seconds, not into the upload step where a deadline will eat it. A label added at edit time ships with every export and every repost. One added at upload exists only in the person who remembered.

Your client's market decides your obligations, not your address. An agency in Karachi or Dubai producing for a client with European customers is inside the EU rules. This is the single most common misreading we encounter, and it is the one that costs the most to get wrong.

Put it in the contract. Three clauses are worth adding now: who is responsible for disclosure on each deliverable, what consent documentation the client must supply before you generate anyone's likeness or voice, and who carries the risk if a product claim in an AI-generated video turns out to be unsubstantiated. Most agency contracts written before 2026 answer none of these.

Keep the untouched original. Store the file exactly as the tool produced it, credentials intact, before editing. Once CapCut has rebuilt the container that provenance is gone, and the original is the only evidence you have of where the asset came from.

Log what you make. Tool, model, date, prompt, reviewer. This is dull and it is the record you produce if anyone asks. It is also, separately, the documentation that supports a copyright claim over the human parts of the work.

If you are the company hiring one

Ask how disclosure is handled before you see the creative. An agency that has not thought about it will improvise at upload time, which is the failure mode.

Do not accept a synthetic testimonial, ever. Not because of labeling, but because a person who does not exist claiming to have used your product is prohibited outright in the US with per-violation penalties, and it breaks platform rules everywhere. An AI presenter describing your product is fine with a label. An AI person claiming personal experience is not fine at all. This distinction is the single highest-value thing on this page.

Substantiation did not change. Nobody has been fined for using AI. They have been fined for what the AI said. Every claim in an AI-generated video needs the same evidence it would need in a filmed one.

Consent is specific. A general appearance release does not cover cloning someone's voice or face. If a real person appears, you need written permission naming AI generation explicitly, with scope, duration and territory.

Budget for the boring layer. Testing your pipeline, keeping originals, and maintaining a consent file is unglamorous work that nobody bills enthusiastically. It is also the entire difference between a manageable question and an expensive one.

A working procedure

What to watch next

2 December 2026: the European grace period ends for tools that were already on the market before August. 1 January 2027: California's rule barring large platforms from knowingly stripping provenance data. Pakistan: whether the Punjab performers bill becomes law, and whether other provinces follow. Google's watermark toggle: whether other providers make visible marks optional too, and how platforms respond. TikTok's spam detection: the definitions and consequences have not been published yet.

The short version

The regulation is less demanding than the headlines suggest. Nothing here stops you making AI video or using it commercially, and disclosure costs you nothing in reach or revenue on any major platform. The whole framework reduces to one habit: when it looks real, say it is not, visibly, early, in plain words.

The technical layer is worth understanding but should not be your compliance strategy. Credentials break in ordinary editing. Watermarks survive but you do not control them. Your visible label is the only part you fully control, and it is the part every regulator and platform actually asks for.

The two places where real money is at stake are not labeling at all. They are fake testimonials, meaning synthetic people claiming experiences nobody had, and claims you cannot substantiate. Both were prohibited long before AI existed. AI only made them cheaper to produce at scale.