AI-Generated Video Disclosure Rules: YouTube, TikTok and Meta

YouTube, TikTok and Meta all ask you to label realistic AI video, each in its own way. YouTube wants a setting turned on in Studio. TikTok requires a label and bans some uses even when they're labeled. Meta asks you to disclose photorealistic video and realistic-sounding audio, and applies its own "AI info" label using C2PA and other industry signals. Underneath all three sits provenance metadata, C2PA Content Credentials, that can trigger a label whether or not you set one. Invisible watermarks such as Google's SynthID add a second signal that detectors can check.

This guide covers what YouTube, TikTok and Meta require as of October 2026, what the provenance signals do and where they fail, and a checklist to run before each upload. These rules move, so each policy summary links to its source. Check the current policy before relying on any detail here. This is not legal advice.

The rules at a glance

All three platforms draw the line at realism. The question is whether a viewer could mistake the content for real footage of real people, places or events. A stylized animation is treated differently from a photoreal clip of a person saying something they never said.

All three platforms read C2PA metadata, so a file carrying Content Credentials may get labeled for you.

YouTube: the realistic-content test

YouTube's disclosure policy requires creators to flag realistic content that was altered or made synthetically. It names three situations:

• A real person appears to say or do something they didn't.
• Footage of real events or places has been altered.
• A realistic scene shows something that never happened.

You declare it with the "AI use" setting in YouTube Studio. The label then appears below the player on long-form videos and as an overlay on Shorts. Animation, beauty filters and AI used for scripts or captions are exempt.

YouTube may also add a label itself. It always does for content made with Veo or Dream Screen, and for files carrying C2PA metadata that marks them as fully generative. Repeated failure to disclose can lead to removal or suspension from the YouTube Partner Program. According to YouTube's announcement, disclosing does not reduce reach or monetization.

The b-roll case catches people out. If you cut generated b-roll into a talking-head video, the person on camera is real but the inserted shots are not. A photoreal generated shot of a scene that never happened fits YouTube's third case. When in doubt, turn the setting on.

TikTok: label it, and know what's banned outright

TikTok's Community Guidelines require a label on AI-generated or significantly edited content that shows realistic scenes or people. Unlabeled content can be removed, restricted or labeled by TikTok. Once you post with a label, you can't remove it, so make the call before you publish.

TikTok also labels some content automatically: anything made with its own AI effects, and content carrying C2PA Content Credentials from other platforms, as described in the TikTok newsroom. If your generator writes Content Credentials and your editor preserves them, TikTok may label the clip even if you forget. If either link is missing, TikTok has no Content Credentials to read, and the label is on you.

Some uses are banned even with a label:

• The likeness of anyone under 18.
• Adult private figures used without their consent.
• Fake endorsements by public figures.

Ads carry a harder consequence: an ad with undisclosed AI gets rejected. If you're producing AI UGC-style ads, read our AI UGC video guide and TurboSurge's breakdown of whether AI UGC is allowed on TikTok and Meta before you put money behind a campaign.

Instagram, Facebook and Threads: Meta's "AI info" label

Meta's policy requires users to disclose organic posts that contain photorealistic video or realistic-sounding audio that was digitally created or altered. Meta says it "may apply penalties" when people don't. As of the Oversight Board's March 2026 decision, Meta still hadn't published what those penalties are.

Note the audio clause. As we read it, it covers real footage with a cloned or generated voice-over, not only fully generated video. Seedance 2.5, Veo 3.1, Kling 3.0 Turbo and MiniMax H3 Max all generate audio with the picture, so check the soundtrack too.

The label viewers see is "AI info", which Meta renamed from "Made with AI" in 2024. Meta applies it using C2PA and other industry signals, as explained in its labeling announcement. The disclosure rule summarized here is written for organic posts. Our sources don't cover Meta's paid-ad rules, so check those separately.

Content Credentials (C2PA): the label inside the file

C2PA is an open standard for provenance metadata, known to most people as Content Credentials. A supporting tool attaches a record to the file describing how it was made. Platforms read that record and can label the content automatically.

• Who writes it: OpenAI for its images, Google for Nano Banana Pro and Gemini Omni, and TikTok for its own content.
• Who reads it: TikTok, YouTube and Meta.
• The weak point: the metadata chain breaks when a file passes through tools that don't support it, as Google acknowledges.

That matters because a typical workflow doesn't post the raw generation. You trim it, add captions and export again. If any tool drops the metadata, the platform has no Content Credentials to read, and a C2PA-triggered label won't appear. That doesn't make the clip exempt. The disclosure rules above apply to what the content shows, not to whether the file still carries metadata.

The same logic applies to start frames. If you make a still with GPT Image or Nano Banana Pro and animate it with a different video model, the credentials on the still describe the still. The video is a new file from a different tool, so don't assume anything carries through. Our image-to-video workflow guide covers that pipeline.

SynthID and other invisible watermarks

An invisible watermark is a signal embedded in the image or video itself rather than in a metadata field. Google's version is SynthID, embedded in the pixels of every Veo frame and every Nano Banana Pro image. You can check for it in two places:

• The Gemini app checks uploaded videos up to 100 MB or 90 seconds.
• The SynthID Detector portal is open to everyone (in English) and also detects partner content, including OpenAI's.

Detection has a clear limit: a detector can only find the watermarks it knows. A clean result means the detector didn't find its own watermark. It doesn't prove the clip came from a camera. We also haven't verified how well these watermarks survive heavy editing, so treat them as a backstop, never as your disclosure.

Here is what each generator attaches, according to our sources:

"Not confirmed" means we couldn't verify it, not that the feature is missing. You can use all eight of these models side by side in ClipSpeed's AI Creator. If documented watermarks matter for a job, pick a Google model or GPT Image 2.5 Flare or Sunburst, but don't assume any mark survives trimming and re-export. Either way, set the disclosure yourself, and for the last row assume no platform will label it for you.

The law is catching up: the EU AI Act and the FTC

Two legal rules also matter for creators and brands that publish AI video. This is a summary, not legal advice.

EU AI Act, Article 50

Since August 2, 2026, anyone deploying deepfakes must disclose them, and providers of generative AI must machine-mark their outputs. Systems already on the market have until December 2, 2026 to add marking. Law-firm summaries, including Baker McKenzie's, cite maximum fines of €15 million or 3%. Note that the deepfake disclosure duty falls on whoever deploys the deepfake, not only on the model vendor. If any of your business or audience is in the EU, check whether it applies to you.

The FTC's rule on fake reviews and testimonials

In the US, the FTC's Consumer Reviews and Testimonials Rule, in effect since October 21, 2024, bans fake reviews and testimonials, including AI-generated ones attributed to people who don't exist. Knowing violations can draw civil penalties of about $53,000 each. Keep it in mind for AI-avatar "UGC" testimonial ads. An AI label on the post doesn't turn an invented customer into a real one. Our guide to using AI video commercially covers rights and licensing in more depth.

A disclosure checklist for every upload

1. List every AI element in the clip. Visuals, voice, music, b-roll, a generated hook shot. One generated shot makes a real video partly synthetic.
2. Apply the realism test. Could a viewer mistake any element for real footage of a real person, place or event? If the only AI involved is animation, a beauty filter, a script or captions, YouTube doesn't require a label. Read TikTok's and Meta's wording for your case.
3. Check the audio on its own. Meta's wording covers realistic-sounding audio that was digitally created or altered, so a cloned or generated voice that sounds real can need disclosure even when the picture is real.
4. Screen for banned uses. On TikTok, no likeness of anyone under 18, no adult private figures without consent and no fake endorsements by public figures, labeled or not. Get written consent from anyone whose likeness you use.
5. Set the label yourself on every platform. Use the "AI use" setting in YouTube Studio, TikTok's AI label before you post (it can't be removed later) and Meta's disclosure for photorealistic content. Don't count on auto-labels.
6. Don't treat metadata as disclosure. Your editor may strip C2PA data, and we couldn't confirm that every generator writes it. Missing metadata means no auto-label, not no obligation.
7. Handle ads and testimonials separately. Undisclosed AI in a TikTok ad gets rejected, Meta's paid-ad rules need a separate check, and AI testimonials attributed to people who don't exist fall under the FTC rule.
8. Keep a record. Save the prompt, model name and generation date for each clip. If a label is questioned later, you can show what was generated and how.
9. Re-read the policies before big campaigns. They change, so bookmark the pages linked above.

The bottom line

The rule is the same in spirit on YouTube, TikTok and Meta: if AI made a realistic person, place, event or voice that a viewer could take as real, label it. All three read C2PA metadata to label content automatically, but that metadata can be lost when a clip passes through tools that don't support it, so the label is your job. TikTok adds a list of uses no label can fix, and the EU AI Act and the FTC rule add legal stakes on top. Run the checklist before every upload, and check the current policy when in doubt.

If most of what you post is real footage, the question is simpler: clips from your own streams, podcasts and videos show things that happened, and YouTube explicitly exempts AI captions. For how the two workflows differ, see AI video generation vs AI clipping, or try ClipSpeedAI's clipping.

Frequently asked questions

Do I have to label AI-generated video on YouTube?

Only if it's realistic. YouTube requires creators to flag realistic altered or synthetic content: a real person appearing to say or do something they didn't, altered footage of real events or places, or realistic scenes that never happened. You use the "AI use" setting in YouTube Studio. Animation, beauty filters and AI used for scripts or captions are exempt. YouTube may also add a label itself, and always does for content made with Veo or Dream Screen or carrying C2PA "fully generative" metadata (YouTube Help).

Does TikTok label AI videos automatically?

Sometimes. TikTok labels content made with its own AI effects and content carrying C2PA Content Credentials from other platforms. Everything else is on you: the Community Guidelines require a label on AI-generated or significantly edited content showing realistic scenes or people, and unlabeled content can be removed, restricted or labeled by TikTok. A label can't be removed after posting (TikTok Community Guidelines).

What is Meta's "AI info" label?

It's the label Meta shows on Facebook, Instagram and Threads for AI-generated or altered content. It was called "Made with AI" until Meta renamed it in 2024, and Meta applies it using C2PA and other industry signals. Separately, users must disclose organic posts with photorealistic video or realistic-sounding audio that was digitally created or altered, and Meta says it "may apply penalties" to those who don't (Meta).

Will labeling my video as AI hurt its reach?

YouTube says disclosing does not reduce reach or monetization (YouTube blog). We didn't find an equivalent statement from TikTok or Meta in the sources we checked, so we can't say either way for those platforms. Skipping a required label, on the other hand, carries risk on all three: removal, restriction or penalties.

What's the difference between C2PA and SynthID?

C2PA is an open standard for provenance metadata (Content Credentials) attached to a file. TikTok, YouTube and Meta read it, but the chain breaks when a file passes through tools that don't support it. SynthID is Google's invisible watermark, embedded in the pixels of every Veo frame and every Nano Banana Pro image. You can check for it in the Gemini app (videos up to 100 MB or 90 seconds) or the SynthID Detector portal. Neither replaces your own disclosure.

Do AI captions on my own footage need an AI label?

On YouTube, no: AI used for scripts or captions is explicitly exempt. TikTok's and Meta's rules target AI-generated or altered realistic scenes, people, video and audio, and captions on real footage don't match that wording as we read it. Check the current policy if you're unsure. If you add generated shots or a cloned voice to the clip, the rules apply to those parts.

Does the EU AI Act apply to AI video creators?

Possibly. Since August 2, 2026, Article 50 requires anyone deploying deepfakes to disclose them, and generative-AI providers must machine-mark their outputs, with a December 2, 2026 deadline for systems already on the market. Whether and how it applies to you depends on your situation, so check the current guidance. This is not legal advice (Baker McKenzie summary).

Related guides

• Can You Use AI-Generated Video Commercially?
• AI UGC Videos: Formats That Work, What Still Looks Fake and How to Test Them
• AI Video Generation vs AI Clipping: When to Generate and When to Clip
• Image-to-Video AI: A Still-to-Motion Workflow That Keeps Product Details Intact
• AI-Generated B-Roll for Talking-Head and Podcast Clips: When to Cut Away and How to Keep It Honest
• How to Avoid Copyright Strikes When Clipping YouTube Videos
• 12 AI Video Mistakes That Make It Look Fake, and How to Fix Each