Mandatory AI Disclosure Rules for Publishing Image-to-Video Generated Content
The specific legal articles, jurisdictions, and platform labelling mechanisms that require disclosure of image-to-video generated content, and the watermarking methods that do and do not satisfy them.
Mandatory AI Disclosure Rules for Publishing Image-to-Video Generated Content
The EU AI Act Article 50(4) transparency obligation takes effect on 25 June 2026. It applies to any AI-generated video that "resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful." That date is your deadline for having a disclosure system in place if you publish image-to-video generated content in the European Union. The legal and platform obligations for disclosing that a video was generated by an image-to-video model do not align neatly across jurisdictions. The gap between a model's claimed watermark compliance and a platform's actual identification system is the single thing most likely to leave you exposed. This guide names the exact articles, jurisdictions, and application dates for AI disclosure rules for image-to-video generated content. It explains how embedded provenance marks, visible watermarks, and the C2PA standard interact with platform identification systems that may ignore them.
EU AI Act Article 52 Generated Video Transparency
Who Must Disclose
Article 50(4) of Regulation (EU) 2024/1689 imposes the transparency obligation on deployers of AI-generated video. The obligation applies to any deployer making AI-generated video available to the public within the EU, regardless of where the deployer is based. "Deployer" means the natural or legal person using the AI system in a professional capacity. A solo filmmaker publishing from outside the EU falls under it if their content is accessible in the EU. The obligation is triggered when the generated content "resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful." An image-to-video clip that turns a still photograph of a real person into a moving video that looks realistic meets that bar.
Disclosure Rules And Two Exemptions
The deployer must disclose that the content is artificially generated or manipulated. The disclosure must be clear and conspicuous. Two exceptions exist. Content that has undergone "human review or editorial control" where a natural or legal person assumes responsibility for its accuracy is exempt. A filmmaker who reviews and takes editorial responsibility for each frame may not need to label it as AI-generated, though the legal test for sufficient editorial control is not yet litigated. Content that is "evidently an artistic, creative, satirical, fictional or analogous work" is also exempt, provided transparency does not hamper enjoyment of the work. A surreal image-to-video piece that is clearly not realistic may qualify.
Model Provider Obligations
The EU AI Act's Article 53, dealing with general-purpose AI model providers, takes effect one month earlier on 2 August 2025. It requires providers to publicly disclose a summary of training data used. That obligation falls on model providers, not deployers. It affects the compliance chain because a deployer relying on a model that has not disclosed its training data may face secondary exposure.
Synthetic Content Labelling Requirements 2026
YouTube
YouTube's Altered and Synthetic Content policy, effective 18 March 2024, requires a label when content contains "realistic" altered or synthetic material that could be mistaken for real. The label is a Creator Studio toggle. Mark the video as containing altered or synthetic content, and YouTube applies a visible label to the video player. Exemptions cover clearly unrealistic content such as animation or special effects. YouTube does not rely on your model's embedded provenance mark to apply the label. If you do not self-declare and YouTube's own identification system flags your content, you risk demonetisation or removal.
Meta
Meta's Manipulated Media Policy, updated April 2024, applies to Facebook, Instagram, and Threads. Meta applies a "Made with AI" label based on industry-standard signals or self-disclosure. The platform looks for C2PA provenance data and other signals. In practice the identification system has a known mismatch with certain watermarking methods. Self-declare.
TikTok
TikTok's Synthetic and Managed Media policy, effective 9 May 2024, requires creators to use a self-disclosure toggle for realistic AI-generated video. TikTok also uses automatic identification via Content Credentials (C2PA) to apply labels. The platform has published its detection criteria, but the system does not read every provenance format. Uploaders who rely solely on their model's invisible embedded mark can find their content unflagged while comparable content from another model is flagged. This creates an unpredictable compliance environment.
| Platform | Detection Method | Watermark Types Read | What Happens When Unread |
|---|---|---|---|
| YouTube | Self-declaration plus own detection system | C2PA provenance data (partial); visible watermarks via frame analysis | Content may be labelled or demonetised without warning; self-declaration is the only reliable route |
| Meta (Facebook/Instagram) | Industry-standard signals including C2PA; own detection system | C2PA provenance data; SynthID (partial); visible watermarks | False negatives are common; content may be labelled incorrectly or not at all; self-declaration recommended |
| TikTok | Creator toggle plus automatic C2PA detection | C2PA provenance data only; no known reading of SynthID or other invisible watermarks | Content may go unlabelled if C2PA is absent; platform may apply label if own detection system flags artefacts regardless of provenance data |
Image-to-Video Watermark Platform Compliance Mismatch
Google Veo 2 And SynthID
A model provider may claim its watermark satisfies platform disclosure rules. Platforms use their own identification systems that may not read that specific provenance format. Google Veo 2 embeds SynthID, an imperceptible digital watermark in video frames. SynthID is Google's technology. It is not a standard that platforms other than Google have adopted. YouTube is owned by Google, so Veo 2's SynthID watermark may be detected by YouTube's systems. Meta and TikTok do not publicly confirm reading SynthID. A filmmaker who generates a clip using Veo 2, relies on SynthID to satisfy their disclosure obligation, and uploads to TikTok may find the content unlabelled and themselves in violation of TikTok's policy.
Runway Gen-3 Alpha And C2PA
Runway Gen-3 Alpha includes C2PA provenance data in its outputs, as stated in its model card published 17 June 2024. C2PA is a broader provenance standard that TikTok and Meta both claim to read. The specification version used must match. TikTok's system reads C2PA version 2.1, published 30 January 2024. If Runway updates its C2PA implementation to a newer version before TikTok updates its identification system, the provenance data may not be read.
OpenAI Sora And Kling 1.6
OpenAI Sora includes C2PA provenance data and adds a visible watermark by default. The visible watermark is the safer option for platform compliance. Any platform's frame analysis can detect it regardless of provenance format. The trade-off is aesthetic: a visible watermark degrades the clip's usability for commercial or artistic purposes. Kling 1.6, accessed through Kuaishou's Chinese-platform interface, has no public model card detailing watermark or provenance data as of 26 September 2026. Uploaders using Kling must verify its disclosure features directly from the platform. Chinese AI disclosure rules under the Provisions on the Administration of Deep Synthesis Internet Information Services, effective 10 January 2023, require deep synthesis service providers to mark generated content. That obligation applies to the provider, not to the uploader. The marks used under Chinese rules may not be recognised by EU or US platforms.
C2PA Metadata YouTube Detection Mismatch
The C2PA (Coalition for Content Provenance and Authenticity) specification version 2.1, published 30 January 2024, is the most widely adopted provenance standard for AI-generated content. Runway, Adobe Firefly Video, and OpenAI Sora all include C2PA provenance data in their outputs. YouTube, Meta, and TikTok all state they use C2PA as one detection signal. The mismatch occurs because each platform implements C2PA detection independently and may not read every field in the provenance data. A platform may validate only the cryptographic signature without verifying that the provenance data accurately describes the video's origin. If a model provider updates its C2PA implementation, adding a new field for the specific image-to-video model version used, a platform that checks only the existing fields may ignore the provenance data entirely.
The practical consequence: C2PA provenance data is not a guarantee of platform compliance. Check the provenance data against each target platform's current detection criteria. Those criteria are not publicly documented in full. The failure case is a clip that carries valid C2PA provenance data from a model that claims compliance, is uploaded to a platform that reads C2PA, and is still flagged as non-compliant. The platform's identification system checked a different field or detected generation artefacts independently of the provenance data. Confirm compliance by uploading a test clip, monitoring the platform's response, and adjusting your disclosure method. Switch to a model with a visible watermark or use the platform's self-declaration toggle regardless of the model's embedded provenance data.
Who AI Disclosure Rules Suit And Who Should Skip
The EU AI Act transparency obligations, synthetic content labelling requirements, and platform-specific identification systems suit the solo filmmaker or video artist who needs to distribute image-to-video generated content commercially in the EU or on major platforms. These rules also suit the legal or compliance person clearing AI-generated video for broadcast, advertising, or platform distribution. They need the exact articles, application dates, and platform detection methods to build a compliance framework. The educator or researcher teaching generative tools may need this information to explain the legal landscape to students.
Someone who generates video only for personal use, never publishing it, keeping clips on a local drive or sharing them privately without platform upload, does not need to follow these disclosure rules. The open-source tinkerer running models locally for experimentation with no public distribution is not a deployer under the EU AI Act and has no platform disclosure obligation.
Before publishing a clip that depicts a real person, obtain explicit consent from that person. No disclosure label substitutes for consent. The greatest risk comes from assuming that a model's claimed watermark compliance is sufficient for all platforms. Test each platform's response to your chosen model's output before relying on its disclosure method.
Common Questions
Do I need to label every image-to-video clip I publish?
No. The EU AI Act exempts content that is evidently artistic, satirical, or fictional, and content that has undergone human editorial control where you assume responsibility. Platforms exempt clearly unrealistic content like animation. Label only clips that could be mistaken for real.
Can I rely on my model's provenance watermark instead of self-declaring?
No. Platforms use their own identification systems that may not read your model's provenance format. Google Veo 2's SynthID is not guaranteed to be read by TikTok or Meta. Self-declare via the platform's toggle for reliable compliance.
What happens if I don't label and the platform detects it?
YouTube can demonetise or remove the video. Meta applies a label anyway, but the label may be wrong. TikTok can remove the content or restrict its reach. Platform detection is inconsistent, so non-disclosure is a risk even with provenance data present.
Do US state deepfake laws apply to non-political image-to-video clips?
No. California, Texas, Washington, Minnesota, and Michigan have enacted deepfake disclosure laws limited to election-related content within specified time windows before elections. A non-political clip published outside those windows is not covered.