Assertions, not verdicts: what travels with a generative video after it leaves your site

A generative video campaign running in seven languages across four country domains produces a lot of assets that nobody watched frame by frame before publication. Two separate obligations land on those files: AI-generated content has to be identifiable, and the provenance of the asset is its own cryptographic record. Teams tend to collapse both into a single checkbox at the end of the pipeline. They are different artefacts, with different failure modes, and only one of them follows the file.

A disclosure flag and a provenance manifest are not the same control

The AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026, with some exceptions. Two of its requirements bear directly on generative media. When using AI systems such as chatbots, humans should be made aware that they are interacting with a machine so they can take an informed decision. Separately, providers of generative AI have to ensure that AI-generated content is identifiable. The first is an interface duty. The second attaches to the asset.

Identifiability inside your own pages is the easy half: a line of copy, a label beside the player, a field in the CMS. The hard half starts when the file leaves. A video reposted elsewhere, re-encoded and cropped, carries none of your page furniture with it. Whatever marks it as generated has to live in or alongside the file. A third date matters for planning: starting on 2 December 2027, high-risk AI systems will be subject to strict obligations before they can be put on the market.

Assertions describe, they do not adjudicate

At the heart of the C2PA specification is the Content Credential, a cryptographically bound structure that records an asset's provenance, also known as a C2PA Manifest. It contains one or more assertions, which are statements about the asset, such as its origin, meaning when and where it was created, its modifications, meaning what happened using what tools, and use of AI, meaning how it was authored. Those three categories map closely onto what a generative pipeline already logs internally.

What the manifest does not do is rule on content. Content Credentials do not provide value judgments about whether a given set of provenance data is true. The specification is explicit that it is not a cure-all for misinformation, but instead seeks to mitigate against its threats in the digital domain. So the record answers who stated what happened to this file, not whether the claims inside it hold. Set reviewer expectations on that boundary before rollout, not after the first dispute.

Per-language variants multiply the surface

Four markets, one codebase, Tatano Energy, 2025–2026: a multilingual e-commerce platform deployed across four country domains, a daily SEO autoblog driven by search trends, and multi-language generative video campaigns. Country domains indexed separately: 4. Languages served: 7. Manual intervention required: 0. Each language variant of a campaign is its own asset, with its own origin and modification history, and a disclosure that has to read correctly in the language a viewer actually sees.

That is why a per-file record beats a per-campaign approval. Signing off on a campaign says nothing about which variant was re-rendered after a late script change, or which one was cropped for a different placement. Write the provenance record in the same step that writes the file, so every variant carries assertions describing its own history rather than inheriting a summary of its siblings. A campaign-level note degrades the moment one asset is regenerated.

Validate the output, not only the label

Misinformation occurs when LLMs produce false or misleading information that appears credible, and overreliance occurs when users place excessive trust in LLM-generated content, failing to verify its accuracy. A manifest that honestly states a video was authored with AI does nothing about a wrong product claim inside it. Labelling and correctness are separate controls, and a visible provenance badge can raise confidence in content that nobody checked.

OWASP's mitigations point in two directions here. Use retrieval-augmented generation to enhance the reliability of model outputs by retrieving relevant and verified information from trusted external databases during response generation. Then implement tools and processes to automatically validate key outputs, especially output from high-stakes environments. For a product video, the key outputs are the claims: what the equipment does, what it is called, what is promised. Grade those per language, before the asset is signed.

Automated validation needs graders rather than reviewers. Design evals that mirror your real-world task distribution, which for a four-market campaign means drawing questions from every language and every asset type you actually publish. Structure questions to allow automated grading: string match on required disclosure wording, code-graded checks on manifest fields, LLM-graded checks on claim fidelity. More questions with slightly lower-signal automated grading beat fewer hand-graded ones. Most use cases need multidimensional evaluation along several success criteria.

The practical split is simple to hold. Disclosure is a product decision about the interface, provenance is a file-level record of origin, modification and use of AI, and claim accuracy is a graded test that neither of the first two performs. Build all three as separate steps in the pipeline that writes the asset, name which one each check covers, and accept that a signed manifest proves authorship history, not truth.

Sources

European Commission (Shaping Europe’s digital future) — AI Act — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

C2PA — C2PA and Content Credentials Explainer (2.4) — https://spec.c2pa.org/specifications/specifications/2.4/explainer/Explainer.html

OWASP Gen AI Security Project — LLM09:2025 Misinformation — https://genai.owasp.org/llmrisk/llm092025-misinformation/

Claude Platform Docs (Anthropic) — Define success criteria and build evaluations — https://platform.claude.com/docs/en/test-and-evaluate/develop-tests

Neurolinks case study — Four markets, one codebase — https://neurolinks.be/work/tatano-energy