A video can show a flood, a protest or a politician saying something inflammatory. A label saying “AI-generated” changes how a viewer should read it. A signed record showing which tool made the video and what edits followed may add more context. Neither, by itself, tells the viewer whether the story surrounding the clip is true.
That distinction matters now that the European Union’s AI Act has started requiring transparency for certain AI interactions and content. The new rules make labels more consequential. They also make it easier to confuse a disclosure with proof.
What the European rules cover
Article 50 of the AI Act applies from 2 August 2026. It puts duties on both the companies providing certain AI systems and the people or organisations using them. Providers must mark certain AI-generated or manipulated outputs in a machine-readable way. Deployers must clearly disclose deepfakes, and certain AI-generated or altered text published on matters of public interest when it has not undergone human review or editorial control. People must also be told when they are interacting directly with an AI system in the cases covered by the Act. The European Commission’s guidelines explain the roles and obligations.
That is narrower than “every picture made with AI needs a warning.” The rules depend on who provides or deploys the system, the kind of content, how it is used and, for some text, whether a person with relevant expertise or a responsible editor has meaningfully reviewed its substance. A quick grammar check alone does not count as that review. Some creative or fictional works can use a disclosure suited to the work, rather than a label that disrupts it. The Commission’s Article 50 questions and answers give examples and explain the exceptions.
There is also a limited transition period: for AI systems already on the market before 2 August, the provider’s machine-readable marking obligation applies from 2 December 2026. That transition concerns marking and detection; it does not suspend every disclosure duty.
A visible label and a technical credential do different jobs
A visible label is for people. A note beside a video, a caption, or an audible disclosure can tell someone what they are seeing without special software. The Commission says a deployer cannot rely on invisible, machine-readable marks alone to fulfil a disclosure duty to viewers.
A technical credential is for software that can inspect a file. Content Credentials, based on the C2PA specification, can carry a signed record of an asset’s origin and changes. A credential might say that an image began as a camera photograph, was edited in particular software, and later had part of its background generated. It can help a newsroom or platform inspect that history. C2PA’s 2026 guidance describes ways to record AI involvement, edits and the area of an image that changed. Read the implementation guide.
The two signals can complement each other. The label explains something to an ordinary viewer; the credential can offer a more detailed, verifiable account to a tool or investigator. A newsroom might display “AI-generated scene” beside a video while preserving its credential so an editor can inspect the source and edits.
But a verified credential answers a limited question: can the recorded history be tied to this file and verified under the credential’s trust system? It does not certify that the event shown happened, that a caption is accurate or that the person who made the assertion deserves trust. C2PA’s own guidance says valid credentials can accompany misinformation. Its harm-modelling document explains what credentials do and do not establish.
Read the claim, then check the source
Suppose a viral clip carries a valid credential showing that it was generated in an AI video tool last week. That is useful information about the clip’s creation. It does not tell you who wrote the accompanying post, whether the post quotes someone accurately or whether the scene is satire. Those are separate questions, and answering them requires evidence beyond the file’s metadata.
The reverse also needs care. A file without a visible label or Content Credential is not automatically genuine, and its absence does not prove that the file was made by AI. A credential may never have been attached, or a distribution or editing workflow may have lost it. C2PA explicitly cautions that the absence of credentials is not a reason to distrust the content on its own.
For a reader, a label should start an investigation rather than end one. Ask what was generated or changed, who made that disclosure, and whether the record has been verified. Then look for the original source, check when and where it was published, and seek independent confirmation of the event.
For a publisher, the practical lesson is to make disclosures visible at the point where people encounter the material and preserve technical provenance through the editing and distribution chain. Describe the intervention precisely: “AI-generated background” is more useful than a vague “AI content” badge when only the background changed. Avoid implying that verification of a file’s history proves the story it is used to tell.
A good label reduces uncertainty about how media was made. A useful credential can let others inspect the record behind it. Trust still depends on the source, the evidence and what a person can independently verify.
