AI Labels Do Not Prove a Video Is False: What to Check

Technology7 min read

A video featuring a familiar voice promises a treatment, announces a public decision or makes a political statement. Beneath it appears a notice about AI. The first instinct may be: “so it’s false.” But the label alone cannot answer that question.

Such a signal may indicate artificial intervention, information supplied to the platform or a technical history of the file. It does not, however, establish whether the person actually said the words attributed to them, whether the information is current or whether the main claim is correct. Those are separate checks.

Start with the question: who applied the label?

The same apparently simple wording — “AI content” — can cover information of differing weight. Before drawing a conclusion, check whether the platform explains where the signal came from.

Type of signal	Where it may come from	What it is for	Key limitation
Visible notice for the public	The person using AI and publishing the material, in situations covered by the regulation	Warns that a deepfake was generated or artificially manipulated	It does not verify the accuracy of the message, its legality or the author’s intent
Platform label	The social network or distribution service	Signals that the platform has received or identified relevant information about the content	It does not automatically say how the signal was obtained, nor is it, in every case, equivalent to the obligation under the AI Act
Provenance data	The app, creator, publisher or technical infrastructure	May retain verifiable assertions about the creation and modifications of the file	It is not a certificate that the claims in the video correspond to reality

The distinction matters: a visible notice is designed for the person viewing the material, while machine-detectable marking serves technical infrastructures. Article 50 of the AI Act treats these functions separately. eur-lex.europa.eu

What may be lost before a video reaches you

The absence of a label does not confirm authenticity. The material may be real, it may be unlabeled synthetic content, or it may be a copy from which information originally available has been lost.

This is particularly relevant after reposting, format conversion, compression, editing or screenshots. Provenance systems, including C2PA standards, can attach verifiable assertions about a file’s origin and history. But this data must be available and preserved throughout distribution in order to remain accessible.

C2PA describes provenance as information about the content’s journey, not as validation of the reality depicted by the file. In other words, a technical history can help answer the question, “Where does this material come from?”, but it cannot, on its own, decide the question, “Is what it says true?” c2pa.org

Isolate the verifiable claim. Formulate the key sentence in your mind: who allegedly said what, when, where and about what.
Find the original publication. Compare the date, author, description and the existence of a longer or unedited version.
Look for the direct source. If an institution, doctor, candidate or public figure is cited, check the relevant primary channels and documents first.
Test the context. An authentic video can mislead through captions, editing, an incorrect location or a false date. Manipulating context does not necessarily require a deepfake.
Use the label as a starting point. It may justify caution and further research, but it is not the final verdict on whether the video is true.

What the AI Act provides, in brief

The regulation does not start from the premise that every use of AI must be disclosed to the public in the same way. Article 50 sets out distinct obligations for distinct situations.

Direct interaction with AI: users must be informed that they are interacting with an AI system, except where this is obvious from the context.
Synthetic content: providers of certain generative systems must design outputs so that their artificial origin can be automatically detected, insofar as solutions are technically feasible, interoperable and reliable.
Deepfakes made available to the public: the actor using the AI system to make the material available to the public must clearly and distinguishably disclose that the material was artificially generated or manipulated, under the conditions set out in the regulation.
Texts on matters of public interest: the rule does not apply mechanically to every text for which an AI tool was used; the regulation takes into account the purpose of publication and whether human editorial review or control exists.

For the purposes of the AI Act, a deepfake is AI-generated or manipulated content that resembles existing persons, objects, places, entities or events and falsely appears to be authentic or truthful. For this reason, not every artificially generated image should automatically be treated as a deepfake. eur-lex.europa.eu

For more information about the AI Act, see also An AI label does not prove that a video is false: how to check a deepfake under the new EU rules, which offers complementary details on this subject.

Who has the obligation?

The roles should not be conflated. The provider of the AI system has obligations related to technical detectability for certain synthetic outputs. The person using the system and making the material public is the central actor for visible disclosure of the deepfake in cases covered by the regulation. The platform may in turn have its own labelling and moderation policies, but its label does not automatically transfer all legal obligations to it.

European Commission guidelines distinguish between strictly personal, non-professional use and professional activities or activities carried out for regular economic benefit. The specific situation and the role of each actor remain important to the application of the rules. digital-strategy.ec.europa.eu

Timeline: two dates, two situations

The transparency obligations in Article 50 apply from 2 August 2026. This is the general reference date for rules on informing users when they interact with AI, the detectability of certain synthetic outputs and disclosure of deepfakes.

There is also a separate deadline for a specific category: Regulation (EU) 2026/1744 sets 2 December 2026 as the date for compliance with the technical marking obligation provided for in Article 50(2), for providers whose synthetic-content generation systems had already been placed on the market before 2 August 2026. This is not a general postponement of all transparency obligations. eur-lex.europa.eu

If you suspect a problem in Romania

For a potentially misleading video, the quickest step is usually to report it on the platform. Before doing so, save the link, the account that published it, the date and time, screenshots of the label displayed — or of its absence — and, if you can find it, the link to the original post.

ANCOM presents itself as the market surveillance authority and single point of contact for the AI Act in Romania, alongside sectoral authorities for certain fields. Jurisdiction may depend on the type of system and the sector. The page consulted does not indicate a single public-facing procedure for every possible labelling case; in a report, it is useful to specify the actor concerned, the material, the platform, when it was encountered and the specific issue observed. ancom.ro

The key point: a label may clarify something about how a video was made or circulated. Checking the truth begins after you read the label; it does not end there.

Frequently asked questions

Does an AI label prove that a video is false?

No. It may indicate artificial generation or manipulation, a signal received by the platform or the existence of provenance data. It does not automatically verify the claims, date, context or intent behind publication.

If I do not see an AI label, is the video authentic?

Not necessarily. It may be authentic, it may be unlabeled synthetic content, or it may be a copy in which provenance data was lost after editing, conversion or redistribution.

What is the difference between a platform label and digital provenance?

A platform label is a visible message applied under the rules of the relevant service. Digital provenance includes technical information about a file’s origin and history. Neither alone certifies the truth of the claims in the material.

When do the transparency obligations in Article 50 of the AI Act apply?

General application begins on 2 August 2026. For certain synthetic-content generation systems already placed on the market before that date, the deadline for compliance with the technical marking obligation in Article 50(2) is 2 December 2026.

Who must disclose a deepfake published online?

In situations covered by the AI Act, the visible disclosure obligation lies primarily with the actor who uses the AI system and makes the deepfake available to the public. The provider of the system may have different obligations, including those concerning technical detectability, and platforms may use their own labels.