An AI label is a useful warning, not a certificate that something is fake. From 2 August 2026, the transparency obligations in Article 50 of the AI Act apply in the European Union. For the public, the most visible change is straightforward: a deepfake—an image, audio recording or video created or altered with AI so that it appears authentic—must be clearly disclosed. At the same time, providers of generative systems must add markings that can be detected by software to synthetic content.
What matters in a Facebook, TikTok, YouTube or WhatsApp feed, or on a local website, is less straightforward: the absence of a label does not mean the material is real. The file may be old, may come from a service outside the scope of the rules, or may have been cropped, re-uploaded or compressed until technical data are no longer available. And a visible label says that the creator or platform has declared the use of AI; on its own, it does not establish whether all claims in the video are true.
What changed in the EU on 2 August 2026
The European Commission’s guidelines on Article 50 show that the obligations are divided between those providing the AI system and those using it professionally to publish or distribute content.
- Chatbots, AI agents and interactive avatars must tell users, from the first interaction, that they are interacting with an AI system, except where this is obvious from the context.
- Providers of generative AI must make AI-generated or manipulated images, audio, video and text detectable through a marking readable by automated systems. The requirement covers synthetic content, not only deepfakes.
- Those using AI professionally to publish a deepfake must clearly inform the person exposed to it that the image, voice or video has been generated or manipulated.
- AI-generated or manipulated text on matters of public interest must be declared as such if published without genuine human review or editorial control and editorial responsibility.
The Regulation applies directly in Romania, as in the other Member States. However, it does not require every retouched photograph, every text corrected by an assistant, or every personal use of an AI generator to carry a visible label.
There is also an important timing nuance. For generative systems placed on the market before 2 August 2026, the technical marking and detection obligation under Article 50(2) has a limited transition period until 2 December 2026. Content created before 2 August 2026 does not have to be labelled retroactively. Therefore, in autumn 2026, markings can be expected to be applied unevenly, especially to older content and existing services.
Deepfake does not mean every AI-made image
In everyday language, “deepfake” has come to mean almost any material generated by AI. The AI Act’s definition is narrower. A deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful to a person.
For example, a recording in which a voice imitates a mayor and announces that water supplies will be shut off in a town may be a deepfake. So may a video in which a candidate appears to make a statement they did not make. By contrast, an AI-generated background in an obviously fantastical video does not automatically become a deepfake in the legal sense relevant to this obligation.
The exception for clearly artistic, satirical, creative or fictional works does not remove transparency altogether: it allows disclosure to be made in a way that does not spoil the experience of the work. An obvious parody has a different context from a clip presented as breaking news in a local group.
Four different things you may confuse with an “AI label”
| Signal | Who can see it | What it may indicate | What it cannot prove |
|---|---|---|---|
| Visible label or audio warning | Any user | That the material has been declared AI-generated or AI-manipulated | That all claims in the material are false or that its origin has been independently verified |
| Machine-readable marking | Platforms and compatible tools | That a system can automatically identify content as synthetic or manipulated | That a user will necessarily see it or that the marking survives every instance of distribution |
| Provenance metadata | Verification applications and services | Information about creation, editing and, sometimes, the file’s history | That a file is authentic if the data are missing; metadata can be removed during export or re-uploading |
| Automated AI detector | A specialised service | An estimate of how likely certain features are to be synthetic | A certain verdict, especially for compressed files, translations, new material or unknown models |
The Commission explicitly states that a person professionally distributing a deepfake cannot rely only on the hidden marking in the file. The disclosure must be perceptible to people—for example, through a visible label or an audio announcement—no later than the first exposure. As part of the voluntary implementation framework, the EU has also made optional icons available. They can help, but using an icon does not by itself demonstrate legal compliance.
Why the absence of a label does not confirm authenticity
The sound rule for users is: “No label” means “I have not been given information”, not “I have been given confirmation”.
- A video may be published by a person acting strictly in a personal capacity. The AI Act generally excludes non-professional activities; that does not automatically make the content authentic.
- The material may have been created before the rules began to apply, or with an older tool still in the technical transition period.
- A file may go through a screenshot, re-encoding, cropping, export or re-upload. During such operations, some provenance data may be lost.
- It may be a forgery without AI: conventional editing, invented subtitles, audio overlaid on real images, a false headline or genuine footage taken out of context.
- Some routine AI edits may fall under the exception for standard editing, where they do not substantially alter the material or its meaning.
Nor do detectors solve the problem with a magic score. The NIST report on synthetic-content transparency treats provenance, watermarking and detection as complementary methods, with different vulnerabilities and trade-offs. In other words, a detector’s verdict should be used as a lead for investigation, not as sole proof.
A short verification protocol before sharing a video
When a video claims to show an urgent situation—such as a fire, accident, announcement by authorities, political statement, fraud or medical alert—check the central claim first, not only the pixels, shadows and lip movements.
- Pause sharing for a few minutes. Pressure such as “send this to everyone now” is a risk signal, not evidence.
- Read the label, if there is one. Look for wording such as “AI-generated”, “AI-modified”, “generated voice” or a platform warning. Note whether the label says “fully generated” or “partially modified”: they are not the same thing.
- Identify the primary source. Who published it first? Is it an official account, an institution, an identifiable journalist, or a page that merely reposts it? A video attributed to a town hall should be checked through that town hall’s official channels, not only through reposts.
- Look for confirmation of the event. For traffic, weather, service outages and other local situations, compare the claim with the responsible institution and several independent sources. A post about a nationwide disruption does not become credible simply because it contains dramatic images.
- Check the date and location. Search key frames with reverse image search, and examine visible landmarks, licence plates, weather, street names and the language in the background. Much viral material is real but old, or filmed in another country.
- Listen to the claim, not just the voice. A voice clone can sound convincing. Ask whether the person actually published the statement, whether a full recording exists and whether the institution confirms the message.
- Treat detection tools as a second opinion. If a detector flags AI, look for verifiable explanations. If it does not, do not consider the case closed.
This type of digital discipline complements established protections. For example, if a video or message pushes you to urgently open a link, make a payment or reveal a code, also follow the steps in InfoBuz’s guide to recognising a phishing message. The deepfake may merely be the convincing packaging for a scam.
What publications, local pages and institutions need to watch
For organisations in Romania, the useful rule is not “add a discreet note at the end”. In the case of deepfakes, the information must be clear and distinct when the public sees or hears the material, not hidden on a terms page or in a comment that is lost when the material is reshared.
In the same context, see also USB-C charger for phone and laptop: how to choose the right power, cable and ports, which provides additional relevant information.
For texts on matters of public interest, the law draws an important distinction. An article for which an editor has checked the substance, corrected factual errors, and for which an organisation accepts editorial responsibility may fall under the labelling exception. A superficial check of spelling or style is not editorial control within the meaning of the Commission’s guidelines. For readers, this means that the absence of a label on a news text may reflect the presence of a responsible editorial team, but the outlet’s reputation, the author and the sources should still be checked.
Especially during election campaigns, civic debates, public-health alerts or emotionally charged local events, a label is a transparency signal that must be read, not a shortcut to the truth. The AI Act raises the standard for those who produce and publish the material. For the public, the defence remains the same: source, context, date and independent confirmation.
Conclusion: the label is the start of verification, not the end
The rules applying from 2 August 2026 make it harder for a professional deepfake to be presented without any warning. This is an important step, and the requirement to combine technical marking with visible disclosure recognises the limits of each method when used alone. But an unmarked video may be authentic, conventionally forged, generated before the new requirements, or simply stripped of its context. Before sharing it, first check who claims it, what event it purports to show, and whether a responsible source confirms it.
Frequently asked questions
Must all AI-generated images have a visible label in the EU?
No. Providers of generative systems must make outputs detectable through markings readable by automated systems. A visible label is required especially for deepfakes distributed by professional users and for certain texts on matters of public interest that lack human review or editorial responsibility.
If a video has no AI label, is it authentic?
No. The absence of a label is not evidence of authenticity. The material may be old, re-uploaded, created by a non-professional user, made using another form of manipulation, or stripped of technical data after editing and compression.
What does machine-readable marking mean?
It is technical information attached to or embedded in content so that compatible applications and tools can identify the material as AI-generated or AI-manipulated. As a rule, it is not enough on its own for the public to notice it.
Does a news article written with AI assistance need a label?
Not automatically. For texts on matters of public interest, the AI Act provides an exception where there is genuine human review or editorial control and where a person or organisation accepts editorial responsibility. Grammar checking alone is not enough.
Can I trust an online deepfake detector?
Use it only as an indication. Detectors can make mistakes, especially with compressed files, altered material, content from unknown models or short recordings. Confirming the source and context remains essential.
