The EU Puts a Label on Synthetic Reality
Europe has officially entered the “show your work” phase of artificial intelligence.
From August 2, 2026, new transparency obligations under the European Union’s AI Act require companies and organisations to disclose when people are interacting with certain AI systems. The rules also cover content generated or manipulated by AI, especially realistic deepfakes and synthetic material dealing with matters of public interest.
That means a photorealistic video of a politician saying something outrageous should not simply appear in somebody’s feed and pretend it arrived from an ordinary camera. It needs a disclosure. Likewise, an AI chatbot should not impersonate a human customer-service representative while quietly plotting its escape from the support window.
The obligations come from Article 50 of the AI Act. They apply differently to companies that build AI systems and those that use them. Providers must design qualifying systems with transparency features, while deployers must make appropriate disclosures when publishing certain outputs.
As The Verge reports, the European Commission has also introduced standard disclosure icons that companies can use. Those particular icons are optional. Disclosure itself is not.
It sounds like a small visual change—a symbol here, a notice there—but it represents a much larger transition. AI transparency in Europe has moved beyond an admirable product feature. It is now a legal requirement.
Who Actually Has to Do What?
The rules divide responsibility between two broad groups: providers and deployers.
A provider develops an AI system and places it on the market. Think of a company offering a chatbot, image generator or video model. A deployer uses an AI system professionally or incorporates it into a service, publication or platform.
One business can occupy both roles.
Providers of interactive AI systems must ensure that people know when they are dealing directly with AI, unless the artificial nature of the interaction is already obvious. A chatbot labelled “AI assistant” probably does not need to announce its robotic identity after every sentence. A synthetic receptionist deliberately presented as a human is another matter.
Providers of systems that generate or manipulate text, audio, images or video must also support machine-readable marking. According to the European Commission’s Article 50 guidelines, those technical markers should enable synthetic material to be detected as artificially generated or altered.
Deployers face a more visible responsibility. If they publish a realistic deepfake, they must disclose its artificial nature. They must also label certain AI-generated or manipulated text published to inform the public about matters of public interest.
In short, the company building the synthetic-content machine must provide the technical plumbing. The organisation pressing “publish” must often provide the warning sign.
Deepfakes Are the Main Character
The word “deepfake” often conjures images of political deception, celebrity impersonations or disturbingly realistic scams. The AI Act addresses that territory directly.
The regulation broadly focuses on AI-generated or manipulated image, audio or video material resembling existing people, objects, places, entities or events in a way that could falsely appear authentic. If a synthetic clip convincingly depicts an event that never happened, a label should tell viewers that reality had some algorithmic assistance.
Ireland’s RTÉ highlighted the new requirements as the rules took effect, drawing attention to both deepfakes and chatbots. The timing matters. Generative tools can now create convincing voices and videos in minutes, while social platforms can distribute them to millions before a fact-checker has finished making coffee.
The law does not declare every AI-assisted image dangerous. Nor does it ban deepfake technology. It attempts to give audiences essential context.
That distinction matters because synthetic media can serve legitimate purposes. Filmmakers use it for effects. Teachers reconstruct historical scenes. Artists remix reality. Accessibility tools can generate speech for people who have lost their voices.
The EU’s approach is therefore not “computer-generated equals forbidden.” It is closer to “computer-generated should not masquerade as documented fact.”
The label does not settle whether content is good, bad or misleading. It tells viewers what kind of object they are looking at.
Not Every AI Touch Requires a Giant Warning
The transparency system includes limits and exceptions. Europe is not demanding a flashing hazard symbol every time someone removes red-eye from a photograph.
The strongest disclosures focus on realistic deepfakes and qualifying text about public-interest matters. The rules also recognise that artistic, fictional, satirical and creative works require different treatment. A disclosure may be presented in a way that does not spoil the audience’s experience, provided it still communicates that AI-generated or manipulated content is present.
That flexibility is important. A historical drama that digitally recreates a deceased public figure is not equivalent to a fabricated news clip designed to influence an election. Both may involve similar technology, but context, presentation and likely audience interpretation differ enormously.
The requirements for public-interest text also contain nuance. AI-generated or manipulated text must generally be disclosed when it is published to inform the public about public-interest matters. However, the rules account for human review and editorial responsibility.
As WeRSM explains, the Commission expects disclosures to appear when users first encounter qualifying content. They should use accessible language, remain understandable and avoid burying the truth inside a magnificent swamp of legal prose.
The objective is clarity, not decorative compliance. A label technically present but practically invisible would defeat the point.
Europe Has Designed Its Own AI Icons
To make disclosures more consistent, the European Commission has produced a family of standard icons for AI-generated and AI-modified content.
The designs can indicate different levels of involvement, such as content generated entirely by AI, partially modified with AI or created with AI assistance. Black, white and transparent versions allow organisations to place the symbols over different visual backgrounds.
Companies do not have to use these exact designs. They may create their own disclosure systems if those alternatives meet the law’s requirements. However, the Commission’s icons offer a ready-made vocabulary and could reduce the visual chaos caused by every platform inventing a different symbol.
That consistency could prove valuable. Users already encounter labels on services such as Facebook, Instagram and TikTok, but platform-specific language and presentation vary. One site may say “AI info.” Another might display “digitally created.” A third may hide the explanation behind a menu that requires the investigative instincts of Sherlock Holmes.
The EU icons attempt to establish a shared visual shorthand.
Still, an icon only works when people understand it. Public education will matter. Platforms must explain what their labels mean, whether a marker denotes fully generated material or a minor alteration, and what conclusions users should—and should not—draw from it.
A label offers context. It is not an automatic verdict on truth.
Invisible Markers Matter Too
Visible labels are only half of the system. Providers must also make qualifying synthetic outputs detectable in a machine-readable format.
Possible techniques include metadata, digital watermarks, content credentials, cryptographic provenance signals and other technical methods. These mechanisms can help platforms and verification tools identify an output even when no human-readable caption accompanies it.
Google already uses SynthID to embed signals into some AI-generated material. Other companies support provenance standards such as C2PA, which can record information about a file’s origin and editing history.
WebProNews reports that Google has backed European efforts promoting clearer identification of generated text, images, audio and video. Its involvement matters because Google operates across the entire synthetic-content pipeline—from Gemini and Imagen to Veo, YouTube and Search.
Yet no technical marker is magical. Metadata can disappear during screenshots, downloads or file conversion. Watermarks may weaken after heavy cropping, compression or editing. Determined actors will try to remove them.
That is why the EU framework encourages a layered approach. Visible disclosures speak to people. Machine-readable markers speak to software. Provenance records provide additional evidence about how content was created and changed.
None is perfect alone. Together, they make deception more difficult and verification more practical.
The Voluntary Code Has Legal Consequences
Alongside the binding AI Act, the European Commission has developed a voluntary Code of Practice on Transparency of AI-Generated Content.
The words “voluntary code” can sound like a polite suggestion attached to a fruit basket. Here, however, the distinction is more substantial.
Signing the code is voluntary. Complying with Article 50 is mandatory.
The code supplies practical methods through which providers and deployers can demonstrate compliance. It includes one section for providers covering marking and detection, and another for deployers covering deepfake and text labelling.
Companies that follow the approved code gain a clearer compliance route. They can point to recognised practices instead of designing an entirely separate framework and then convincing multiple authorities that it works.
Organisations that do not sign may still comply. However, they must demonstrate that their alternative measures offer an equivalently adequate level of transparency. Regulators can assess those systems individually.
This structure lets the EU establish legal outcomes without prescribing one permanent technical solution. That makes sense in a field where last year’s cutting-edge watermark can quickly become this year’s easily defeated parlour trick.
It also gives industry participants a role in improving implementation. Signatories can collaborate through task forces, compare techniques and respond to emerging ways of stripping or confusing markers.
The law sets the destination. The code offers a reasonably well-lit route.
Old Systems Receive a Limited Breathing Space
The obligations began applying on August 2, but the transition is not identical for every product.
New systems placed on the market from that date must meet the applicable requirements immediately. Generative AI systems already available before August 2 receive a limited grace period, until December 2, 2026, for the machine-readable marking and detection obligation.
That allowance recognises an engineering reality: adding durable provenance mechanisms to an existing model and its entire output pipeline is not necessarily a weekend patch.
However, this is not a universal four-month holiday from transparency. The Commission’s official FAQ explains that the grace period specifically concerns the marking obligation for previously marketed generative systems. Other applicable Article 50 duties should not automatically be treated as postponed.
Material generated before August 2 does not require retroactive labelling. The Commission nevertheless encourages organisations to label older content where practical.
That line avoids an absurd compliance treasure hunt through years of online archives. Requiring every platform to identify and relabel all historical synthetic content would be technically daunting and sometimes impossible.
From now on, though, the expectation is clearer. New synthetic material should carry information about its origin as it moves into the public sphere.
The age of “we’ll figure out provenance later” is rapidly closing.
The Price of Ignoring the Label
Companies that breach these transparency requirements can face fines of up to €15 million or 3 percent of their total worldwide annual turnover, whichever is higher.
For major technology companies, 3 percent of global revenue is not loose change discovered beneath the office sofa. It creates a serious incentive to build compliance into products rather than attaching a tiny disclaimer after regulators call.
National market-surveillance authorities will handle much of the enforcement, while the European AI Office oversees areas assigned to the Commission. In Ireland, the regulatory framework includes the country’s newly established AI Office, adding domestic machinery to the broader European system.
As RFI reports, regulators are gaining broader powers to inspect advanced systems and act against organisations that violate the rules. Depending on the investigation and legal basis involved, authorities may seek information, examine documentation and impose sanctions.
Enforcement will determine whether the labels become meaningful or merely cosmetic. Authorities must decide how visible a disclosure should be, how resilient technical marks must remain and when a company has taken reasonable steps.
They must also enforce the rules consistently across 27 member states. If identical content receives different treatment in Ireland, France and Germany, Europe risks replacing one transparency problem with a regulatory patchwork.
The Commission’s guidelines and shared icons aim to prevent that fragmentation.
A Label Helps, but It Cannot Manufacture Trust
Europe’s rules address a genuine problem: people increasingly struggle to tell what was recorded, what was generated and what was subtly manipulated.
Labels can help. They can expose synthetic impersonations, give audiences useful context and provide platforms with common detection signals. They may also discourage some casual deception because unlabeled AI content now creates legal risk.
But labels cannot solve everything.
A disclosure does not stop a convincing falsehood from going viral. Bad actors may ignore the rule, remove technical markers or distribute material from outside the EU. Users may misunderstand labels and assume that anything marked “AI” is false—even when the underlying information is accurate. The opposite danger also exists: authentic material can be dismissed as synthetic, a tactic often called the liar’s dividend.
Platforms also need reliable systems for correcting mistakes. False AI labels can damage artists, photographers and journalists whose genuine work gets flagged by imperfect detectors.
The EU has therefore taken an important step, not delivered a final cure. Transparency must work alongside media literacy, platform accountability, authentication tools, independent journalism and effective enforcement.
Still, the direction is unmistakable. Europe wants synthetic media to carry provenance rather than arrive wearing a fake moustache and insisting it has never met an algorithm.
From August 2 onward, AI can still create, edit, entertain and persuade. It simply has to introduce itself.
Sources
- The Verge – Europe’s AI labeling and transparency rules are now in effect
- RFI – EU steps up AI regulation as new transparency rules come into effect
- RTÉ – New EU rules on AI deepfakes and chatbots
- WeRSM – Europe’s AI Labels Are Now Mandatory
- WebProNews – Google Joins EU Generative AI Code of Practice to Label AI Content
- European Commission – Code of Practice on Transparency of AI-Generated Content
- European Commission – Guidelines on Transparency of AI-Generated Content
- European Commission – Article 50 Transparency Obligations FAQ
Kingy Launch Brief
Put the week’s verified AI launches in your inbox.
Get a source-checked briefing on consequential AI launches, with a clear try, watch or skip verdict. Beehiiv will ask you to confirm your address, then you can choose the subjects you want to follow.
Free · Choose your subjects · Double opt-in · Unsubscribe anytime