EU AI labels are no longer just a proposal for the future. Transparency obligations under the European Union’s AI Act began applying on 2 August 2026, setting rules for certain conversations with artificial intelligence and for material that has been artificially generated or manipulated. For ordinary people, the important question is not whether every photograph will carry a giant warning. It is when a person must be told that the thing they are seeing or hearing is synthetic.
The law addresses several different situations, and treating them as one universal watermark requirement causes confusion. The European Commission’s official guidance distinguishes duties for providers of AI systems from duties for people and organisations deploying those systems. Some outputs need machine-readable marking, some require disclosure to the audience, and some situations have exceptions. This matters for the creators, publishers and audiences trying to understand what the rule actually changes.
Which EU AI labels apply to chatbots?
The most familiar situation is a conversation with a customer-service bot or another system that interacts directly with a person. Under Article 50, providers must generally ensure that people are informed when they are interacting with an AI system, unless that fact is obvious to someone reasonably well informed, observant and circumspect in the circumstances.
A customer opening a clearly named chatbot may already understand what they are using. Someone receiving messages that appear to be from a human representative may not. Whether disclosure is necessary depends on that context, which makes interface design and presentation important parts of compliance rather than decorative details.
There are additional specific provisions for systems that categorise people using biometric data, and for emotion-recognition systems. The law is not a blanket permission to operate such tools merely because a notice is displayed. Other rules and restrictions may still apply, including prohibitions and data-protection obligations.
LiveAIWire previously looked at how chatbot disclosure can affect persuasion. The connection is straightforward: knowing that the other party is an AI can change how a person evaluates the message. Transparency is useful even when the words on the screen remain identical.
What must be marked when AI generates pictures or sound?
The Act also imposes duties on providers of systems that generate synthetic audio, images, video or text. Those providers must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, as far as technically feasible and subject to the requirements and exceptions in the legislation.
That is a technical obligation on the system side. It is not the same as demanding that every generated picture carry a visible red banner across its centre. A machine-readable marker may be embedded in ways that software can detect. How effectively those markings survive cropping, re-encoding, screenshots or other editing is a separate practical issue.
The law provides an exception where the system performs an assistive function for standard editing or does not substantially alter the input data or their semantics. Consequently, an automatic enhancement of an existing image should not casually be lumped together with the creation of a wholly fabricated scene. The precise circumstances matter.
The text of Article 50 draws distinctions that are easy to lose in social-media discussion. A watermarking obligation for providers, a transparency statement for the audience and the responsibilities of someone publishing content are related, but they are not interchangeable legal tests.
Why realistic deepfakes are treated differently
People and organisations that deploy AI systems to generate or manipulate image, audio or video content constituting a deepfake have a disclosure obligation. A viewer hearing what appears to be a real politician, musician or family member may otherwise believe that the recording is authentic. The rule aims to make the synthetic nature of that content clear to the people exposed to it.
The Act defines a deepfake in relation to content that appreciably resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. That definition is more specific than using ‘AI-generated’ as a synonym for ‘deepfake’. A stylised illustration of a fictional creature will not necessarily raise the same concern as a convincing fake recording of a real event.
There are qualifications for artistic, creative, satirical and fictional works. The disclosure requirement in those settings is limited so that it does not hamper the display or enjoyment of the work. That is not a general exemption from transparency; it is a direction to provide disclosure in a way suited to the context.
The harm can be personal as well as political. LiveAIWire has covered regulatory action involving intimate deepfakes in the UK. The European transparency rules are a different legal measure, but both illustrate why a fabricated recording can have consequences long after the original file has circulated.
What happens when AI writes something people rely upon?
The rules also address AI-generated or manipulated text published for the purpose of informing the public on matters of public interest. Under the Act, deployers generally must disclose that the text was artificially generated or manipulated. However, an important exception can apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication.
This is a significant distinction for publishers. It recognises a difference between unexamined automated text being presented as public information and material subjected to human editorial responsibility. It does not mean that attaching a person’s name to a completely unchecked output automatically satisfies every other duty a publisher may have.
The law does not make a reader’s broader questions disappear. Is a claim accurate? Where did a quotation originate? Has a photograph been cropped misleadingly even without AI? Good disclosure does not replace those checks. A labelled fake can still be deceptive in context, and an unlabelled genuine image can still tell an incomplete story.
For creators, there is also a practical records issue. It may become useful to retain information about where synthetic material originated and which edits were made, particularly when multiple tools are used in one workflow. That is an operational suggestion, not a separate record-keeping rule imposed by every part of Article 50.
Does this mean everything online is now trustworthy?
No. The regulation establishes transparency duties for specified actors and outputs within its scope. It does not guarantee universal compliance or provide a magical detector for material produced elsewhere. Content can be reshared across borders, altered after initial creation or separated from the notice that accompanied its first publication.
It also matters when the rules began applying. The Article 50 obligations generally started on 2 August 2026. The legal text includes transitional treatment for certain AI systems that were placed on the market before that date, with a later compliance deadline for that specific situation. A claim that every existing system had to be retrofitted overnight would be too broad.
Readers should ask two practical questions about a suspicious clip: who published it, and what reliable evidence supports its authenticity? A disclosure label, when present, is valuable information. Its absence is not proof that a dramatic recording is real. This is especially true when an image has travelled through several platforms.
The EU’s wider approach also intersects with questions of how models are trained and who owns the materials they use. LiveAIWire’s coverage of the EU copyright and AI-training debate concerns a different issue. Whether a model may use protected work during training is not the same question as whether its finished output must be disclosed to an audience.
The most useful change may be a clearer expectation
For a consumer, these rules are likely to matter most at points of uncertainty. A call that sounds human, an apparently convincing video, or a public-interest article with an unclear origin can lead someone to trust an account they would evaluate differently if its production method were known.
For organisations, the challenge is to identify which role they occupy in the chain. An AI model provider, an app developer, a marketing agency and the publisher of a final advertisement may face different duties. Labelling one stage of the process is not necessarily enough to satisfy the obligations attached to another.
The aim is not to persuade readers that synthetic material is automatically bad. AI can create useful illustrations, translations and accessible interfaces. The more durable principle is that a person should not be misled about whether a relevant interaction or apparently authentic recording was generated by a machine. As the technology becomes harder to spot by sight or sound, that distinction matters more, not less.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.