The EU has drawn a clearer line for AI labels

The deadline now comes with instructions

Europe's AI transparency rules start to apply on 2 August. The European Commission has now published the final guidance that explains what those rules mean in practice.

The basic idea was already in the AI Act. People should know when they are dealing directly with an AI system, and they should be able to recognise certain synthetic or manipulated content. The new document matters because it draws some of the awkward lines: when a chatbot warning is needed, what counts as a deepfake and how much human editing is enough to change the labelling duty for text.

There are two different jobs here. Providers build or put an AI system on the market. Deployers use it in a professional setting. A model company may need to make generated content detectable, while a publisher using that model may have to place a clear label where a reader can see it.

The distinction is easy to lose. It is also where many compliance decisions will begin.

What providers have to make visible

A system that directly talks with a person — a chatbot, agent or avatar, for example — should identify itself from the start of the first interaction. There is an exception when the AI nature of the exchange is obvious to an average, reasonably observant person. The Commission says that exception should be read narrowly.

Providers of generative systems also have to add machine-readable marks to synthetic audio, images, video and text so the material can be detected as generated or manipulated. That is the technical layer. It is meant for software and platforms, not only for the person looking at the content.

Not every output is covered. The guidance excludes source code, short strings of numbers or symbols, machine-to-machine material never exposed to people and some non-final work inside closed production environments. Standard editing can also fall outside the marking duty when it does not substantially change the input or its meaning.

That still leaves judgement calls. Removing noise from an image is not the same as inventing an event that never happened. Between those examples sits a great deal of ordinary creative work.

A deepfake label has to reach a person

Deployers have a separate disclosure duty. If an image, recording or video closely resembles a real or plausibly real person, place or event and could falsely appear authentic, the guidance treats it as a deepfake.

The disclosure must be clear and perceivable by the time someone first encounters the content. A visible label or an audible notice can do that. Relying only on hidden, machine-readable metadata cannot.

Context matters. An audience watching an obvious fictional production may not expect every background or effect to be real. Artistic, satirical and fictional works have a more flexible disclosure rule so a notice does not ruin the work. That is not a general escape for deceptive material presented as evidence.

People also have to be informed when they are exposed to emotion-recognition or biometric-categorisation systems, whether the analysis happens live or later.

Human review means more than a tidy-up

The most useful clarification may concern text about public affairs, health, science, finance and other matters people reasonably debate.

AI-generated or manipulated text published to inform the public on those subjects needs a clear label when it has not undergone human review or editorial control. A real review can remove that labelling duty. A spell-check, grammar pass or procedural sign-off cannot.

The Commission describes human review as a deliberate examination of the substance by someone with relevant knowledge and professional judgement. Editorial control means an entity has real authority to approve, change or reject the content on substantive grounds, including fact-checking and source quality. A person must also carry ultimate legal responsibility for publication.

That makes the rule less about whether a machine touched a sentence and more about whether a human editorial process genuinely stands behind it. It is a meaningful distinction, though one that authorities may eventually have to test with evidence rather than a policy statement on a website.

What is confirmed, and what is still open

The confirmed part is the timetable and the scope described in Article 50. Most duties apply from 2 August 2026. A limited grace period runs to 2 December for the machine-readable marking obligation on generative systems already placed on the market before the August date. Older content does not have to be labelled retroactively.

The Commission says the rules will reduce deception and help people calibrate their trust. That is a policy expectation, not a measured result. A label can be ignored, a technical mark can be stripped and a badly placed notice can satisfy a process without helping anyone.

The guidance explains the Commission's interpretation; it does not replace the legislation. National market-surveillance authorities will handle most enforcement. The AI Office and the European Data Protection Supervisor have narrower roles. Fines under the Act can reach €15 million or 3% of worldwide annual turnover, with proportionality for smaller companies.

The hard questions now move into implementation. What will count as obvious AI interaction? When does editing become generation? How will a company prove that a review was substantive? And will labels remain attached when content moves across apps and platforms?

The final guidance gives teams a much better map. It does not remove every grey area. That was probably never realistic.