REGULATORY INSIGHT

The EU AI Act transparency obligations — what became enforceable on 2 August 2026

22 August 2026

On 2 August 2026 the European Commission's AI Office, together with national authorities, began enforcing the AI Act. The transparency obligations that took effect on the same date reach every operator using AI to generate or alter content — including operators who build and run their systems entirely outside the EU.

Three obligations, now enforceable

The Commission's announcement sets out three in concrete terms.

• Disclosure for interactive systems — chatbots and other interactive AI systems must tell users they are dealing with AI, not a human.
• Deepfake labelling — images, video, or audio that have been edited or generated using AI must be labelled.
• Machine-readable marking — AI-generated or altered content must carry machine-readable marks so that it can be detected more easily.

The first two are visible to a person. The third is different in kind. Because the requirement is that the mark be machine-readable, a label rendered on screen does not satisfy it. The information has to be carried in the content itself, in a form a machine can read.

The reach extends well beyond the EU

Article 2(1)(a) applies the Regulation to providers placing AI systems on the market or putting them into service in the Union, "irrespective of whether those providers are established or located within the Union or in a third country".

Article 2(1)(c) goes further, reaching providers and deployers in a third country "where the output produced by the AI system is used in the Union". The system need never be located in the EU. If its output is used there, the Regulation applies.

For an operator outside the EU, the question is therefore not where the system runs, but whether any path exists by which its output reaches users in the Union.

Labelling is the easy part

The difficult part in practice is not meeting the obligation. It is being able to show, afterwards, that you met it.

The content you distributed on a given day did carry a machine-readable mark. Six months later, can you demonstrate that to a third party?

Implementing a marking function is comparatively straightforward. But compliance is assessed on whether the implementation demonstrably worked, not on whether it was built. Marking can fail during an incident. A configuration change can quietly take one distribution path out of scope. When that happens, the question becomes whether your records let you establish which content, over which period, was affected.

Under an enforcement regime, an enquiry from an authority or a complaint from a third party is what triggers the demand for an explanation. What you produce at that point is not a specification. It is the operational record.

Absorb it in process, or carry it in the system

There are two ways to meet a requirement of this kind. One is procedural: write the runbook, and have people perform the checks and keep the records. The other is structural: build the system so that the records needed for proof accumulate on their own.

The procedural route starts faster and breaks down as you scale. Distribution paths multiply, staff change, exceptions accumulate, and the records stop being consistent enough to reason over. The structural route front-loads the design work but keeps you able to establish the affected range afterwards. Most of the intellectual property we hold sits on the structural side.

Worth checking now

• Whether any path exists by which your system's output is used in the Union
• Whether AI-generated or altered content carries a machine-readable mark
• Whether a failure to apply that mark is recorded — and if not, whether the affected range could be established afterwards
• Whether disclosure of AI involvement covers every interactive path, not just the primary one

Note that the rules for high-risk AI systems have been deferred to 2 December 2027, and those for systems embedded in regulated products to 2 August 2028. The transparency obligations are being enforced ahead of them.

← Back to Insights