Do your AI agents meet EU AI Act Article 50?

7 checks for each agent

1. List every agent that talks to people: chat, email, voice, comments, support replies.
2. For each of those agents, write down where the person is told they are dealing with an AI system, and when: at the first interaction, not later.
3. List every agent that generates audio, images, video or text that leaves your organisation. Tools that only assist standard editing, or do not substantially change the input, are exempt from the marking rule.
4. For each of those agents, check that the output is marked in a machine-readable way as AI-generated, not only with a visible label.
5. Find any image, audio or video your agents make that resembles real people, objects, places or events and would falsely appear authentic. That is a deepfake under Article 50 and needs a disclosure. For evidently artistic, satirical or fictional work, the disclosure can be limited so it does not spoil the work.
6. Find AI-written text you publish to inform the public on matters of public interest. It needs a disclosure unless a person reviewed it and a person or organisation holds editorial responsibility.
7. If you deploy emotion recognition or biometric categorisation, check that the people exposed to it are told how it works.

Which dates apply?

The European Commission's AI Act Service Desk gives 2 August 2026 as the date the Article 50 rules apply. It gives 2 December 2026, amended by the Digital Omnibus, as the deadline for some providers whose content-generating systems were already on the market to meet the marking rule in paragraph 2.

Article 99 sets the fines. Breaching Article 50 falls under a lower maximum than prohibited practices, so figures quoted for prohibited practices do not apply to Article 50. Read the amounts in Article 99 itself.

How do you show a disclosure happened?

Article 50 asks you to disclose and to mark. It does not tell you how to keep proof that you did. If you are ever asked, it helps to show, for a given conversation or output, that the disclosure or the marking was in place when it went out.

• Keep the disclosure text and the marking method under version control, with the date each version went live.
• Record, for each agent, which version it used, and when.
• Keep that record somewhere the agent cannot edit.

Where TrustGate fits

TrustGate is not an Article 50 tool and does not mark content. TrustGate returns a signed allow, deny or escalate decision before the action runs, for actions that go through it. The receipt binds the agent, action, target and exact arguments that were evaluated. For actions that go through it, that record can show which agent sent what, to which target, and whether it was allowed to.

To test a workflow of your own, use the free review kit.

Related: what an agent execution boundary is, evidence of human oversight under the EU AI Act and how to make an AI agent audit trail tamper-evident, which covers the record-keeping that Article 12 asks of high-risk systems.

What this page is not

• Not legal advice. Read the article and ask your counsel.
• Not a full list of your obligations. Other parts of the Act, such as the rules for high-risk systems, may also apply.
• Not a way to make a workflow compliant. It is a list of questions to answer.