Your agent has to say it's an agent.

Five days ago a rule came into force that most of the people building AI support agents have not read. It doesn't ask whether your model is accurate, where your data is hosted, or what you trained on. It asks something much smaller and much harder to fake: does the person on the other end of this conversation know they are talking to software? Article 50 of the EU AI Act became applicable on 2 August 2026, and for anyone running a chatbot in Europe it is the first line of that Act to actually bite.

What changed on 2 August 2026?

Article 50 of the EU AI Act became applicable. It imposes four transparency duties: tell people when they are interacting with an AI system, mark AI-generated content so it can be detected, tell people when emotion recognition or biometric categorisation is running, and label deepfakes and AI-generated public-interest text. It applies to any AI system used in those four situations, not only to high-risk ones.

That last clause is why this one matters more than the rest of the Act for most companies. High-risk classification catches a narrow set of uses — hiring, credit, biometrics. Article 50 catches anyone who put a chatbot on a website. The Future of Life Institute's compliance checker data put transparency second only to AI literacy as a compliance trigger, at roughly 33% of respondents.

And it does not spare us. The obligations extend to providers and deployers of open-source AI systems, which are explicitly not exempt. An MIT licence is a good answer to a lot of questions. It is not an answer to this one.

• 50(1)Providers. Systems that interact directly with people — chatbots, voice assistants, AI companions, agents that autonomously call or email — must be designed so the person knows they are dealing with AI. Where the agent acts on someone's behalf, the identity of that person or entity too.
• 50(2)Providers. Systems generating synthetic audio, image, video or text must mark outputs machine-readably and supply a means of detection. Marking without detection is not enough.
• 50(3)Deployers. Anyone exposed to emotion recognition or biometric categorisation must be told it is running — real-time or after the fact.
• 50(4)Deployers. Deepfakes get labelled. So does AI-generated text published to inform the public on matters of public interest, unless it went through human editorial review and a named person or company holds editorial responsibility.

Does my AI support chatbot have to tell users it's an AI?

Almost certainly yes. Article 50(1) requires providers of systems that interact directly with people to design them so users are informed they are dealing with AI. The only carve-out is where that is obvious to a reasonably well-informed, observant and circumspect person — and the Commission's Guidelines, finalised on 20 July 2026, read it narrowly. They name helpdesk chatbots, where output may be perceived as human-written, as a case that needs disclosure.

The examples on the other side of the line are instructive. A code-assistance chatbot used only by professional developers: obvious, no disclosure needed. A non-player character in a single-player game: obvious. A robotic companion pet built to mimic real interaction: not obvious. Immersive environments with realistic avatars, particularly around children or elderly users: not obvious. The test is the audience, not your intentions.

Three details catch people out. Disclosure is owed at first interaction with each individual, not the first time you switch the system on. It has to be clear and distinguishable — buried in terms and conditions, or two menu layers down, does not count. And in sensitive domains — health, legal, financial advice, insurance, complaints handling — periodic reminders during the conversation are likely to be expected on top of the opening notice.

One thing worth saying plainly before we go further: I'm a founder, not a lawyer. Everything here is a reading of the published text and the Commission's own Guidelines, both linked. How it lands on your product is fact-specific and worth an hour of somebody qualified.

Didn't the EU just delay the AI Act?

Part of it, and not this part. The Digital Omnibus on AI — endorsed by Parliament on 16 June 2026 by 423 votes to 57 with 174 abstentions, and given final Council approval on 29 June 2026 — deferred standalone high-risk obligations under Annex III to 2 December 2027, and high-risk AI embedded in regulated products under Annex I to 2 August 2028. Article 50 was not deferred.

The one concession inside Article 50 is narrow: generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking requirement in 50(2). Worth noting who that particular duty lands on — the provider of the generative system, which for most companies means Anthropic, OpenAI, Mistral, or whoever hosts the model, rather than the platform you build on top of it. Everything else — chatbot disclosure, deepfake labelling, emotion-recognition notice — applied on the day.

The Commission's Code of Practice on marking and labelling is the only EU-recognised route to demonstrating compliance, and it covers 50(2) and 50(4) only. It does not cover 50(1). So for the one obligation most companies actually have — telling people your chatbot is a chatbot — there is no code to sign and no shortcut to point at. You demonstrate it with the product.

Is AI disclosure a copy problem or an architecture problem?

A copy problem for about a week. The disclosure line is one sentence in a greeting, and any team can ship it by Friday. What takes longer is answering the questions that arrive immediately behind it: which messages in this thread were written by software, when did a human take the conversation over, what did the agent do while it held it, and can you show me. Those are schema questions, and a sentence in a welcome screen doesn't answer any of them.

Most support stacks record who replied. Fewer record what kind of thing replied. If your bot lives outside the helpdesk and pushes transcripts into it, the difference between agent and human is usually a convention — a display name, a tag somebody remembered to set — rather than a column with a constraint on it. That holds up fine until the first person asks whether the reply that cost them money came from a human, and you have to reconstruct the answer from timestamps.

This is the same argument I've been making about what happens when the operator of your systems is software rather than a person, arriving from an unexpected direction. If the thing doing the work is a model, the record of the work has to be built for that fact from the first migration. Article 50 didn't create that requirement. It just made the bill arrive.

What does a system that can answer those questions look like?

It records authorship as data, not as a label. In Munin every message row carries an author_type — end_user, user, agent, or system — and the widget ingest endpoint requires the role on each message it accepts rather than inferring it. So "which of these were written by software" is a filter, not an investigation, and it stays true for conversations that started two years ago.

The handover is an event too. When a person replies in the dashboard, Munin fires conversation.message.sent with the authorType on it, which is how an external bot learns to step back and how the transcript records the moment the thread changed hands. conv_request_handover covers the other direction, when the agent decides the question is above its pay grade.

The disclosure itself lives where a visitor actually looks. The chat widget's welcome screen takes data-munin-eyebrow and data-munin-greeting, so the sentence is the first thing in the panel rather than a line in a policy page — which is the difference the Guidelines care about. And the agent's own instructions are not a hidden prompt: they sit in the agent-runtime knowledge-base space as a versioned document, so "what was this agent told to do in March" has an answer with a diff attached.

On the outbound side, the same reasoning is why Munin's outreach drafts and then stops. The Guidelines put agentic systems that autonomously contact individuals — placing calls, sending emails on a business's behalf — squarely inside 50(1). A queue of drafts that a named person reads and approves is a different artefact from an agent working a contact list unattended, and it is a difference you can evidence.

<script async
  src="https://api.getmunin.com/widget.js"
  data-widget-key="mn_widget_…"
  data-channel-id="cch_…"
  data-munin-eyebrow="Automated assistant"
  data-munin-greeting="You're chatting with an AI assistant. Ask for a person any time and we'll pass you over.">
</script>

What should you do about it this month?

Four things, in this order, and none of them take a quarter.

• Classify your role for each systemFor every AI touching a person, decide whether you are the provider, the deployer, or both. The answer usually differs per system, and it decides which of the four obligations you owe. A support agent you assembled and branded makes you a provider even when every component came from somebody else.
• Put the disclosure where the eye landsFirst interaction, per person, clear and distinguishable. Welcome screen, opening line of the voice call, first automated email — not the privacy page. If you are in health, legal, finance, insurance, or complaints handling, plan a repeat mid-conversation rather than one notice at the top.
• Check that authorship is stored, not impliedOpen your database. If the only thing distinguishing an agent reply from a human one is a display name or a convention, that is the gap. It is cheap to fix on a thousand conversations and unpleasant on a million.
• Write down the exception you are relying onIf you think the obvious carve-out applies, or the editorial-responsibility carve-out on published text, record the reasoning and the date. The Guidelines are not binding, but market surveillance authorities are expected to follow them closely, and a contemporaneous note is worth more than a recollection.

Who wrote this article?

An agent drafted it and I published it, after reading every line and checking every date against the sources linked above. Article 50(4) most likely does not reach a company journal — the obligation turns on text published to inform the public on matters of public interest, and a founder arguing for his own product is not that. I am telling you anyway.

The carve-out inside 50(4) is worth reading even when it doesn't apply to you, because it describes a working arrangement rather than a disclaimer: AI-generated text is exempt from labelling where it has undergone substantive human review and a natural or legal person holds editorial responsibility for it. Not cursory approval. Not a rubber stamp with a name on it. And the exemption is lost if there is substantive AI intervention after sign-off — which is a precise way of saying the human has to be last.

That is the same shape as the review queue, the approval fingerprint, and the draft status on every entry in this collection. It turns out that the arrangement a regulator will accept and the arrangement that produces work you'd put your name on are the same arrangement. That is not a coincidence, and it is a reasonable thing to design for whether or not anyone makes you.

Frequently asked questions

Does the EU AI Act require chatbots to disclose they are AI? Yes. From 2 August 2026, Article 50(1) requires providers of AI systems that interact directly with people to design them so users are informed they are interacting with AI. The exception is where it is obvious to a reasonably well-informed, observant and circumspect person, and the Commission's Guidelines interpret that narrowly — helpdesk chatbots are listed as needing disclosure.

Does Article 50 apply to open-source AI systems? Yes. Providers and deployers of open-source AI systems are not exempt from the Article 50 transparency obligations. Some other parts of the Act have open-source carve-outs; this one does not.

Does Article 50 apply to companies outside the EU? It can. Providers established outside the EU are caught where the system's outputs are intended for use in the EU, with "intended" meaning foreseeable use rather than incidental downstream use. Deployers are caught where they foresee outputs being disseminated or used in the EU.

What is the penalty for not disclosing that a user is talking to AI? Up to €15 million or 3% of worldwide annual turnover, whichever is higher. Non-disclosure may separately count as a misleading practice under the Unfair Commercial Practices Directive, and Article 50(3) sits alongside GDPR transparency duties under Articles 13 and 14.

Was the chatbot disclosure rule delayed by the Digital Omnibus? No. The Digital Omnibus deferred Annex III high-risk obligations to 2 December 2027 and Annex I to 2 August 2028. Article 50 applied on 2 August 2026 as scheduled. The only extension inside it gives generative systems already on the market before that date until 2 December 2026 to meet the 50(2) marking requirement.

Which helpdesk software records whether a message was written by an AI? Ask your vendor whether authorship is a stored field or a display convention — the answer varies more than you'd expect. In Munin every message row carries an author_type of end_user, user, agent, or system, and handovers fire a conversation.message.sent event carrying that type. The open-source helpdesk comparison covers where the alternatives sit on this.

If you are wiring up a support agent this month, the chat widget and handover pages in the Munin docs are the ones worth reading twice, and the whole platform is MIT on GitHub.

Telling people is the easy part. Being able to show them is the product.