The head of the customer contact centre wants a chatbot on the council website, so residents can get answers about bin collections and parking permits in the evening too. Two floors up, a communications officer asks ChatGPT for a first draft of a news item about roadworks. She corrects a date and publishes it.
Since 2 August 2026, both situations fall under Article 50 of the AI Act. What differs is who has to act, and when a disclosure is needed. This article is written for municipalities and other public bodies in the EU.
#What Article 50 asks of a municipality
Article 50 of the AI Act (Checked 16 November 2026) contains four transparency obligations. Two fall on the provider of an AI system, two on the organisation using it. The AI Act calls that organisation the deployer, and public authorities are named in the definition.
Under paragraph 5, the information has to be clear and distinguishable at the latest at the first interaction or exposure, and meet accessibility requirements. The Digital Omnibus left the 2 August 2026 date in place, as we explain in the AI Act timeline after the Digital Omnibus.
An employee using ChatGPT is not a deployer in their own right. In its questions and answers on Article 50 (Checked 16 November 2026), the European Commission says that employees acting under the instructions and control of an organisation are not separate deployers. The municipality is. That makes Article 50 a policy question, not something for the communications officer to settle alone.
#A chatbot for residents: often the municipality is the provider
The duty in paragraph 1 falls on the provider. Most people think of the supplier, but the definition in Article 3 is wider. A public authority is a provider when it has an AI system developed and puts it into service under its own name or trademark. A chatbot a supplier built for the municipality, presented on the website as the council's assistant, makes the municipality responsible for the notice.
The notice can only be left out when it is obvious to a reasonably well-informed, observant resident that they are talking to AI. The Commission asks for that exception to be read narrowly, because it takes transparency away from people. A chatbot with a friendly name and a staff photo will not look like AI to many residents. Say it in the first line of the conversation, in plain words, and make sure a screen reader reads it out.
Settle who does what during procurement. The supplier delivers the technology; the municipality decides how the chatbot introduces itself to residents.
#A text for residents: when do you disclose that AI wrote it?
Paragraph 4 requires a disclosure for text generated or manipulated by AI when it is published to inform the public on matters of public interest. The Commission lists three conditions: the text is published, it is meant to inform the public, and it concerns a matter of public interest. Its examples of such matters include politics and democratic processes, public health, and public administration and services.
For a municipality, that covers much of what it publishes. A news item about roadworks, a page about waste collection, a summary of a council decision on the website. The news item from the opening falls under it.
What is not published stays outside this duty. An internal memo, a draft briefing for a councillor or an email to a colleague needs no disclosure. Those documents are still subject to the GDPR and to your own rules, such as which tools are allowed and which data must stay out. How ChatGPT fits the rest of the AI Act is covered in ChatGPT and the AI Act.
#A spelling check is not human review
No disclosure is needed when the text has gone through human review or editorial control and a person or organisation holds editorial responsibility. That exception is narrower than it sounds.
The Commission describes human review as a deliberate examination of the substance by someone with relevant knowledge and professional judgement. Editorial control means control by a responsible editorial entity, such as an editor-in-chief, with the authority to approve, change or reject the text on substantive grounds, including fact-checking and checking sources. Spelling and grammar checks explicitly do not count. Correcting a date and hitting publish, as the communications officer did, is not review.
Have the project lead for the roadworks check the substance, and record that the communications team signs off on the publication. Then no disclosure is needed. Skip that step, and the item gets a line saying AI produced it.
#How to make AI text recognisable
The law does not prescribe wording. On 10 June 2026 the European Commission published a code of practice (Checked 16 November 2026) on marking and labelling AI-generated content. It comes with free EU icons (Checked 16 November 2026) for text that is fully AI-generated and for text that AI partly modified. Both the code and the icons are voluntary. The Commission adds that the Article 50 duty is not, and that using the icons does not by itself establish compliance.
A short, fixed sentence with the item does the job as well. What matters is that residents see it when they first read the text, and that it reads the same on every page.
#What changes on 2 December 2026
On 2 December the transition period ends for generative AI systems placed on the market before 2 August 2026. Their providers must then mark output in a machine-readable format, as paragraph 2 requires. That is work for the makers of the tools, not for the municipality using them.
For the municipality, nothing about Article 50 changes that day. The chatbot and published texts have been covered since 2 August. According to the Commission, content generated before 2 August 2026 does not need to be labelled retroactively.
On fines, the AI Act provides that breaching Article 50 can lead to a fine of up to 15 million euros, or 3 percent of worldwide annual turnover for an undertaking if that is higher. To what extent public authorities can be fined is left to each member state under Article 99(8).
#What to record in your policy and AI register now
Article 50 is about uses, not tools. The same Copilot licence can draft an internal memo, which needs nothing, and a web page, which does. Four arrangements make it workable.
- Map where AI writes for residents. The website, newsletters, social media and the chatbot. For setting up the register itself, see an AI register for municipalities.
- Record who the provider is for each use. For your own chatbot that is often the municipality; for a general tool, the supplier.
- Agree who reviews. A colleague with subject knowledge for each topic, and a team that signs off publication. Without that step, the text gets a disclosure.
- Pick one standard disclosure. A sentence or an EU icon, the same everywhere, so residents recognise it.
What belongs in an AI policy about contact with residents is set out in AI policy for municipalities.
These arrangements only hold if the municipality knows which AI tools are in use, and for what. With Decide you record a decision per tool in the AI register, with its use cases, the people responsible and an indicative AI Act classification; which obligations apply depends on how you use the tool. See shows in counts which tools are actually in use, including the ones nobody has decided on yet. For the people who keep those arrangements on track, there is AI governance for privacy officers.