---
title: "AI Act transparency: what chatbots and agents must disclose"
description: "On 12 November the EU Official Journal published the Commission's guidelines on Article 50 of the AI Act: who is a provider and who a deployer, what chatbots and agents must disclose, marking and detection of generated content, deepfakes and texts on matters of public interest, and the 2 December deadline."
url: "https://www.noze.it/en/insights/ai-act-transparency-chatbots-agents/"
lang: "en"
alternate: "https://www.noze.it/insights/trasparenza-ai-act-chatbot-agenti/"
published: "2026-11-20"
author: "noze S.r.l."
keywords: "AI, Cybersecurity, AI Act, Article 50, Transparency, AI Agents, Chatbots, Deepfakes, Watermarking, Code of Practice, Compliance, Regulation 2024/1689"
---

# AI Act transparency: what chatbots and agents must disclose

On 12 November the EU Official Journal published the Commission's guidelines on Article 50 of the AI Act: who is a provider and who a deployer, what chatbots and agents must disclose, marking and detection of generated content, deepfakes and texts on matters of public interest, and the 2 December deadline.

20 November 2026

![Sophia, a humanoid robot with a realistic face and a transparent skull cap showing its circuitry, lit in red and blue backstage at Web Summit 2018](https://www.noze.it/images/insights/trasparenza-ai-act-chatbot-agenti.jpg)

Sophia, the Hanson Robotics humanoid robot, at Web Summit 2018. Image: Web Summit, CC BY 2.0 (Wikimedia Commons).

On 12 November the Official Journal of the European Union, C series, published the **Commission’s guidelines on the transparency obligations in Article 50 of the AI Act** (C/2026/6005). The 43 pages set out definitions, criteria and examples for anyone who builds or uses chatbots, agents, generative systems, emotion recognition or biometric categorisation. The Commission had put the English version online on 20 July, before the obligations started to apply on 2 August.

The guidelines are not binding: only the Court of Justice of the EU can give the authoritative interpretation of the regulation. They are, however, also written for the competent authorities, so that Article 50 is applied consistently and uniformly.

## Four obligations and who is responsible

| Paragraph | Systems                                                          | Who      | Obligation                                                         |
| --------- | ---------------------------------------------------------------- | -------- | ------------------------------------------------------------------ |
| 50(1)     | systems that interact directly with people                       | provider | inform people that they are interacting with an AI system          |
| 50(2)     | systems that generate or manipulate audio, images, video or text | provider | mark outputs in a machine-readable format and make them detectable |
| 50(3)     | emotion recognition and biometric categorisation                 | deployer | inform the people exposed                                          |
| 50(4)     | deepfakes and texts published on matters of public interest      | deployer | disclose that the content is generated or manipulated              |

Paragraph 5 applies to all of them: the information must be given in a clear and distinguishable manner, at the latest at the first interaction or exposure, and must meet the applicable accessibility requirements. The obligations add up: the provider of an assistant that generates images in a conversation may have to comply with both paragraph 1 and paragraph 2, and a deployer who uses it professionally to produce a deepfake may also have the obligation in paragraph 4.

## Provider or deployer

The guidelines spend a lot of space on roles, because they decide who has to do what.

- **Modifying and rebranding makes you a provider.** A company that modifies a third party’s generative system, for example with new training data, and puts it into service under its own name becomes the provider of the new system. So does a public body that develops a chatbot in-house and uses it itself under its own name.
- **The deployer is whoever decides the use.** What counts is authority over the system, meaning the decision on purpose and manner of use, not technical control. Employees working under the organisation’s instructions are not separate deployers. A company that simply commissions an advert from an agency, without deciding whether and how the agency uses AI, is not a deployer.
- **Hosting platforms are not deployers.** Those who only distribute content generated by others, without authority over the use of the AI system, have no obligations under Article 50, but are strongly encouraged to preserve marks and labels. Very large platforms become providers or deployers when they integrate or use AI systems for their own purposes.
- **Open source.** AI systems released under free and open-source licences remain subject to Article 50. Components that are not a system on their own, such as software, data, models, tools and services, are not.

Article 50 does not apply directly to general-purpose AI models, but the Commission encourages developers of these and other generative models to build identification of interactive systems and marking of content in at model level, to make the work of downstream providers easier.

## Chatbots and agents: paragraph 1

Paragraph 1 covers systems intended to interact directly with people in a two-way exchange: voice assistants, chatbots, avatars, robots, AI companions, robotic pets, social media bots and also **coding agents**. It does not cover recommender systems, spam filters, machine translation and transcription, text and code autocompletion, back-end decision support systems or machine-to-machine calls.

On **agents** the guidelines are explicit (point 31). An agent that interacts with people while booking, buying, handling correspondence or negotiating must disclose its artificial nature **and the person on whose behalf it is acting**. If the provider cannot know in advance whether the agent will meet people, the agent should be designed at the level of its architecture to disclose itself whenever that is reasonably likely. It should also disclose itself to whoever instructs it, at key moments such as authorisations, notifications and validations, and in every new interaction. In a footnote the Commission points to attestations in European Digital Identity Wallets and the proposed European Business Wallets as one way to identify an agent verifiably.

The form of the notice is free, provided it meets paragraph 5 and takes account of minors, older people and people with disabilities where the system may meet them. Some solutions are not enough on their own:

- a notice only in the terms of use, in a URL or in the documentation;
- a machine-readable mark the user cannot see;
- generic references such as “assistant”, or an interface that makes the system look human;
- a general note such as “the services on this site use AI”;
- technical descriptions such as “this system uses LLMs”.

In most cases a clearly visible notice before the first interaction is likely to be sufficient; the Commission recommends placing it next to the input field. Periodic reminders are likely to be needed with vulnerable people, in sensitive or immersive contexts and where the risk of being misled is higher, such as financial advice, insurance, legal assistance, healthcare, complaint handling and AI companions. The system must in any case disclose itself whenever someone asks whether it is an AI or when the exchange suggests the person may be confused.

The exception for “obvious” interaction is to be read narrowly: it covers a coding assistant reserved for professional developers or an internal assistant used by trained staff (HR, legal, procurement, compliance, IT), while a customer support chatbot open to the public must disclose itself.

## Marking and detection: paragraph 2

The provider of a generative system must do two things: **mark** outputs in a machine-readable format and ensure they are **detectable**. One without the other is not enough. Detection tools must be publicly accessible, based on industry standards and ideally able to run locally on the device. Until harmonised standards exist, the provider may use its own or a third-party solution, as long as it is interoperable with those of others.

The techniques mentioned in recital 133 include watermarks, metadata, cryptographic provenance methods, logging and fingerprints, also in combination. There is no need to keep the full provenance chain of the content. The solution must be effective, reliable, robust against modifications and attacks and interoperable, in line with the state of the art. The limit of what is “technically feasible” is objective and does not depend on the resources of the individual provider. Marking should not process information about the creator of the content.

Outside paragraph 2:

- short outputs such as single words, captions, alt text and interface labels;
- **source code**, comments included, as well as SQL, infrastructure as code, YAML, JSON, scripts and configurations;
- agent-to-agent communication that does not reach people, and an agent’s intermediate steps, such as reasoning or browser actions;
- intermediate work in closed production workflows, where only the final result has to be marked.

Standard editing does not need marking either: grammar fixes, formatting, compression, small colour adjustments and AI-generated **translations**, which may however fall under paragraph 4 when they are texts on matters of public interest. Summaries, paraphrases that change style or meaning, face and object replacement and realistic speech synthesis in a specific person’s voice do need it.

For industrial and business-to-business applications, excluding systems aimed at the public and consumers, there is an exemption with three cumulative conditions: strictly technical output, intended for a small predefined group of people inside the provider’s and deployer’s organisation, and not shared outside the company, with safeguards against misuse such as cloud isolation and role-based access controls.

## Deepfakes and texts: paragraph 4

A **deepfake** is a generated or manipulated image, audio or video that appreciably resembles real or plausible persons, objects, places, entities or events and could falsely appear authentic or truthful. Among the Commission’s examples, a realistic synthetic avatar of a chief executive congratulating staff is a deepfake, and so is an advertising image of a product that makes it look better than it is; a dragon or an elephant driving a car is not realistic and falls outside. The label must be visible or audible without technical tools: the deployer cannot rely on the provider’s machine-readable mark. For evidently artistic, creative, satirical or fictional works the obligation is lighter, but if content also has an informative character the full obligation prevails.

For **texts** the obligation covers those published to inform the public on matters of public interest, including public administration and public services. One of the Commission’s examples is an AI-generated summary of an article about a city council decision. Texts on a company intranet are not published.

No label is needed if there is **human review or editorial control** together with a natural or legal person holding **editorial responsibility**, meaning final legal responsibility for the publication; their identity and contact details should be made public. The review must go into substance, by people with the relevant knowledge and professional judgement, and includes at least fact-checking. A spell check, the mere existence of an editorial policy, an automated review or a cursory approval are not enough. If AI substantially changes the text after approval, the exception no longer applies.

## Code of practice and deadlines

The **Code of Practice on Transparency of AI-generated Content**, published on 10 June, was found adequate by the Commission on 8 July for paragraphs 2, 4 and 5. By the end of July about 190 organisations had signed it. Signatories are checked on how they implement the code. Non-signatories must demonstrate compliance by other means, for example a gap analysis against the code, and may receive more requests for information. The EU has also published a set of icons for labelling content, whose use is optional.

Fines go up to 15 million euros or 3% of worldwide annual turnover in the previous financial year, whichever is higher. For SMEs the lower of the two applies.

Article 50 has applied since 2 August 2026 to all systems, including those already on the market. The Digital Omnibus (Regulation (EU) 2026/1744) grants a transition period only for marking and detection (paragraph 2) of generative systems placed on the market or put into service before 2 August, which must comply by **2 December 2026**. For systems that are partly interactive and partly generative the extension covers only marking: the notice of interaction with AI has been due since 2 August. Content produced before 2 August does not need to be marked or labelled retroactively, but a text generated before and published on or after 2 August must be labelled.

## Our take

For anyone building agents, point 31 is the most practical. An agent that writes emails or replies on behalf of an office, when there is a person on the other side, must present itself as an agent and say who it works for, and it should do so towards whoever instructs it at authorisation steps as well. When the counterparties cannot be foreseen, the guidelines ask for the agent to be designed at the level of its architecture to disclose itself: it has to be built into how the agent works, not only into its instructions.

For the public sector, the section on texts is the one that matters. In [IntelliPA](https://www.noze.it/en/products/intellipa/) drafts of administrative acts show where the text comes from and the officer in charge verifies them before signing. The guidelines make clear that this check counts as human review if it goes into substance, with verification of facts and sources, and that a substantial AI change after approval voids the exception. The public body must also hold editorial responsibility for the publication.

For providers of generative systems already on the market, there are twelve days left until 2 December. What was already due from 2 August is in the [AI Act checklist](https://www.noze.it/en/insights/ai-act-checklist-obligations-august-2026/); the Omnibus timeline is in [the Digital Omnibus towards formal adoption](https://www.noze.it/en/insights/ai-act-omnibus-adoption/).

## Sources

- [Official Journal of the EU, C/2026/6005, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ:C_202606005)
- [European Commission, Guidelines on transparency obligations for providers and deployers of AI systems](https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems)
- [European Commission, Code of Practice on Transparency of AI-generated Content](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content)
- [European Commission, opinion on the assessment of the code of practice](https://digital-strategy.ec.europa.eu/en/library/commission-opinion-assessment-code-practice-transparency-ai-generated-content)
- [European Commission, EU icons for labelling AI-generated content](https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content)
- [Regulation (EU) 2024/1689, AI Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng)
- [Regulation (EU) 2026/1744, Digital Omnibus on AI](http://data.europa.eu/eli/reg/2026/1744/oj)
