AI texts and AI chatbots in your online store: what do you have to label?

Research as of 1 October 2026. This guide covers texts and conversations: product descriptions, category and SEO copy, FAQs, newsletters and blog posts that an AI writes or revises, and AI chats or voicebots in customer service. Different standards apply to AI images; they are covered in the guide Do I have to label AI images in my online store?, and video and audio in Labelling AI videos in your online store and in ads. This guide reflects the current state of interpretation and is general information, not legal advice. For the wider picture, see Article 50 for online stores.

What does Article 50 of the AI Act require for texts and chatbots?

Article 50 of the AI Act (Regulation (EU) 2024/1689) contains several transparency obligations, each with its own addressee. For texts and chats, these paragraphs matter:

Under Article 3(4), a deployer is anyone “using an AI system under its authority”; that is you when you use a text tool or a chat widget in your store. Under Article 3(3), a provider is anyone who develops an AI system “or that has an AI system … developed” and places it on the market or puts it into service under its own name or trademark.

The difference from images. For images, the obligation falls on the store directly as the deployer: every AI image that is a deep fake needs a visible label, advertising included. For texts it is the other way round: under the guidelines, advertising and product copy is usually outside Article 50, and texts on matters of public interest only fall under it without human review. And for a chatbot, the obligation is addressed to the provider of the system, not to the store using it; in practice, you still have to make sure the notice reaches your customers.

The Digital Omnibus (Regulation (EU) 2026/1744) has not changed this. Article 50 was only amended in paragraph 7 (Article 1 point 20), Article 3 only in point 14, with new points 14a and 14b (Article 1 point 4). The deadline of 2 December 2026 in the new Article 111(4) (Article 1 point 39(b)) only concerns providers of systems placed on the market before 2 August 2026, and only the machine-readable marking under paragraph 2. For chatbots, the guidelines make clear that even where a system also generates content, the notice under paragraph 1 has applied since 2 August 2026 (para. 153). A consolidated version of the AI Act was not available on 1 October 2026.

Does your chat widget have to say that it is an AI?

The first sentence of Article 50(1) reads: “Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.”

The European Commission’s guidelines on Article 50 (annex to C(2026) 5054 final of 20 July 2026) examine four elements for this (para. 30):

1. An AI system. Excluded are “simple non-AI automated response mechanisms (e.g., traditional out-of-office emails, rule-based quick message answers)”.
2. Intended to interact. This means an exchange in both directions, expressly including via “online interfaces, phone or e-mail”.
3. Directly. There is no direct interaction “if customer service representatives use AI assistance tools to help them communicate with natural persons”, in other words if your staff only use AI to help them write. The mere possibility that a human intervenes is not enough to get round the obligation, though. Where AI answers and human content are mixed, the AI answers have to be disclosed “unless those AI outputs have been properly reviewed and sent by humans as the main interlocutors with the natural persons”.
4. With natural persons. Purely machine-to-machine processes are excluded.

As examples within the scope, the guidelines name “AI-enabled voice assistants, chatbots/conversational agents in various contexts (e.g. public service, customer support, complaints management, e-commerce, finance, healthcare, education etc.)”. For typical store tools, this gives:

Who owes the notice: the widget provider or your store?

On the wording and the guidelines, the provider. According to para. 28 of the guidelines, Article 50(1) “is addressed to providers of AI systems directly interacting with natural persons”. The Commission’s FAQ on Article 50 (page as of 24 July 2026) lists only paragraphs 3 and 4 as deployer obligations, and the Bundesnetzagentur, Germany’s market surveillance authority for the AI Act, places paragraph 1 under provider obligations (“Anbieterpflichten”) in its overview.

The AI Service Desk at RTR, the Austrian regulator for broadcasting and telecommunications, writes in its FAQ on chatbots, translated from German: “This design obligation falls on the provider of the chatbot; if you merely use a bought-in chatbot as a deployer, you should make sure that the information provided for by the provider actually remains perceptible in use.” The AI Service Desk informs and advises; it is not a supervisory authority.

The Wettbewerbszentrale, the German industry body against unfair competition, goes further in its guide (version 2.0, as of 29 July 2026). It first names the software manufacturer as the provider, but then continues (our translation): “If company U acquires a chatbot from a software manufacturer, names it Uwe and integrates it on its website, the company is responsible for the labelling as the provider.” The guide does not give a reasoning based on the legal text.

When you are certainly the provider. The guidelines expressly name as providers a company that develops a chatbot “in-house” and puts it into service for its own use under its name, and a company that modifies an existing system, “e.g. with new training data”, and puts it into service under its own name (para. 11). If you have an agency develop a chatbot for you and use it under your name, the wording “has … developed” in Article 3(3) fits.

Assessment. On the wording of Article 3(3) and RTR’s answer, there is much to suggest that a store which only integrates a finished chat widget, gives it a name and fills it with its own content such as FAQs remains a deployer (our own reading). The guidelines do not say where adjustments start to make you a provider, beyond the example of “new training data”. In practice the dispute hardly matters: on every one of these views, the notice has to reach your customers at the start of every conversation. The only difference is whether you are the addressee of the obligation yourself or make sure that the provider’s solution stays visible in your store.

When is the AI interaction “obvious”?

No notice is needed if the AI interaction is obvious to a reasonably well-informed, observant and circumspect person. The guidelines interpret this exception narrowly (para. 45):

• The yardstick is the audience the system is likely to talk to (para. 44). If the general public, including children, older people or people with little AI experience, can use the chat, the provider cannot rely on the exception (para. 45).
• “The general awareness of consumers and other natural persons that AI systems (including chatbots and agents) exist does not imply that they recognise them in interactions.”
• The exception should be limited to cases “where there is almost no doubt left about the nature of the interaction”. Voice (robot voice or human-sounding) and design, such as a profile picture showing a human, also feed into the assessment.
• As an example where the exception does not apply, the guidelines give “AI chatbots embedded in online platforms or assistance support tools (helpdesks) whereby users directly interact and receive AI outputs (e.g. replies to queries or other AI-generated content) they may perceive as human-generated”.

RTR sums it up as follows (translated from German): “For chatbots for the general public, e.g. in customer service, this exception is regularly not met.” The Wettbewerbszentrale considers it possible that a bot’s name or appearance already makes clear that it is a bot, but advises (our translation): “Anyone who wants to be sure that no false impressions arise should label the bot if in doubt.” For a store chat with consumers, you should not rely on the exception.

When and how must the notice appear?

Timing. Article 50(5): “The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The information shall conform to the applicable accessibility requirements.” For a chat, para. 143 requires the information “at least” to be provided “once at the start of an interactive session”; the guidelines give the example “when launching a conversation with a chatbot”. Under para. 40, a single prominent notification before the first interaction is likely to suffice in most instances. In riskier contexts, such as health, financial or legal questions, complaints handling or vulnerable users, the guidelines consider periodic reminders likely to be necessary. If a customer asks whether they are talking to a human, the system has to disclose its AI nature (para. 40).

Form. Paragraph 1 does not prescribe any particular technique (para. 34). As examples, the guidelines give a chatbot that starts a conversation by mentioning that it is based on AI (para. 36), plain-language banners such as “You are interacting with an AI system”, a first-turn greeting from the chatbot and a persistent label near the input field (para. 37). Wording examples from the German-language bodies, translated:

• RTR: “You are interacting with an AI system”.
• Wettbewerbszentrale: “You are chatting with an AI assistant here.”

What is not enough. Under para. 38, the following are insufficient on their own: disclosures only in terms and conditions, URLs or documentation; machine-readable markings that cannot be perceived in the conversation; unclear signals such as a mere reference to an “assistant”, or human-like representations that may mislead; general statements such as “Services on this website use AI”; purely technical descriptions such as “this system uses LLMs”. The Wettbewerbszentrale also advises against labelling a bot only as supported by AI (“von KI unterstützt”), because a human customer service team can also be supported by AI at times.

Accessibility. Recital 132 on chatbots ends with the sentence: “Such information and notifications should be provided in accessible formats for persons with disabilities.” If the chat may also reach children, the notices should be child-friendly and easy to understand (para. 34). Which accessibility rules apply to a store is explained in the guide Is your AI label accessible?

What applies to phone voicebots?

An AI voicebot that takes calls falls under Article 50(1); the guidelines expressly mention the phone as a means of interaction (para. 30) and name AI voice assistants as an example. On voicebots, RTR says (translated from German): “That is, if a conversation is handled by an AI, it would have to be disclosed that one is interacting with an AI.”

For the form, the guidelines (para. 37) recommend “explicit spoken statements at the beginning of the interaction (e.g. “This is an AI-powered assistant”)”, supplemented in longer calls by reminders, especially after interruptions or when the role of the system changes. Audio cues can help, “but are not considered sufficient by themselves”. Whether the voice sounds like a robot or like a human feeds into the question whether the AI is obvious (para. 45).

A real person’s voice. If the voicebot speaks with the cloned voice of a real person, such as the owner, the question of an audio deep fake arises as well (Article 50(4), first subparagraph, Article 3(60)). RTR gives as an example the case where the AI generates a well-known human voice; the guidelines list cloned voices of a podcast’s presenters as a deep fake, but not an AI voice for fictional characters where there is no deception as to the identity of the narrators. This obligation falls on you as the deployer. More on audio in Labelling AI videos in your online store and in ads.

Do you have to label AI texts in your store?

The second subparagraph of Article 50(4) reads: “Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”

The guidelines break this down into three elements (para. 131):

• Published: accessible by “an indeterminate, fairly large number of unrelated, potential readers”. Not published are, for example, “private, interpersonal correspondence (for professional purposes)” and internal texts. A reply to a single customer email is therefore not a published text.
• Informing the public: the text is meant to communicate knowledge, opinions or facts; short texts that do not do so materially do not inform the public.
• Matters of public interest: topics relevant to society, such as politics, public administration, justice, fundamental rights, public security, public health, environmental protection, consumer safety and economic, financial, scientific or cultural developments that may be a subject of public debate.

As examples outside the obligation, the guidelines give, among others, “AI-manipulated text that is part of a company’s advertisement or product descriptions (not including any claims related to e.g. health, consumer safety or sustainability)”. The Bundesnetzagentur’s German summary puts it the same way. As an example within the obligation, the guidelines give, among others, “AI-manipulated parts of a lifestyle-website article comparing the effects of various diets on a particular disease in middle-aged women”.

The Wettbewerbszentrale reaches the same result (our translation): “This obligation is unlikely to be very significant in (advertising) practice.” And: “Advertising texts are therefore likely to require labelling only in exceptional cases at most, since they will often lack the relevance to the public referred to.”

When does the obligation for texts not apply?

Even a text on a matter of public interest does not need a label if two conditions are met together (para. 133):

1. Human review or editorial control. This means a deliberate examination of the substance by people with relevant expertise; “Fact-checking the accuracy of the content is a minimum requirement that should be part of that review” (para. 134). Not enough are “Superficial, solely formal or procedural checks (e.g. spell-checking or grammatical correction)”, the mere existence of an editorial policy, automated review processes or a cursory approval (para. 135). If an AI changes the substance of the text after sign-off, the exception no longer applies (para. 136).
2. Editorial responsibility. A natural or legal person must hold the ultimate legal responsibility for the publication; their identity and contact details should be publicly available in an easily findable location, online for instance in the terms and conditions or other user-facing legal information (para. 138). For a store, it is natural that the company holds this responsibility and is named in the legal notice (our own reading).

The guidelines consider the conditions met, for example, for an AI-generated sustainability report reviewed by professionals and for an AI translation reviewed by a human. They are not met for articles reviewed by another AI system where a human only performs a superficial grammar check. For advertising texts that are relevant to the public in an individual case, the Wettbewerbszentrale considers it sufficient “if appropriately trained people take over the final check and examine texts for errors” (our translation).

Signatories of the Commission’s Code of Practice on Transparency of AI-Generated Content commit, for such texts, to an internal policy that names at least the responsible person with name, role and contact details and the organisational measures for the review; individual reviews do not have to be documented (Commitment 4). The code only binds signatories, but it is a useful orientation.

Which store texts are affected?

From the guidelines’ examples, typical store texts can be classified as follows. This is a reading of this guide, not a decision:

The guidelines state the exception for advertising and product copy and the proviso for health, consumer safety and sustainability, but do not explain how far that proviso reaches. If you word such claims with AI, you had better review them for substance anyway, also because of unfair competition law; for texts on matters of public interest, the exception then applies as well.

How do you label a text where needed?

The disclosure has to be clear and perceivable, without technical tools and without the reader having to do anything (para. 132), and at the latest at first exposure; the guidelines give the example “at the start of an AI-generated or manipulated text publication on matters of public interest” (para. 143 and examples after it). Signatories of the code place the icon or an equivalent label “for example above or at the top of the text, near the headline of the text, or in the colophon at the beginning of the text” and may label only the AI part (Sub-measure 1.2.2(f)). The EU icons are also intended for published text (Annex 1 of the code); details in The EU icons for AI images. The Wettbewerbszentrale shows two text examples placed before the article, translated: “This text was edited with the help of an AI.” and “AI-generated text”.

For older texts, para. 154 says: texts generated and published before 2 August 2026 do not need to be labelled retroactively. “Conversely, if texts that have been generated or manipulated before 2 August 2026 are published on or after that date, they need to be labelled.” A blog post within the obligation that was written before the cut-off date and only goes online now therefore needs a notice, unless a person has reviewed it and holds responsibility for it.

And the machine-readable marking?

Article 50(2) also covers text: providers of systems that generate text have to mark the outputs in a machine-readable format. According to the guidelines’ examples on paras. 90 to 92, this includes AI summaries and rewrites that change meaning, style or structure, but not spelling and grammar corrections or AI translations; short sequences such as alt texts are excluded altogether (para. 68). This is an obligation of the providers of the text tools, not of the store; RTR writes that it does not concern you “directly” as a deployer (our translation). The exception is if you develop your own text system, or modify one with new training data, and use it under your name (paras. 11, 15).

What applies regardless of Article 50?

Article 50(6) leaves other transparency obligations unaffected, and for chatbots the guidelines stress that the information obligations under consumer protection law apply whether or not the AI interaction is “obvious” (para. 50). For your store, this means:

• What your chatbot says, you say. On 12 May 2026 (4 UKl 3/25, final according to the NRWE decisions database), the Higher Regional Court of Hamm ruled on the AI chatbot of a provider of beauty treatments, which attributed specialist medical titles to its managing directors working as doctors that they did not hold. The court treated the answers as commercial practices of the company: the behaviour of chatbots and social bots is also a commercial practice “of its operator, because the operator has sufficient influence on the system and sets it in motion” (our translation; operator here in the sense of the UWG). It rejected the argument that consumers know AI makes mistakes: a large part of the consumers addressed “place particular trust in the accuracy of the computer-generated answer” (our translation). The company is answerable under Section 8(2) UWG for the contractor who programmed the chatbot and put it into operation. The Wettbewerbszentrale advises setting chatbots “narrow limits” and checking regularly whether they keep to them (our translation); in practice, that means clear instructions on prices, delivery times, health and legal questions, and a test before going live.
• Wrong product information. If an AI text contains untrue or deceptive statements about essential characteristics, such as composition, fitness for purpose or possible uses, it is misleading in Germany under Section 5(2) no. 1 UWG. An AI notice does not cure this.
• Mandatory information stays mandatory. If you offer goods with their characteristics and price in a way that lets consumers buy, Section 5b(1) no. 1 UWG requires all essential characteristics of the goods; product-specific information obligations apply alongside. Who wrote the text makes no difference.
• Fake reviews. In Germany, the following are always prohibited towards consumers (Section 3(3) UWG): claiming that reviews come from real buyers without reasonable and proportionate checks (annex no. 23b), and submitting or commissioning fake consumer reviews or endorsements (annex no. 23c). If you make reviews available, information on whether and how you ensure that they are genuine counts as essential information (Section 5b(3) UWG). In Austria, the equivalent annex nos. 23b and 23c to the UWG are in all cases misleading (Section 2(2)), together with Section 2(6b). The Swiss UWG does not mention reviews expressly (as of 1 January 2025); Article 3(1)(b) against incorrect or misleading statements applies there. If you have an AI summarise genuine reviews, the summary has to be accurate (our own reading).
• AI literacy. Under the amended Article 4 of the AI Act (Article 1 point 5 of the Digital Omnibus), deployers take measures “to support the development of AI literacy of their staff” who work with AI systems; no specific level of AI literacy is required of any individual.
• Data protection. The information obligations under data protection law continue to apply alongside Article 50(1) (para. 52).

In Germany, the market surveillance authority for the AI Act is the Bundesnetzagentur; the Wettbewerbszentrale also considers breaches of the AI Act to be possible breaches of unfair competition law (Who enforces the EU AI image labelling rule in your store?). For stores based in Switzerland, the AI Act only applies under the conditions of Article 2(1)(c) (Labelling AI images as a Swiss shop).

Checklist for your store

This list is a recommendation of this guide, not the legal position.

Chat, email agent and voicebot

1. List all tools that talk or write to customers by themselves: chat widget, AI email agent, hotline.
2. Check for each one that the AI notice appears at the start of every conversation, on mobile as well as desktop, and that no setting hides it. Keep a dated screenshot.
3. The notice names the AI expressly, for example “You are chatting with an AI assistant”; not just “assistant” or “supported by AI”.
4. Do not give the bot a human name with a photo of a person unless the AI notice cannot be missed.
5. For a voicebot: spoken statement at the start, reminders in long calls and after a transfer.
6. Decide which topics the bot does not answer, and test typical questions before going live.
7. Clarify with the widget provider whether the notice is built in and who maintains it.

Texts

8. Review AI texts for substance, not only for spelling: product data, claims, facts.
9. Claims on health, safety and sustainability and blog posts on such topics are reviewed by a knowledgeable person; the company holds responsibility and is named in the legal notice. Otherwise: a notice at the start of the text.
10. After sign-off, no substantive AI changes without a new review.
11. No reviews or testimonials written by AI.

Documentation

12. Record for each tool: provider, purpose, where the notice appears, who reviews texts, dates of creation and publication. RTR expressly recommends documenting when AI content was generated and when it was published.

Easy AI Act Image Labels checks and labels images only; it has no function for texts, chatbots or voicebots.

What is still open?

• Role with a bought-in chatbot: RTR and the Wettbewerbszentrale see the store differently, as a deployer with a recommendation of care or as the provider. The guidelines do not address this case expressly and are non-binding (para. 5); only the Court of Justice of the EU can give a binding interpretation of the AI Act.
• Threshold for becoming a provider: the guidelines do not say whether your own name, your own knowledge content or your own instructions to the bot already amount to a modification within the meaning of para. 11.
• Proviso for product copy: the guidelines do not explain how far claims “related to e.g. health, consumer safety or sustainability” bring product and advertising copy within the obligation.
• Newsletter and blog: whether a newsletter sent to many subscribers is “published”, and where public interest begins for guide articles in a store, has not been decided.
• Unfair competition law: no court has decided whether a missing AI notice in a chat is a breach of unfair competition law on its own. The bodies checked cite no court ruling on Article 50(1) or the second subparagraph of Article 50(4).

In short

An AI chat or voicebot in your store has to say clearly at the start of every conversation that an AI is answering. On the wording and the guidelines, the obligation falls on the provider of the system; as a store, you make sure the notice stays visible according to RTR, and according to the Wettbewerbszentrale you are the provider yourself for a bot you have bought and integrated. The result is the same: a notice at the start, no human appearance. AI texts in a store, that is product descriptions, category and SEO copy, FAQs and advertising newsletters, are usually outside Article 50 under the guidelines; texts on health, the environment or consumer safety are within it, unless a person has reviewed them for substance and the company holds responsibility for them. Regardless of this, you are responsible under unfair competition law for every statement in your texts and your chatbot, and fake reviews are prohibited, with or without a label.