AI Act Article 50: How to Disclose Bots and Label AI Content

Since 2 August 2026, every chatbot, AI receptionist, AI avatar and synthetic product video aimed at people in the EU carries a legal duty to be honest about what it is. Article 50 of the AI Act is the part of the regulation that touches the most businesses — and, unlike the high-risk rules, it was not postponed by the Digital Omnibus (Jones Walker).

The fix usually takes an afternoon: one sentence in a greeting, one label on a visual, one setting that stops your editing software from stripping metadata. This guide shows exactly what Article 50 requires, which party carries which duty, how to phrase disclosures that don’t hurt conversion, and how the EU rules compare with the US and Central Asia.

What Article 50 requires — four duties in one table

Source: AI Act, Article 50. Breaches can cost up to €15 million or 3% of worldwide annual turnover (TechPolicy.Press); for SMEs the lower amount applies.

Note the split. If you build your own bot, you are its provider and own paragraph 1. If you buy a bot from a vendor and run it on your website, the vendor should have built disclosure in — but you are the one facing customers, and the configuration (greeting text, language, channel) is usually in your hands. In practice: make sure it happens, whoever is formally responsible.

Chatbots and voice agents: disclosure that works

The “obvious” exception — and why not to rely on it

Article 50(1) does not apply where it is obvious, to a reasonably well-informed and observant person, that they are talking to AI. A widget labelled “AI assistant” with a robot icon may qualify. A friendly voice called “Anna” answering your clinic’s phone does not. Since the exception is judged from the user’s perspective and invites argument, the safer standard is: disclose always, briefly, at the start.

How strong the illusion can be: a Russian voice-AI vendor reports that around 90% of people do not realise they are speaking with a robot (Neuro.net). That is precisely what the law is aimed at.

Scripts you can copy

Website chat (EN):

Hi, I’m the virtual assistant of Brightway Dental — an AI. I can answer questions about treatments, prices and opening hours, and book appointments. Want to talk to a person? Just type “human”.

Voice agent (EN):

Hello, you’re speaking with the AI assistant of Brightway Dental. This call is recorded to handle your request. How can I help?

Voice agent (DE):

Guten Tag, Sie sprechen mit dem KI-Assistenten der Zahnarztpraxis Brightway. Das Gespräch wird zur Bearbeitung Ihres Anliegens aufgezeichnet.

Other channels: messengers, email, SMS and avatars

The rule follows the interaction, not the channel. A WhatsApp or Telegram bot discloses in its first message, just like a website widget. An AI agent that answers support emails on its own should say so in the signature or the first line (“This reply was written by our AI assistant; a colleague reviews escalations”). An AI avatar in a video call or an interactive product demo needs the same one-line notice at the start. Two grey zones are worth handling explicitly:

• Human-sent, AI-drafted messages. If an employee reviews and sends an email drafted with AI, the person is interacting with your employee, not with an AI system. No Article 50(1) notice is required, although your AI transparency page can mention the practice.
• Handover between bot and human. When a person takes over a conversation, say so (“You’re now chatting with Maria from our team”). It is not required by Article 50, but it keeps the disclosure honest in both directions and avoids customers assuming they are still talking to a machine.

Three details matter more than wording. First, disclose in the user’s language — a German caller should not get an English notice. Second, combine the AI disclosure with the recording notice: in Germany, recording a conversation without consent is a criminal offence under §201 StGB, punishable by up to three years. Third, offer a way to a human. Article 50 does not require it, but it defuses most complaints and supports GDPR rights around automated decisions.

How the EU compares with other markets

Sources: Retell TCPA playbook, FCC 24-17, GC AI on state laws, Kursiv. If your agents call US numbers, consent is the bigger issue than disclosure; see legal AI phone calls: TCPA, AI Act and GDPR.

Synthetic media: labelling images, video and audio

What counts as a deepfake

The AI Act defines a deepfake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. A photorealistic “customer” in a UGC-style ad, a cloned voice of your founder, a product shown in a kitchen that never existed — all qualify. A clearly stylised illustration usually does not.

How to label

• Visible, at first exposure. A small “AI” / “KI” / “IA” mark on the image or at the start of the video (TechPolicy.Press). A line in the end credits is not enough (Orrick).
• Advertising gets no shortcut. Evidently artistic, creative, satirical or fictional works may disclose in a way that doesn’t spoil the work. The Commission’s final guidelines of 20 July 2026 indicate that advertising usually does not fall under this lighter regime (Faegre Drinker).
• Consent for real people. Labelling does not replace permission to use someone’s face or voice. Get a written release for every digital twin, with a defined scope and term — for example, 24 months of use for named campaigns. In the US, likeness rights are a patchwork: the federal NO FAKES Act cleared a Senate committee in June 2026 but has not become law, while state laws such as Tennessee’s ELVIS Act (in force since 1 July 2024) already protect voice and likeness (Holland & Knight, GovTrack).

Machine-readable marks: the provider’s job, your workflow

Under Article 50(2), the tools you use must mark outputs. The transparency Code of Practice expects at least two layers — typically signed metadata plus an invisible watermark — and providers of systems released before August 2026 have until 2 December 2026 to comply (Reed Smith, CSA). Many tools already do this: Google’s Veo and Nano Banana Pro embed SynthID in every output, Adobe Firefly attaches Content Credentials.

Your part is not to destroy those marks. Export settings in editing and compression tools often strip metadata by default. Add a check to your production checklist: C2PA metadata present in the delivered file.

Platforms label for you — whether you like it or not

• YouTube requires disclosure of realistic synthetic content and labels videos automatically using C2PA metadata and detectors; systematic non-disclosure can lead to removal and demonetisation (YouTube Help).
• Meta has detected AI in ads automatically since 1 June 2026 and shows an “AI info” label in “About this ad” (Meta).
• TikTok requires labels for realistic AI content; an automatically applied C2PA label cannot be removed, and ads need an AIGC label or disclaimer (Common Thread).

So the real choice is not “label or not” but “label deliberately and well, or get labelled by the platform”. For the creative side, see AI UGC ads: costs and disclosure rules; for copyright and likeness, AI-generated content: copyright and labelling.

AI-generated text

Article 50(4) also covers text — but narrowly. AI-generated or manipulated text published to inform the public on matters of public interest must be disclosed, unless it has undergone human review or editorial control and a person or organisation holds editorial responsibility. A company blog written with AI assistance and edited by a responsible human typically needs no mandatory label. Publishing an AI transparency statement that explains how you use AI is still a good practice — it is cheap and builds trust.

Emotion recognition and biometric categorisation

If you deploy a system that recognises emotions or categorises people biometrically, Article 50(3) requires you to inform the people exposed to it. But check Article 5 first: emotion recognition in the workplace and in education has been prohibited since 2 February 2025. A call-centre tool that infers agents’ moods is not a transparency question; it is banned.

Implementation checklist

Typical mistakes

• Hiding the disclosure in terms and conditions. It must be clear and given at the first interaction.
• English-only notices for German, Dutch or French users.
• Human-sounding personas without disclosure — “Anna from reception” is fine only if Anna says she is an AI assistant.
• Stripping metadata in post-production, then wondering why platforms flag content inconsistently.
• Labelling only in the end credits of a video.
• Assuming the Omnibus delayed everything. Transparency has applied since 2 August 2026.

Does disclosure hurt results?

The evidence suggests that hiding AI is the bigger risk. In a study by Ipsos and Syracuse University (May 2026, 3,000 viewers), only 13% of viewers could reliably identify AI ads — yet human-made ads delivered 14% stronger short-term sales (McGraw Hill). McDonald’s Netherlands withdrew an AI Christmas commercial within days after a public backlash (Today), and some brands now advertise “No AI” as a trust signal (DesignRush). The lesson: disclose plainly, and make sure what you disclose is good enough to stand on its own.

Conclusion

Article 50 is the most practical part of the AI Act: tell people when they are dealing with a machine, label realistic synthetic media where they first see it, keep machine-readable marks intact, and don’t run emotion recognition on staff. Done well, disclosure is one sentence and one label — and it protects you from fines, platform penalties and the reputational cost of being “found out”. The broader timeline is in our EU AI Act guide for SMBs.

Lotus Libre builds voice agents and AI content production with disclosure, consent records and C2PA marking built into the pipeline. Want us to check your bots, scripts and ad workflow against Article 50? Book a free 15-minute audit.

This article is general information and not legal advice. For decisions about your specific situation, consult a qualified lawyer.