Art. 50 AI Act: labelling AI content from 2 August 2026 — a practical guide

Do a quick inventory: website chatbot, generator graphics in social, a newsletter written with GPT. If you have any one of those three — from 2 August 2026 you have at least one disclosure duty. This is the most visible AI Act rule: its absence is visible to the naked eye, on your own site.

In short
  • Art. 50 introduces four transparency duties: chatbots disclose they are AI; synthetic content is labelled; deepfakes are clearly revealed; emotion-recognition systems inform the people they concern.
  • The information duty applies from 2 Aug 2026; technical machine marking (watermarking) — 2 Dec 2026 to verify.
  • This is not a lawyer’s problem — it is a UI, metadata and publishing-process problem. It can be closed in 7 days.
Duties

What you have → what you must do: four duties in one table

What you haveWhat Art. 50 requiresWhat it looks like in practice
Customer chatbot / voicebotClear information that the interlocutor is dealing with AI — unless obvious from context.Message in the first reply or a permanent chat chrome element. Voicebot: announcement at the start of the call.
Images, video, audio from generators published externallyLabelling that content is AI-generated or manipulated — in a machine-readable format.Label at publication + file metadata. Full technical watermarking from Dec 2026.
Realistic likenesses of people, places, events (deepfake)Clear, visible disclosure that the content was artificially generated or manipulated.Visible annotation on the content itself, not only in the caption.
Emotion recognition / biometric categorisationInforming the persons the system acts upon.Rare for SMEs — but if you use e.g. conversation sentiment analysis, the duty is yours.

The fifth case everyone asks about: AI-generated text. Disclosure applies to text published to inform the public on matters of public interest — unless the content underwent human review and editorial responsibility sits with a named person or entity. For typical content marketing that means: human editing and a byline are your exception. But the graphic next to the text has no such exception.

Before / after

What it looks like in the interface

For a chatbot the whole change is often one line — here is a variant that meets the rule without killing conversion:

AI assistant · digitaboo — you are speaking with an artificial intelligence system
Bot
Hi! I’m an AI assistant. I can answer questions about our offer — if you prefer a human, type “consultant”.
Customer
How much does an audit cost?

Two design principles we use in deployments: disclose at the start of the interaction, not in the terms (the rule is about informing the person the system interacts with — a footer link does not do that) and neutral language instead of apologetic — conversion research does not forgive “warning, this is only a bot”.

Edge cases

Questions that come up in every workshop

LinkedIn post written with AI help — label it?

A company post you edited and stand behind with your name — as a rule, text disclosure does not apply (human control + editorial responsibility). But a generated graphic in the same post — must be labelled as synthetic. The most common mistake we see: firms police the copy and publish generator images with no trail.

What about content published before 2 August?

Priority goes to active touchpoints and new publications. Rational order: chatbot → publication templates → new content → only then any archive review. Do not burn resources retro-labelling everything retro scope — to verify.

Internal assistant for employees?

Art. 50 targets interactions with persons and public content. Internal tools sit under AI policy and GDPR — but hygiene is the same: staff should know they are talking to AI and what they may paste into it.

Sales email written by GPT?

Commercial correspondence is not “informing on matters of public interest” — the text duty does not cover that case. Voluntary transparency can still be a trust advantage, not a cost — we write about that regularly.

The point

This is not a lawyer’s problem

A lawyer will tell you that you must label. They will not tell you where in the UI to place disclosure without killing conversion, which metadata to stamp on a generator file, or how to rebuild the publishing process so labels happen by default, not by memory. Art. 50 sits exactly at the junction of law, UX and engineering — which is why so many firms stall between legal and marketing, neither of whom feels ownership.

Plan

Deployment in 7 days

  1. Days 1–2 — touchpoint inventory. Every place AI meets a human or an audience: chatbots, voicebots, image/video generators, content templates, mail automations.
  2. Day 3 — decisions per channel. For each point: in scope?, disclosure formula, UI placement, who implements.
  3. Days 4–5 — implement. Chatbot message, labels and metadata in the publishing flow, template updates.
  4. Day 6 — policy and people. Add labelling rules to the AI policy + 30-minute training for content and support (also a building block for Art. 4 — yesterday’s article).
  5. Day 7 — evidence. Short note: what was inventoried, what was deployed, when, who owns it. Your material when the question comes — from an authority or a client.
Transparency Fix · fixed price · 5 working days

We can run this plan for you

Review of all AI–customer touchpoints, decisions per channel, ready disclosure formulas (PL/EN), implementation guidance for the team, and an evidence memo. One review, fixed price, before 2 August.

Book a 30-minute call
Download

Compliance checklist: Art. 4 + Art. 50 before 2 August

12 control points with date and owner fields — including a per-channel labelling section.

Download checklist (PDF)
FAQ

Common questions

From when exactly does Art. 50 apply?

Information and labelling duties — from 2 August 2026. Technical machine-marking requirements (watermarking) — from 2 December 2026, after Digital Omnibus changes.

Is “AI-generated” in the post description enough?

For deepfakes — no: disclosure must be clear and visible on the content. For ordinary synthetic content, a publication label + metadata is a sensible minimum.

Who is liable — us or the tool vendor?

Duties split between the system provider (technical marking) and the deploying entity (disclosure to recipients). If you publish — you own disclosure on your side.

This article is informational and is not legal advice. Duty scopes are based on Art. 50 of Regulation 2024/1689 and the Digital Omnibus package status as of July 2026.

RW

Rafał Wiatrowski — for 25 years has built and deployed technology: from code, through CTO roles, to GenAI strategy for large organisations. Designs production AI systems (agents, RAG, MCP) and helps companies move from pilots to accountable deployments. Executive MBA (Carlson School of Management).

Jeden artykuł tygodniowo. Zero lania wody.

Subskrybuj →