EU AI Act · for businesses that use AI

Article 50, in plain terms.

Three rules about telling people when something is AI. Which ones apply to you depends on what you build and what you publish. Here is each one, what it asks, and what we do about it.

Article 50 applies from 2 August 2026.

In July 2026 the EU delayed large parts of the AI Act. The rules for high-risk AI moved to December 2027 and August 2028. The Article 50 transparency rules were not delayed. If you have read that the AI Act was postponed, this part was not.

Reg (EU) 2026/1744 · in force 27 July 2026 · amends Art 113 of Reg (EU) 2024/1689

The three rules

Read the “who this is” line first. Most webshops are hit by 50(1) and 50(4). 50(2) applies if you build the AI tool itself.

ARTICLE 50(1)

Say it’s a bot

Who this is: anyone running a chat assistant customers talk to.

What you must do
Tell people they are talking to an AI, unless it would be obvious to a reasonable person anyway.
Where we help
We check your live site for a chat widget, look for a disclosure in English, Dutch and German, and record what we found. Works on any website — you do not need Shopify.
What’s still open
Nothing much. This one is clear.
SETTLED

Applies from 2 Aug 2026

ARTICLE 50(4)

Label AI content you publish

Who this is: anyone publishing AI-generated or AI-edited images, video, audio or text.

What you must do
Disclose that the content was artificially generated or manipulated.
Where we help
You declare which products or content are AI once, at source. We apply the visible label, check it actually renders to a visitor, and write every decision to a tamper-evident register with a one-click evidence pack.
What’s still open
Whether ordinary AI product photography is caught at all. The rule targets content that would falsely appear authentic, and where the line sits for a retouched product shot is not settled.
THRESHOLD UNSETTLED

Applies from 2 Aug 2026

ARTICLE 50(2)

Mark what your tool makes

Who this is: you build the AI tool that generates the content.

What you must do
Mark your outputs so a machine can detect they were AI-generated.
Where we help
We record which of your content already carries a provider mark and which does not, so you can see the gap. We do not treat a missing mark as a failure before the December date.
What’s still open
Marks often do not survive publishing. We tested this: uploading to Shopify strips the embedded C2PA credential entirely.
SETTLED

2 Aug 2026 · tools already on sale before that date have until 2 Dec 2026

What we don’t claim

  • We do not detect AI. We know content is AI in three ways only: you tell us, the file carries a provider mark, or your generation tool tags it. Where none of those apply, we record the origin as unknown — never as “not AI”.
  • Following a code of practice is not proof. The EU confirmed in July 2026 that the Article 50 codes of practice carry no presumption of conformity. What you can show is the record of what you decided and when.
  • A register is not a legal opinion. We record decisions and the evidence for them. Whether a given obligation applies to your business is a question for your own counsel.
  • When we cannot check something, we say so. A scan that was blocked is recorded as blocked, not as a clean result.

Find out where you stand

We will scan your storefront for AI chat disclosure and undeclared content, and send you what we found — including anything we could not check.

Request an exposure scan See a live register