Regulation (EU) 2024/1689 · Article 50 · Transparency

The delay you read about didn't cover this.

Headlines said the EU pushed back the AI Act. Two obligations moved to 2027 and 2028. The transparency rules did not move. They apply to anyone publishing AI-made content into the EU, and they start on 2 August 2026.

until Article 50 applies
2 August 2026
ObligationAppliesStatus
Transparency & disclosure Art. 50 · chatbots, deepfakes, published AI text
2 Aug 26
Live in
Machine-readable marking Art. 111(4) · legacy systems on market before 2 Aug 26
2 Dec 26
Next
High-risk, stand-alone Annex III · hiring, credit, education, essential services
2 Dec 27
Deferred
High-risk, embedded in products Annex I · medical devices, machinery, toys
2 Aug 28
Deferred

The Digital Omnibus on AI was adopted by Parliament on 16 June 2026 and Council on 29 June 2026. It deferred the two bottom rows by 16 and 12 months. It left the top row exactly where it was.

Six questions · about 90 seconds

Find out which duties land on you.

Article 50 is five separate obligations, and most teams trigger two or three of them without knowing. Answer honestly. Nothing is stored, nothing is sent until you choose to send it.

Question 1 of 6 · Reach

Do people in the EU see what your AI produces?

The Act follows the output, not the office. A US or UK team publishing to an EU audience is in scope.

Question 2 of 6 · Your role

Did you build the AI, or are you using someone else's?

This is the fork that decides everything. Building, fine-tuning or putting your own name on a system makes you a provider with heavier duties. Using ChatGPT, Claude or Midjourney makes you a deployer.

Question 3 of 6 · Conversation

Does an AI talk to your customers directly?

Support widget, DM autoresponder, sales agent, voice bot, booking assistant.

Question 4 of 6 · Synthetic media

Do you publish AI images, video or audio that could pass for real?

Product shots, AI models or presenters, voice clones, face swaps, retouching that changes what happened. Obviously fantastical work sits outside this. Photoreal does not.

Question 5 of 6 · Published text

Do you publish AI-written text on matters of public interest?

News, politics, health, money, safety, the environment. Marketing copy about your own product generally isn't this. A health brand's advice blog usually is.

Question 6 of 6 · Inference on people

Do you read emotion or sort people by biometrics?

Sentiment from faces or voices in calls, ad testing that reads expressions, categorising people from biometric data. Rare, high consequence.

Your result

Get the labelling pack before Saturday.

The plain-English version of what you just read, plus the wording to actually put on things. Built by an operator who ships AI-assisted content daily, not by a law firm selling you a retainer.

That didn't go through. Nothing was saved, so nothing is lost. Try once more, and if it fails again the wording is published below anyway, no email needed.

The pack appears on this page the moment you submit. The email is a copy, plus the December update when it lands. One click to leave.

Here it is.

Copy what you need right now. The same pack is in your inbox as a copy, and the December update follows when it lands.

Chat opener · Art. 50(1)

You're speaking with an AI assistant, not a person. Ask for a human at any point and I'll hand you over.

Image or post caption · Art. 50(4)

Made with AI. This image is generated, not photographed.

Video overlay · Art. 50(4)

AI-generated

On the asset, in the first frames. A caption alone does not satisfy "visible at first exposure".

Byline for AI-assisted public-interest text · Art. 50(4)

Drafted with AI assistance. Reviewed and published by [name], who is responsible for its accuracy.

Naming a person is what carries the exemption. "Reviewed by the team" does not.

The test that decides whether you're done

Could someone tell it's AI without clicking, hovering, or reading small print? If not, it isn't done. That is Art. 50(5) in one sentence.