- No retroactive labelling: content finished before 2 August 2026 does not have to be labelled after the fact — European Commission guidelines of 20 July 2026, section 8.4, paragraph 154.
- But for text, publication counts, not creation. Written before, published after = must be labelled.
- The guidelines are not binding. The regulation itself contains no transitional rule. Courts are not bound by them.
- The wave: as of 21 Aug 2026, no documented cases are traceable. The warnings started in March — from the people who bill for them.
Article 50 of the EU AI Act has applied since 2 August 2026. Publish AI-generated content, and you have to label it. Since spring, one genre has filled the timelines: "Expert warns of a wave of warning letters." "Is a wave of cease-and-desist letters looming?" "What website operators must do now."
Three weeks after the deadline, I went looking for how many of those letters are actually documented. The answer is why this article exists.
The wave everyone warns about #
Search for warning letters about AI labelling and you will find dozens of posts. Law-firm blogs, consultant pages, agency guides. Nearly all follow the same shape: name the deadline, quote the fines (up to 15 million euros or 3 per cent of global annual turnover), point at earlier waves, offer advice.
What none of them contains: an actual case. No file number, no named company, no court. Instead the comparison runs to influencer advertising disclosure — where that same pattern really did produce hundreds of proceedings. "There was a wave back then" quietly becomes "there will be one now."
I am not claiming there are no warning letters, or that there never will be. I am saying: three weeks after the deadline, not one case is publicly documented — while the warnings about them are too many to count. The ratio is the story.
Even the Wettbewerbszentrale — the body that actually sends warning letters in Germany — has so far mainly done one thing: publish a free guide to AI labelling in February 2026. People planning a wave rarely publish the instructions for avoiding it first.
What the rules actually say — paragraph 154 #
The question I get most often: do I now have to go back and label all my old AI images?
No. And the reference is specific: on 20 July 2026 the European Commission published its final guidelines on Article 50. In section 8.4 "Entry into application", paragraph 154, page 50, it states that content finally generated or manipulated before 2 August 2026 does not need to be labelled retroactively. That covers AI-generated content under Article 50(2) as well as deepfakes under Article 50(4) subparagraph 1.
European Commission, guidelines on Art. 50 AI Act, 20 July 2026 — section 8.4 "Entry into application", paragraph 154, page 50. Voluntary retroactive labelling is recommended, not required.
So much for the relief. It only holds if you read three things that most guides leave out.
The three real traps #
First: the relief is not in the law. It is in guidelines. Commission guidelines are not legally binding — they express the current interpretation and do not bind courts. The regulation itself contains no explicit transitional provision. Anyone saying "that's allowed, it's in the law" is wrong.
Second — the point almost everyone skips: for text, the date of publication counts. For image, audio and video, what matters is when the content was generated; a later first publication is harmless as long as the version was finished before. For text the guidelines draw the line differently: created and published before the deadline means clear — created before but published after means it must be labelled.
If you batch-produced blog posts with AI in July and schedule them for September, you are publishing content that must be labelled. Check your content backlog — that is where the risk sits, not in your archive.
Third: touching it makes it new. Legacy content stays legacy only while it stays untouched. Any AI edit after 2 August creates a new version, and that version falls under the obligation. The old image can stay as it is. The old image run through an AI upscaler cannot.
How to label it properly #
Good news for everyone waiting on official wording: the AI Act prescribes none. What it requires is the outcome — it must be clearly and distinguishably recognisable that content was artificially generated or manipulated. How you phrase that is your call.
Two levels have emerged in practice — and the distinction matters more than it looks:
| Content | Label | Where it belongs |
|---|---|---|
| Fully AI-generated | AI GENERATED |
Visible in the image, e.g. top-right corner — plus the caption |
| AI-edited | AI MODIFIED |
Same — the gradation is more honest than a blanket "AI" |
| Video | Notice at the start | On screen, repeated in longer videos |
| Audio | Spoken notice | At the beginning, repeated if long |
| Text | Visible notice | With the text; providers must additionally mark it machine-readably |
Two details from practice. Your label has to be understandable for your audience — in German-speaking markets "KI" is considered safer than "AI" for exactly that reason. And double labelling — visible in the image and in the caption — is the safest option. It survives someone copying the image out of your article.
One distinction that often gets muddled: providers of AI systems must mark output machine-readably — watermarks, digital signatures. Deployers, meaning you as the user, must label visibly. If your tool embeds an invisible watermark, that does not replace your visible label.
Who can actually act #
No documented cases does not mean nobody could. The route runs through competition law rather than fines: a breach of the transparency obligations can be treated as a breach of a market-conduct rule under German unfair-competition law. That opens the door for competitors, the Wettbewerbszentrale and associations, and consumer protection bodies.
For small businesses that is the realistic risk — not the EU fine. A warning letter costs four figures, arrives fast, and comes with a cease-and-desist undertaking carrying a contractual penalty. Germany's Federal Network Agency became the competent market surveillance authority when the national AI implementation act entered into force on 29 July 2026. It can request documents and demand remedies — but it has only just been stood up.
Where I stand #
I label everything. Every image in this article carries the notice; every AI contribution to my work is stated in the open. Not because I fear a warning letter that so far does not exist — but because it looks better than the alternative.
Say openly what is AI-generated and you look in control. Hide it and have to explain later, and you look caught out. — Operator note
The reflex to see a threat first in every new rule costs more than the rule does. Labelling is trust you can buy cheaply: two words under an image, and your audience knows where it stands. That is not bureaucracy, it is a free credibility move your competitors are currently leaving on the table.
What to do today, in this order: check your content backlog (the text trap), fix a standard label (two phrasings are enough), build it into the workflow so nobody forgets. That takes a morning. Anything beyond it is something somebody is selling you.
The wider picture is in the cornerstone: → EU AI Act · 2 August 2026 — what companies must do NOW.
This article is not legal advice. I am an operator, not a lawyer. It reflects the research position as of 21 August 2026, with references you can check yourself. For binding answers on your specific case, ask someone admitted to practise.
FAQ
Do old AI contents have to be labelled retroactively?
No. Under the European Commission's final guidelines of 20 July 2026 (section 8.4, paragraph 154, page 50), content finally generated or manipulated before 2 August 2026 does not have to be labelled retroactively. Important: for text, the date of publication counts, not the date of creation.
Have there been warning letters over missing AI labels?
As of 21 August 2026, no documented cases are publicly traceable. They are possible: under unfair-competition law, competitors, the Wettbewerbszentrale and consumer associations can act. Nobody has evidenced an actual wave — the warnings started in March 2026.
How exactly do I label AI content?
The AI Act prescribes no wording. It only has to be clearly and distinguishably recognisable that content was artificially generated or manipulated. In practice: AI GENERATED for fully generated content, AI MODIFIED for edited content. Visible in the image plus caption, on screen at the start of a video, spoken at the start of audio.
What happens if I re-edit an old AI asset?
A new version is created, and that version must be labelled. Legacy content stays legacy only while you leave it alone. An upscaler or retouching pass after 2 August is enough to reset it.
How high are the fines?
The range for breaches of the transparency obligations is up to 15 million euros or 3 per cent of global annual turnover. For small businesses that is unrealistic — the practical risk is a civil warning letter with four-figure costs.