Who is on the hook, and what has to be done
Article 50 of the AI Act governs transparency, and it puts obligations on two different roles. This is the most important distinction in the whole topic.
| Paragraph | On whom | What has to be done |
|---|---|---|
| (1) | provider | An AI system that talks directly to a person must be designed so the person knows they are talking to a machine |
| (2) | provider | Synthetic audio, image, video and text output must be marked in a machine-readable format, so that the machine origin is detectable |
| (3) | deployer | With an emotion recognition or biometric categorisation system, the people concerned have to be informed |
| (4) | deployer | When publishing a deepfake, and with AI text on a matter of public interest meant for the public, it has to be disclosed that it is machine-made |
| (5) | both | The information must be clear and distinguishable, at the latest at the first interaction, and accessible |
In practice this means that machine-readable marking is not your job if you are only using the tool — it falls on the provider of the generating system. What lands on you is the visible disclosure: labelling deepfakes and public-interest text, and, if you run a chatbot under your own name, the notice.
Dates: Article 50 has applied since 2 August 2026. There is a single grace period: systems placed on the market before 2 August 2026 have until 2 December 2026 to introduce the machine-readable marking required by paragraph (2). For chatbot disclosure and deepfake labelling there is no reprieve.
The exceptions — precisely, because this is where people get it wrong
On paragraph (1): no disclosure is needed if it is obvious from the situation to a "reasonably well-informed, observant and circumspect" person.
On paragraph (2): no machine-readable marking is needed if the system only performs a standard editing assistance function, or if it does not substantially alter the input or its meaning. So a spell checker does not need marking.
On paragraph (4) there are two exceptions:
- For a work of art, creative, satirical or fictional work, the obligation is limited to disclosing in a way that does not hamper the enjoyment of the work. So you are not exempt — you may simply label more discreetly.
- For text, if human review or editorial control took place and someone bears editorial responsibility for publication, labelling is not mandatory. This exception fits most company blogs and newsletters — provided somebody really did review it and does take the responsibility.
The official EU icon set
In the summer of 2026 the Commission released three official icons, each in four variants (black, white, and both at 50% transparency), in SVG and PNG, free to download.
| Icon | When to use it | Example |
|---|---|---|
| Basic | Where AI contributed to producing the deepfake or the published text | a video captioned "the voices were synthetically produced", with the basic icon next to it |
| Fully AI-Generated | Where the whole content is machine-made, without a human creative element or editorial control beyond the prompting | a fully generated video of a fictional event; a machine-written news summary |
| Partially AI-Modified | Where existing, human-made content was partly altered by a machine | someone's face swapped in an authentic photo; an empty flat furnished by machine in a photo |
Placement has rules too: the icon must be clearly perceptible at the latest on first encounter; nothing may be layered over it; and it has to stay visible when the content is shared or downloaded. That last one is the point most often lost — a caption placed in the surrounding frame disappears the moment somebody crops the video.
And the trap that has to be said out loud: using the icons is voluntary, while the labelling obligation is not. The icon on its own does not create legal compliance — the responsibility stays with whoever publishes.
According to the Commission's own user testing, the result improved on every measured dimension once a text label was added next to the basic icon. In other words: icon plus words, not icon alone.
The code of practice: voluntary, but this is the evidentiary route
In June 2026 a code of practice on the transparency of generated content was finalised. It has two sections: one for providers on marking and detectability, the other for deployers on labelling deepfakes and machine-written text. By the end of July roughly 190 organisations had signed it.
Its legal status is precisely worded: joining is voluntary, but the labelling obligation is a legal obligation. The Commission and the board confirmed that the code is a suitable voluntary instrument for demonstrating compliance.
That is where the stakes are: anyone who complies by some other route has to prove that their measures are adequate — and that can be assessed separately in each member state by the market surveillance authorities. Whoever signs gets a single EU yardstick.
When "satire" does not protect you: a 2026 regulatory decision
This is the most important practical example in the article, and it was born barely two weeks before Article 50 started to apply.
An Italian satirical TV programme used AI to make deepfake footage of a well-known news anchor: his real likeness, in the studio from which he presents the evening news, with sentences put in his mouth by machine dubbing. The videos ran on television, on the programme's website, on social media and on the streaming platform.
The channel's defence stood on two legs: they had placed the necessary warnings, and besides, it is common knowledge that the programme is explicitly satirical — and in any case the AI Act's disclosure duty only applies from 2 August 2026.
In July 2026 the Italian data protection authority rejected both arguments and issued a warning. Three points of the reasoning are instructive:
- Satire protects you when it shows. The sentences put in his mouth, the authority said, "appear real at an immediate glance", and they do not carry enough exaggeration to make it obvious that this is a fake.
- Context makes it worse. The fact that he appears in his own news studio makes the processing "particularly insidious" from the point of view of personal identity.
- The caption was not enough. According to the authority the warning was not accompanied by a sufficiently clear disclosure — "at least measured against the technological knowledge of an average or less attentive audience" — and what would have been needed is a disclosure that also reaches the less attentive viewer and the one who joins a programme already running.
The infringement established was not under the AI Act but under the GDPR: the principles of lawfulness, fairness and transparency, plus data protection by design and by default. The authority did not fine, on the grounds that the question counts as legally novel.
The lesson in one sentence: the AI Act's satire exception only softens the AI Act's labelling duty. With a deepfake of a real person, data protection law applies separately and at full strength, and there is no "obviously satirical" exemption there.
The technical side: what works and what does not
Three solutions are in circulation, and all three have their limit.
Embedded provenance metadata. The best-known standard family of this kind attaches content credentials to the file. There are two practical limits. One is that the standard's own security documentation states it: it offers no protection whatsoever against the marking being removed entirely — an attacker can simply tear it off. From this follows an asymmetry worth remembering: an existing, valid marking says something; a missing marking says nothing at all. The other limit is the certificate requirement: a validated signature needs a certificate on the trust list, and a self-signed one does not validate as trusted — this is where a small business gets stuck.
A 2026 academic analysis goes further than that: the authors argue the specification misses both of its own security goals, timestamps can be swapped unnoticed, many verifier programs do not handle revocation correctly, the same image is valid on one tool and invalid on another — and some signed material became unverifiable within months because certificates expired.
Invisible watermarking. It survives the screenshot, but according to a 2026 measurement diffusion-based regeneration essentially destroys it: for the methods examined the detection rate fell from 99.8% to 7.4%, and with guided regeneration to zero, while image quality was preserved.
A visible caption or icon. It is not forgery-proof, but it is the only one that does what the law asks for — because the law prescribes a disclosure aimed at the viewer, not cryptographic proof.
The summary sentence: metadata dies at the screenshot, a watermark usually survives the screenshot, and a watermark dies at regeneration. The visible caption survives none of them technically — but it was never meant to.
A detailed comparison of the labelling techniques — what each can do, how long each lasts — is the subject of watermarks and provenance, with the measurements.
What to do tomorrow morning
1. Go through where machine-made content shows up in the company. Website chatbot, social media images, newsletter, video, customer support replies.
2. Decide for each channel whether you are a deployer or a provider. If the chatbot runs under your name, you are quite probably a provider — and then the disclosure is yours.
3. For text, check the editorial exception. If someone really reviews it and takes the responsibility, no labelling is needed. Write down who that someone is.
4. Where there is a deepfake, label it — with an icon and with words. Somewhere that stays visible after sharing and downloading.
5. If a real person appears in it, look at the data protection side too. The Italian case is precisely about this. And realistic intimate content produced without consent becomes an outright prohibited practice from 2 December 2026, in the highest fining band.
What to take away
Labelling has been mandatory since 2 August 2026, there is a free EU icon set for it and a voluntary code of practice, which is the simplest evidentiary route to compliance. What falls on you as a user is the visible disclosure, not the machine-readable marking.
None of the technical solutions is forgery-proof, and the standards' own documentation says so. The law, however, does not ask for cryptographic proof either — it asks that the viewer finds out. The Italian decision's yardstick is the best compass here: it has to reach the less attentive viewer too.
What applies to you today out of the whole AI Act is summarised with dates in the current state of the EU AI Act, quarter by quarter; and for the ethical and business side of labelling text, when you have to declare AI writing takes it from there.