Since 2 August 2026, Article 50 of the EU AI Act has required a double marking of content generated or manipulated by AI (European Commission FAQ): a machine-readable technical trace, embedded in the metadata or as a watermark, which falls to the system provider, for all synthetic content; and a visible disclosure, perceptible to anyone looking at the image, which falls to whoever publishes content that could pass as authentic, including public-interest text published without editorial control. Neither obligation has a company-size threshold, and the fine can reach €15 million or 3% of worldwide turnover. A grace period runs until 2 December 2026, but it only covers the technical marking of systems already on the market before 2 August. My own trade has lived since 2017 under a single mandatory disclosure, for a single action. This piece puts the two texts side by side.
Two marks for a generated pixel, one line for an altered body
The contrast comes down to two texts. On the AI side, the EU regulation demands precision: interoperable and reliable technical marking for providers, clear and perceptible labelling from the first exposure for deployers when content could pass as authentic, European icons that are proposed but optional, and a code of practice published on 10 June 2026. Nothing is retroactive: content produced before 2 August does not have to be marked after the fact.
There is another text, the one that has framed my trade since 2017. Decree no. 2017-738, in force since 1 October 2017, requires the mention “Photographie retouchée” (retouched photograph) whenever a model's figure has been slimmed or filled out by software. No other requirement exists. The fine reaches €37,500, raised where applicable to 30% of the advertising spend, and it targets the advertiser, never the studio that handled the layers. Next to the technical precision now demanded for a generated pixel, the lightness of that regime, one sentence on the image with no equivalent technical requirement, stands out by contrast.
The 2017 text stops at the model's body
On my screen, the figure is one slider among many. The Liquify tool, the one that narrows a waist with a mouse movement, arrived with Photoshop 6.0, in 2000: a quarter of a century of practice. I can narrow a waist, and I can just as easily lighten skin, move a light source, add a catchlight to an eye or rebuild an entire background. The 2017 text stops at the body: the face, the complexion, the lighting and the set never fall within its scope.
Its scope of distribution is broader than it is usually summarised, however: the text covers billboards, press, direct mail and printed advertising, as well as online communication to the public; television is the one channel missing. What remains narrow is the gesture it targets: the figure, and nothing else.
Even the most cautious players in the sector read it that way. Getty Images banned any retouched figure from its image bank in direct response to the French law, in its contributor contracts from 1 October 2017, while continuing to accept retouching of complexion, hair colour or nose shape. Nine years on, that perimeter has not moved on this ground.
2023: the face enters the scope, for influencers only
An extension came in 2023, for influencers. Article 5 of the law of 9 June 2023 finally brings the face within the scope of the disclosure, with two distinct wordings: “Images retouchées” (retouched images) when image processing has slimmed, filled out or altered someone's appearance, and “Images virtuelles” (virtual images) when a face or a figure has been created by AI. The disclosure is only due in the context of commercial influence, though: a creator who alters their face outside a commercial partnership stays outside the text. The penalty changes scale: up to two years' imprisonment and a €300,000 fine, raised to €1.5 million for a legal entity. The legislator therefore knew how to toughen the sanction while widening the perimeter. It did so for one subset of visual production only, never for the whole.
The State connects the texts without drawing the conclusion
The connection is not my own construction: it appears in the Senate record. On 16 April 2026, Senator Dany Wattebled questioned the government on the rise of generative AI in advertising and fashion, citing the 2017 decree in the statement of his oral question no. 1070S. The reply given on 30 April by Anne Le Hénanff, minister delegate for artificial intelligence, relies on the 2023 influencer law as the existing framework, then refers the generalisation of transparency to the application of Article 50 at European level. The three regimes are thus linked in a single official exchange. No extension of the 2017 text comes out of it.
In September 2017, the year the decree came into force, Emily Ratajkowski publicly denounced retouching of her lips and chest on the cover of Madame Figaro, with no public reaction from the magazine at the time. The case fell outside the law, as the figure was not at issue, and nothing required a response. That is the gap I have just described, embodied on a French magazine cover, in the same year as the text meant to regulate the practice.
One could conclude that this text has never been enforced anywhere. No sanction under this regime was identified in the public sources consulted for this piece. An absence of trace is no proof of an absence of control. I prefer to stick to what I can observe directly: a perimeter that stayed narrow, and a trade that has never had to disclose most of what it does.
The objection that holds, and the ground where it falls short
A stronger argument than the previous ones can be raised against me. Article 50 does not primarily target body image. It answers a risk of large-scale disinformation, political deepfakes included, where the stakes go far beyond whether a model looks thinner on a cover. Comparing the two regimes term by term may look like a shortcut to a reader who knows the European text in detail. The objection holds. It does not excuse us from looking at the narrower ground where the two texts genuinely overlap: commercial, advertising and editorial imagery, where a machine will tomorrow generate what retouching software shaped yesterday, with a traceability requirement that the retouched body never received.
AIxH's view
I do not know what a watermark will change about the way a viewer looks at an image. I do know that it took nine years to turn a figure slider into a line of disclosure, and that the European icons remain optional while the obligation does not. Within the Content & social media service of AIxH, social media agency in Luxembourg, we treat this transparency as long-term groundwork: a human brief, AI production openly acknowledged, human validation before publication, and consistent labelling across all the visual production it accompanies. You will judge, in time, whether labelling changes a viewer's eye more than the disclosure changed mine.