An ordinary street scene, reworked by a model into a snowy night. Snow on the road, breath fogging from passers-by, the reflections in the shop windows repainted. The image shows no seam. Once it is posted, few ask where it came from, no one knows the snow is fake, and no one wonders whether that streetlamp was ever lit.
Since August this year, such things have clear rules in Europe. The rules do not sit in a platform’s back office. They are written into a law: the EU Artificial Intelligence Act, Regulation (EU) 2024/1689, adopted in 2024 and applied in stages, with the transparency provisions near the end of the schedule.
Article 50 Takes Effect
Article 50 is about transparency. It does not judge whether the content is good or the image beautiful. It asks one thing: let people know a machine made this, rather than a person shooting it or writing it. That is a right to know, not an aesthetic standard.
The article applies from 2 August 2026. In a notice at the end of July, the European Commission said that from that day the AI Office, together with national authorities in the member states, would begin enforcing the relevant rules, transparency included.
Obligations are split by role. Those providing a generative system must ensure the output can be recognised by machines. Those using a system to generate or modify content must say so publicly. Two kinds of actor, two sets of actions, and neither substitutes for the other.
The Machine Has to Read That It Is Generated
Paragraph 2 is specific: audio, images, video and text generated by a system must carry a machine-readable mark in the output and must be detectable. This is not a suggestion. It is the core of the provision.
The mark is not a small label in the corner of the frame. It is information written into the file structure. It has to be effective, interoperable, robust and reliable, and it has to survive editing, transcoding and redistribution, rather than falling apart at the first touch.
For ordinary users, this means a downloaded image may carry the sentence “I am generated” inside the file itself. You cannot see it, but software can read it, and so can platforms, which is what makes tracing possible after the fact.
Text is harder. A line of a few dozen characters has little room to hide a mark, and changing a few words can scramble it. Images and video have far more pixels to work with, and they hold up better under repeated detection.
Paragraph 2 also leaves a gap. Systems that only assist with editing, without substantially altering the input or its meaning, fall outside the marking obligation. That exemption is there for retouching, colour grading and cut-out tools.
Old Systems Still Have a Transition
What about systems already on the market? The European Commission’s position is that systems placed on the market before 2 August 2026 may have a transitional period, until 2 December 2026 at the latest, to come into compliance, rather than being rebuilt in August.
That grace is not unconditional. It hangs on an omnibus legislative package still moving through procedure, and the Commission’s own text says “if adopted”. Until that legislation completes, the date is a target, not a certainty.
So anyone selling into Europe should watch two dates, not one: 2 August, and the deadline before the year ends. The first governs new systems, the second governs old ones, and the two batches travel on separate tracks.
One more point is easily missed: content generated and already made available before 2 August does not need to be labelled retroactively. The rules cover outputs after that date, not old accounts, so existing catalogues do not all have to be redone.
A Voluntary Code, Not a Law
On 10 June, the European Commission published the Code of Practice on Transparency of AI-Generated Content. The name is long, so the nature needs saying: it is signed voluntarily. It is not law, and signing or not signing does not by itself make anyone unlawful.
The code has two parts. One addresses providers, covering the machine-readable marking in paragraph 2. The other addresses deployers, covering the visible disclosure in paragraph 4. Both require the information at the user’s first contact, not buried in settings.
It does not prescribe a technical route, only a direction: layered marking. Signed metadata in one layer, an imperceptible watermark in another, each covering for the other, so that when one is lost a verifier can still keep checking.
Signing has a practical benefit. According to the Commission, adhering to a code deemed adequate can serve as one way of demonstrating compliance. Not signing is also possible, but then you have to show compliance another way, and the burden of proof stays with you.
At the end of July, the Commission published the first list of signatories, with more than 180 organisations on it. The list runs from large companies to intermediaries that build tools and distribute content, broader than expected, and not just a handful of model makers.
Watermarks Fear Screenshots and Compression
The technical difficulty is obvious to anyone who edits images. A watermark lives in pixels or in file structure. One screenshot, one heavy compression, one format conversion can wash it out, and once it is gone it is hard to restore and impossible for a verifier to read.
Metadata is more fragile still. Most social platforms rewrite files on upload and strip surplus fields, so the signature disappears. The image you see is still there; its ID card is gone, and the chain breaks at that step.
Hence the layered approach the industry pushes: one layer in the file, one in the pixels, and content credentials tying source, edit history and generating tool into a single chain that a verifier can read in one pass, without guessing layer by layer.
Those credentials are built on a standard called C2PA, promoted by a group of companies and organisations. In May this year, OpenAI announced that its image products had become C2PA-compliant and added Google’s SynthID watermark at the same time: two backup routes, one through the file, one through the pixels, neither replacing the other.
Chinese Teams Have to Comply Too
For Chinese creative teams, this is not distant news. If a work is going into the European market, the rules travel with the work. Being made in Hangzhou or Chengdu does not exempt it, and contract terms change accordingly.
A poster, an ad film, a character image for a game: any of these, once placed or listed locally, can fall inside Article 50. When the client signs, the question gets asked: how do you plan to handle the labelling?
The more practical problem is cost. Detection needs tools, records need systems, and process change needs people. For a large company that is one new role. For a team of five, it can be a full day of work.
What technology can do is write “this is generated” into the file so machines can read it. What it cannot do is make people believe. Labelling settles knowledge. Trust still has to be earned by the content itself, a little at a time, and it cannot be hurried.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.