On 10 June 2026 the European Commission published its final Code of Practice on the transparency of AI-generated content, seven weeks before Article 50's marking and disclosure duties come into force on 2 August. That sequence is telling. It suggests how much is still unsettled beneath a date that has never moved. Vera Schmitt and colleagues (2026) read Article 50(2)'s demand for 'dual transparency', content at once machine-readable and legible to a person, not as a box ticked at the end of a pipeline but as something that has to be designed into a system from the start. Their fact-checking and synthetic-data cases suggest today's models cannot produce it on demand. Transparency has to be built in rather than bolted on, and little on the market is built that way yet.
The clearest illustration is text. Thomas Souverain (2025) takes the four qualities Article 50 asks a mark to have, reliable, interoperable, effective, robust, and tests real language-model watermarking against each. None of the methods he examines clears all four at once. The ones that come closest are woven into the model during training, not stamped onto the output afterwards. A mark applied late is a mark that can be taken off, and that lesson reaches well past text.
Pictures make the same point from another direction. Jie Cao and colleagues (2025) survey the field and find that even a light pass through an image editor or a diffusion model strips or blurs a watermark while the picture itself survives untouched. Removal, on their account, is not an edge case but the central unsolved problem of the area. The gap is not only theoretical either. When Rijsbosch, van Dijck and Kollnig (2025) audited generators already in deployment, they found just 38% marking their output adequately and 18% labelling deepfakes. Even the imperfect tools that exist are used unevenly.
The cost of getting this wrong is real. Shuai Wu and colleagues (2026) set it out: synthetic evidence convincing enough to move markets, mislead clinicians and sway a courtroom, answered by a call for cryptographic provenance behind a visible label. That is close to what Article 50 asks for. It is also what none of the work above can yet promise. The obligation becomes binding on 2 August. The Code meant to make it workable, out seven weeks earlier, stays voluntary.