Executive briefing · week of 23 August 2026. The marking half of Article 50 shipped on 2 August. The verification half — the one that decides whether any of it survives an audit — has roughly a hundred days left.
On 2 August, the European Union's AI Act reached a date that had been circled for a year: the transparency obligations in Article 50 became applicable. Chatbots have to tell you they are chatbots. Synthetic media has to be marked. Deepfakes on matters of public interest have to be labelled. A great deal of compliance work closed out that week, and a great deal of it was real.
But the date most enterprises should have circled is not behind us. It is 2 December.
Under the transition arranged through the AI Omnibus, generative systems already on the market before 2 August have until 2 December 2026 to implement the machine-readable marking obligation in Article 50(2). That is the harder half. A disclosure is a sentence in an interface. A mark that a downstream party can actually detect and verify is a system — one that has to survive the ordinary violence the internet does to files: re-encoding, compression, cropping, screenshots.
Four answers to the same question
This is where the week's second story matters. Roughly 190 organisations — Anthropic, Google, Meta, Microsoft, OpenAI among them — signed the Commission's Code of Practice on Transparency of AI-Generated Content, which the Commission and the AI Board confirmed as adequate, earning signatories a presumption of compliance. That is a genuine achievement on the provider side. But those signatories did not converge on a mechanism. Anthropic embeds a digital identifier in Claude's EU text. OpenAI marks ChatGPT text with Google's SynthID; Gemini uses the same. Meta detects and labels other companies' images with invisible file markers. Adobe, Firefly, DALL·E, Sora and Imagen embed C2PA manifests.
Four answers to the same question. Now put yourself in the seat of a compliance officer at a company that deploys models from three of those vendors. You have inherited three or four disclosure formats and no common way to confirm, months later, that a specific asset was ever marked at all. The Code fixed the provider's presumption of compliance. It did not fix the deployer's verification problem.
The scepticism, read carefully
Which brings us to the week's third story — the one that broke into the mainstream when Forbes asked, on 16 August, whether AI watermarks "stick around, or are just for show." The scepticism has evidence behind it. Published University of Maryland testing found adversarial edits removed invisible image watermarks from around 93% of images with no visible quality loss. Text watermarks weaken under paraphrase and translation. And C2PA metadata is trivially stripped: any tool that re-encodes a file removes the manifest, platforms do it automatically on upload, and a screenshot bypasses it entirely.
Read quickly, that looks like an obituary for the whole Article 50 marking apparatus. Read carefully, it is something else: a description of a verification problem, and an argument for how to build correctly.
Marking is not DRM
Here is the reframe worth holding onto. The critique measures watermarks against digital rights management — as if the goal were an unremovable seal. Article 50 never asked for that. It asks that AI-generated content be detectable and that the disclosure be traceable. The object is an auditable record — a paper trail for accountability — not copy protection.
Grade a watermark as DRM and it fails. Grade it as one layer of an audit trail and the design becomes obvious: a file-level mark is fragile but informative, so you never rely on it alone. You pair it with a record that lives independently of the file — timestamped, tamper-evident, and verifiable on demand — so that when the mark is stripped, the claim "this was generated by X, on this date" still rests on something.
That is the lane the market has left open. Almost all of the published commentary is about marking — how a provider stamps its output. Almost none is about verification — how anyone downstream confirms a mark after the file has been through the internet. The fragility everyone is now worried about is precisely a verification problem, and verification is the half of Article 50 that decides whether any of the marking holds up in an audit.
A testable definition of done
For teams working the 2 December deadline, that turns a vague obligation into a testable one. "Article 50 done" has a definition of done: given any asset your system generated, a third party can verify — after a re-encode, a crop, a screenshot — that it was AI-generated, by you, on a date. If a routine resize breaks that, you have shipped a label, not a system.
AIACT50's reading
There are roughly a hundred days to build the second thing. The organisations that treat marking and verification as one architecture, rather than marking as a checkbox, are the ones that will still be compliant when someone actually checks.
Sources: European Commission — Article 50 Guidelines and Transparency Code of Practice · Cooley and Faegre Drinker on the 2 December transition · Forbes (16 Aug 2026) on the watermark debate · CBC and eyesift on per-vendor mechanisms · resemble.ai and sesamedisk (2026) on watermark and C2PA fragility. Public sources only — no Market Monitor input available this week.
If you want to see how a file-level mark and an independent, verifiable record are combined into one implementation — the verification half, not just the marking:
Explore AI Act 50 →