What the June 10 Code of Practice Locked In
The European Commission published the final Code of Practice on marking and labelling AI-generated content on June 10, 2026. The Code is a voluntary path to complying with the AI Act's Article 50 transparency obligations, which apply on August 2, 2026, and it specifies four instruments: signed metadata, watermarking, free detection tools, and a common EU label icon. More than 187 participants shaped the drafting, and further implementation guidelines are due before August 2 — so reading the Code text and stopping there is not a plan.
Article 50's Two Tracks: Machine-Readable Marking and Labels
Article 50 splits into two obligations of different character. One embeds a machine-readable mark into AI-generated content; the other applies a human-perceptible label to deepfakes and to AI text on matters of public interest. The first lives in the file itself, carried by signed metadata or watermarking; the second surfaces on screen through the common EU icon. Because the two live at different layers, they need different implementation points.
To Claim the Exception, You Have to Record It
The Code grants a labelling exception for text produced under human review and editorial responsibility. That exception does not assert itself, though. If nothing records who reviewed which draft, when, and under whose editorial responsibility, you have no basis to prove the exception to a supervisory authority. An organization that never documents the exception's conditions has effectively surrendered it.
From CMS Fields to Review Logs: A Provenance Pipeline Checklist
Pin the target numbers first. A workable starting bar: 100% label coverage across public-interest and deepfake publications, 100% review-record retention, and zero marking-verification failures per month. Because image, text, and audio carry marks by different specs, count coverage per medium — otherwise a shortfall in one hides inside the average.
Three failure patterns recur. First, the publishing pipeline has no provenance metadata field at all, so there is nowhere to embed signed metadata. Second, the human-review exception's requirements are never written down, so even reviewed text cannot claim the exception. Third, marking diverges by medium — watermarking on images, nothing on text, some other scheme on audio — until the coverage tally means nothing.
Put the recovery branch at the verification step. If marking verification fails just before publish, block the release and route it back to the generation stage to reinject metadata; if a draft classified as an exception has an empty review record, default to forcing the label on. Only strip the label when the exception is verifiably solid — that way unverified content never leaks out unlabeled.
Harden the operations checklist into three stages. (1) Add provenance fields to the CMS so there is a home for the generating tool, model ID, reviewer, editorial owner, and label type. (2) At generation, inject signed metadata and watermarking, and auto-attach the common EU label icon to anything classed as public-interest. (3) In the review log, keep the review timestamp, the reviewer, and whether edits were applied — that is your evidence for any exception claim.
Before rollout, run one scenario per medium so you confirm that marks and labels actually land on text, images, and audio alike. Reverse-check your own output with the free detection tools to see whether the watermark reads back, and you have earned the verifiability the Code assumes.
Run the improvement loop on metrics. Tally the top marking-verification failure types each month and fix the CMS field or injection rule behind them; when the implementation guidelines arrive after August 2, update the label-icon spec and the exception-documentation template to match. Alarm on the moment label coverage drops below 100%, and you can pull back before high-volume publishing drifts out of conformance.
Takeaways at a Glance
What August 2 demands is a pipeline, not a declaration. Stand up the three stages — provenance fields in the CMS, signed metadata plus watermarking plus the EU label icon at generation, and review-log retention — and watch 100% label coverage, 100% review-record retention, and zero marking-verification failures as your metrics. Do that, and the voluntary path from a Code shaped by 187-plus participants holds even at publishing scale.
References
Code of Practice on marking and labelling of AI-generated content — EU digital-strategy