Marking AI Content: What December 2 Actually Requires: Not a New Transparency Rule. The Last Day of a Grace Period, and the Field Printed the Wrong Year

This article is for educational purposes and does not constitute legal advice. Organizations subject to the EU AI Act should consult qualified counsel about their specific obligations.

December 2 is not a new marking rule. It's the last day of a grace period, and a lot of commentary has printed the wrong year for it.

The Date the Field Keeps Getting Wrong

Start with the error, because it shows how thin the reading has been. Much secondary commentary reports the Article 50(2) marking deadline as December 2, 2027. It's December 2, 2026. The mistake is explicable: the Omnibus moved other obligations to that later date, one year apart on the same day. EveryAILaw caught it while verifying the grace period against the Official Journal text, and recorded the correction on August 2, 2026.

If your compliance calendar says 2027 for content marking, it's carrying a year you don't have.

What Actually Falls Due

Article 50's transparency duties have applied since August 2, 2026, and were never deferred: Article 50 sits in Chapter IV, and the Omnibus deferral that pushed the high-risk regime to 2027 and 2028 reached Chapter III.

December 2 closes something narrower. Regulation (EU) 2026/1744 inserted a new Article 111(4), giving generative systems already on the market before August 2, 2026 a four-month transitional period to meet the Article 50(2) marking requirement.

So the question isn't "do we mark." It's "was this system on the market before August the second," and if it was, your grace period ends on December 2. Article 50 carries its own penalty band: up to EUR 15 million or, for an undertaking, 3% of worldwide annual turnover, whichever is higher, under Article 99(4)(g), with lower ceilings for SMEs and small mid-caps.

Machine-Readable Is the Load-Bearing Phrase

Article 50(2) requires providers of systems generating synthetic audio, image, video or text, general-purpose systems included, to mark outputs in a machine-readable format, detectable as artificially generated or manipulated. The technical solutions must be effective, interoperable, robust and reliable.

Machine-readable rules out a visible caption as the sole mechanism. Detectable means something downstream has to find the mark and conclude the content is synthetic. And robust is not decorative: a watermark nobody can read, a manifest nothing checks, metadata a re-encode strips, each satisfies "we marked it" and fails a duty written about detectability rather than marking.

The exceptions are narrow: Article 50(2) excepts limited assistive editing or insubstantial changes. "We only used it to tidy the draft" is a real carve-out; "we only used it a bit" is not.

So the failure mode is a pipeline problem, not a model problem. A provider marks correctly at generation, the content crosses an export, a CMS, a platform, and arrives stripped.

The Code of Practice Is Not a Shield

This is where the reading goes wrong in a board deck. The Commission has issued final Article 50 guidance and facilitated a Code of Practice on transparency of AI-generated content. The Code is usually called voluntary, inviting a conclusion it doesn't support: that signing it discharges the obligation.

Look at what the Omnibus did to Article 50(7). As replaced, the Commission facilitates codes of practice for detection, marking and labelling, assesses their adequacy for Article 50(2) and (4), and may impose common rules by implementing act if a code is inadequate.

That's a structure, not a menu. The Code isn't an alternative to the duty, it's a route toward satisfying one, and the Commission holds a backstop if that route stops leading anywhere. Adherence is evidence of good faith, not discharge.

Four Questions Before December 2

None need a lawyer. All need an owner.

Which of our generative systems were on the market before August 2, 2026? This decides whether December 2 means anything to you.
Does the provider mark, and how? Get the mechanism named. "We're compliant" is not a mechanism.
Does the mark survive our own pipeline? Test end to end, through every export and platform the content crosses. Almost nobody runs this one, and it fails most.
Who owns the result when detection fails? If nobody does, that's the finding.

The Part Marking Cannot Do

Marking is a machine control on a machine boundary, and there it works. The EU is right to require it.

What it doesn't do is tell anyone whether the person who received that content noticed the mark, checked it, or waved it through under deadline. A marked document and an unmarked one land on the same desk, and the difference only matters if someone reads it.

That gap is the People vector, and it's what PAICE (People + AI Collaboration Effectiveness) measures: whether a professional verifies the claim carrying the decision, escalates when a source can't be confirmed, and corrects rather than defends. Marking is necessary and not sufficient, and a behavioral assessment is no substitute for a provenance mechanism. They answer different questions.

December 2 asks whether your synthetic content is detectable. Fair question, real deadline, smaller half of the problem.

Want to know whether your people read the marks they're given? Learn about PAICE for organizations or take an assessment to see it firsthand.

Marking AI Content: What December 2 Actually Requires