Nobody Ever Asked CEN-CENELEC for a Watermarking Standard

The trade press has spent the summer reporting that Article 50 watermarking compliance is stalled behind a harmonised technical standard that is running late. The statute shows the standard was never requested, and the Digital Omnibus has meanwhile bound every provider of a pre-August generative system to Article 50(2) marking by 2 December 2026. None is coming. Regulation (EU) 2026/1744, in force since 27 July 2026, inserts a new Article 111(4) into the AI Act, and it reads, in full, "4. Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026."

That obligation attaches to everyone. Signing the Commission's voluntary Code changes nothing about it.

The presumption of conformity stops before Chapter IV

Article 40(1) is the machinery that lets a harmonised standard stand in for the law, and its presumption covers "the requirements set out in Section 2 of this Chapter", meaning Chapter III on high-risk systems, plus certain Chapter V obligations. Article 50 sits in Chapter IV. The machinery does not reach it, and the Commission's own standardisation requests confirm the design. M/593 of 22 May 2023 asked CEN-CENELEC for ten deliverables, of which the nearest is "transparency and information provisions for users of AI systems", and the mandate's Annex maps that item to Article 13 deployer information, not to Article 50. M/613 of 23 June 2025 moved the deadline to 28 February 2027 without touching scope.

The work programme agrees. I pulled JTC 21's working-group list from jtc21.eu on 19 August 2026. WG2 under Emilia Tantar carries conformity assessment, quality management and risk management. WG3 under James Davenport carries dataset governance, bias management, evaluation and logging. Searching the page for watermark, content marking, synthetic content, deepfake or Article 50 returns zero hits. There is no publicly listed work item.

What the AI Act provided for instead, in Article 50(7), was a code, and that instrument arrived on schedule. The Code of Practice on Transparency of AI-generated Content went final on 10 June 2026, ahead of the 2 August application date, and its Section 1 is implementable engineering rather than aspiration. Signatories apply two layers of machine-readable marking where no single technique is good enough alone. Measure 1.1 requires digitally signed, time-stamped metadata, which C2PA satisfies today, with imperceptible watermarking or fingerprinting plus logging as the second layer. Two carve-outs exist, for generative systems in closed physical products and for free-form text under 200 tokens. Henna Virkkunen, Executive Vice-President for Tech Sovereignty, Security and Democracy, said on 10 June, "Europeans have a right to know whether what they see, hear or read has been made or altered by AI, especially when such content can shape public debate."

2 December is statute. 2 February 2027 is a promise between signatories.

The string "2 February 2027" appears exactly zero times in the Official Journal text of Regulation (EU) 2026/1744. That date is the Code's interoperability commitment, satisfiable through a public industry standard for routing detection queries or a publicly readable signpost mechanism, and it obliges signatories only. Treating it as a regulatory deadline gets the power relationship backwards.

Tech Jacks Solutions wrote on 12 May that "Absent a harmonized standard, providers are building against the regulation's text and whatever technical guidance the EU AI Office has issued, not against a validated benchmark." True, and Clayton Lewis at TechTimes made the same observation on 21 July. Both framed the absence as delay. The mandate texts show it is design. A late standard ends a wait. A standard nobody ordered ends nothing, and an unsigned provider still waiting will reach 2 December with no presumption of conformity by any route and no documentation.

So the decision for a compliance or engineering lead at a non-signatory is concrete. Stop tracking JTC 21 for this obligation. Read Article 50(2) itself and the Code's Section 1, the only public reference points available, and choose before 2 December between signing and writing down your own two-layer marking rationale against the bare statutory text.

Some caveats. The official standards.cencenelec.eu register would not open, so the negative rests on jtc21.eu and four corroborating sources. None of the ten major law-firm briefings I reviewed raises the missing-mandate point, and no Commission statement connects the Code to the absent standard. That connection stands on the statutes alone.

One more reason to work from primary text. The Commission's AI Act Service Desk still carries a live disclaimer on its Article 99 page, "This provision has been amended by the Digital Omnibus on AI. The text displayed on this page has not yet been updated to reflect those amendments." The fine levels are unchanged, but the explainer lags the law it explains. Verify against the Official Journal, not the tooling built on top of it.

Verification note: prior coverage (Tech Jacks Solutions, TechTimes) treats the watermarking standard as delayed; this piece shows it was never mandated. The binding 2 December 2026 date is the new Article 111(4), inserted by Regulation (EU) 2026/1744; the 2 February 2027 interoperability date appears nowhere in the regulation and binds Code signatories only. Five-minute check: open the OJ text of Regulation (EU) 2026/1744, search for "2 February 2027" (zero occurrences), then read Article 1 point (39)(b). Load-bearing claims were verified against the OJ text, the M/593 and M/613 mandate PDFs, and jtc21.eu on 19 August 2026.

Sources: Regulation (EU) 2026/1744. AI Act, Regulation (EU) 2024/1689, Articles 40 and 50. Code of Practice on Transparency of AI-generated Content. JTC 21 work programme. Commission standardisation FAQ. Press release IP/26/1328.