Europe’s generative-transparency clock has a red mark: 2 December 2026. That day ends the grace period for Article 50(2) AI Act machine-readable marking for generative systems already on the market before 2 August 2026. Providers who launched after that date never enjoyed the window: they had to comply from day one.
The so-called Digital Omnibus framed that four-month extension as technical relief—not political relief—for providers with hard-to-remark legacy stacks. Other Article 50 duties (end-user disclosure, deepfake identification in sensitive contexts, and so on) have been enforceable since 2 August. Confusing “watermark grace” with “all of Art.50 on pause” is an expensive compliance error.
On 17 September 2026, the European AI Board’s 9th plenary confirmed there will be no further general postponement. The signal to industry is unambiguous: the calendar stands. Labs and platforms hoping for another year-end extension must execute marking plans, not last-minute lobbying.
The associated sanction regime can reach €15 million or 3% of the preceding financial year’s worldwide annual turnover—whichever is higher. For hyperscalers that sounds like manageable risk; for thin-margin European scale-ups it can be existential. The incentive to sign codes of practice and document controls is therefore economic as well as reputational.
A second chained milestone is 2 February 2027: the target for interoperability of watermark-detection mechanisms among transparency-ecosystem signatories. Without detectors that recognize one another across providers, marking fragments into silos and loses value for journalists, platforms, and electoral authorities.
Neural Pulse separates three layers often blended in headlines: (1) end-user disclosure; (2) machine-readable marking / watermarking; (3) C2PA-style provenance on files. The Omnibus mainly eased layer (2) for pre-2-Aug cohorts; it does not erase (1) or replace good practice on (3).
U.S. providers with EU users—Anthropic, OpenAI, Google, Meta, Microsoft—share the same countdown, with different technical approaches but the same political date. The Commission and national market-surveillance authorities will have to prioritize: they cannot audit everyone on 3 December.
For deployers (firms integrating third-party models) the lesson is contractual: require Art.50 marking evidence in the MSA, detector access where it exists, and update clauses if the provider changes methods. “The model is from outside the EU” is not a defense if the output is offered on the European market.
The Board also reinforces the transparency Code of Practice as a voluntary path to show conformity. Signing is not an automatic shield against fines, but it organizes controls, audits, and shared language across ~190 organizations—the subject of a sibling article in this Neural Pulse batch.
Practical risks toward December: detector false negatives, adversarial watermark scrubbing, and multimodal content assembled with unmarked tools. Art.50 does not promise a perfect oracle; it promises a duty of reasonable, documented means.
Member states with election calendars or high media polarization will treat 2 December as an information-security date as much as a digital-market date. Expect national guidance and, likely, symbolic first investigations in 2027.
Neural Pulse closes with a minimum checklist for 2 December: inventory of in-scope generative systems; marking status by model cohort; internal detection tests; a response plan if a regulator asks “how do I know this is synthetic?”; and a named owner—not an ownerless Slack thread.
Fewer than ten weeks remain. In European regulation, that distance is a sprint, not a sabbatical semester.
