The countdown to the transparency obligations for AI-generated content has entered its decisive phase. Article 50(2), (4) and (5) of the AI Act becomes applicable on 2 August 2026, and the code of practice meant to facilitate compliance has received its adequacy assessment: on 8 July through the Commission’s opinion, and the following day through the AI Board’s assessment. In a communication of 16 July 2026, circulated to those who took part in the drafting process, the AI Office has announced that the deadline for inclusion in the first publication of signatories has been extended to 27 July 2026, that the initial list is expected on 31 July, and that the Commission’s guidelines on the scope and practical application of the obligations are due at the beginning of next week. In the meantime, the Commission has put in writing a warning that matters more than any deadline: adherence to the code does not constitute conclusive evidence of compliance.
The dates, and a divergence worth noting
The Commission’s frequently asked questions on signing the code, last updated on 10 June 2026, still give 22 July 2026 at 18:00 CEST as the deadline for inclusion in the list of initial signatories. The extension to 27 July and the 31 July publication date come instead from the AI Office’s communication of 16 July 2026 to those who took part in drafting the code and, as of today, do not yet appear on the Commission’s public pages, which place the publication of the list generically in July 2026. In practice, however, the divergence resolves itself: whoever submits the form by 22 July makes the first list on either reading, whereas anyone relying on 27 July is relying on a deadline the Commission has not yet published. The prudent course is to work to the earlier date and to watch the official pages for updates, because what binds is the date the institution publishes, not the one circulating by email. Signing remains possible afterwards in any event, on an ongoing basis: what is lost by missing the deadline is not the ability to adhere, but presence on the first list.
What is signed, and in which parts
The code is structured in two sections, mirroring the distinction Article 50 draws between actors. The first concerns providers of generative AI systems subject to paragraph 2, and covers the marking of generated or manipulated content in a machine-readable format together with the related detection mechanisms. The second concerns deployers subject to paragraph 4, and covers the disclosure and labelling of deepfakes and of certain generated or manipulated text published for the purpose of informing the public on matters of public interest. The two sections may be signed separately, according to the role actually performed, whereas individual commitments cannot be picked apart: each section is accepted as a whole. An organisation acting both as provider and as deployer is invited to adhere to both. Providers of generative AI models and technology providers of marking and detection solutions may also sign section 1 alone, even though paragraph 2 does not bind them directly. The form must be signed by a senior executive with sufficient authority to bind the organisation, and the signature may be withdrawn by letter from an executive of equivalent seniority.
What signing produces, and what it does not
The benefit of signing is real but must be described precisely. For signatories, supervision will focus on monitoring adherence to the code, with greater predictability and reduced administrative burden, applied uniformly across the Union regardless of the place of establishment and of the competent market surveillance authority. That is an advantage of certainty and method, not a safe conduct. The page accompanying the Commission’s opinion says so without hedging: adherence to the code does not constitute conclusive evidence of compliance with the obligations. The code facilitates the demonstration of compliance and makes it uniform, but it does not replace it, and the duty resting on the organisation remains the one in Article 50, not the one in the code. It is the same distinction, between an adequate instrument and a presumption of conformity, that we had occasion to examine when the adequacy assessment was made known, and that the Commission’s own text now confirms in the plainest terms.
Those who do not sign
Not signing does not amount to non-compliance. The Commission makes clear that adherence is voluntary and that any enforcement would concern a breach of Article 50, not the absence of adherence to the code. Non-signatories remain bound, however, to demonstrate compliance by other adequate means, whose suitability will be assessed by the market surveillance authorities, and must be able to explain and document how their chosen measures ensure that the obligations are met. The Commission suggests they carry out a gap analysis against the measures set out in the code, and warns that they may receive a greater number of requests for information or access. The cost of not signing is therefore not a penalty: it is a heavier evidentiary burden and a more intense exposure to supervision.
What the Digital Omnibus promises, and why it is not yet law
One element eases the pressure straight away, and rests on the AI Act as it stands: content generated or manipulated and already made available before 2 August 2026 need not be marked or labelled retroactively. The second must be handled with care, because it is routinely presented as though it were settled. The Digital Omnibus on AI contains a targeted grandfathering rule for the marking and detection obligations in Article 50(2), covering generative AI systems placed on the market or put into service before 2 August 2026, to which it would grant a transitional period to reach conformity by 2 December 2026. The text was approved by the European Parliament on 16 June 2026 and by the Council on 29 June. As of today, however, the regulation amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 has not been published in the Official Journal, and it will enter into force only on the third day following publication: the EUR-Lex record for the AI Act currently lists, among the modifications to the regulation, nothing but linguistic corrigenda. Until then the grandfathering rule is not applicable law and no planning can be built on it: what binds remains Article 50 as it stands, with its date of 2 August 2026. Anyone relying on the extension is relying on a text that has been adopted but not yet published, and for a provider who must demonstrate compliance on 3 August, that difference is not a matter of style.
It remains the case, in any event, that from 2 August all systems falling within the scope of Article 50 and placed on the market or put into service in the Union must comply, regardless of when they were first placed on the market, and that the timing of the code’s publication does not move the deadline set by the AI Act itself. On the technical side, finally, the code stays exposed to the evolution of standards: the AI Office has announced it will consider formal updates at least every two years, including in light of emerging standards, which returns to the familiar point of European marking resting on standards Europe does not govern alone.
For anyone deciding over the next fortnight, the operational picture is settled: establish which section of the code matches the role actually performed, identify the executive entitled to bind the organisation, submit the form by the deadline the Commission confirms on its own pages, and keep in mind that what signing buys is predictability in supervision, not exemption from proof. Choosing not to sign breaks no rule, but it means building, and documenting, on one’s own the path the code already sets out in writing.
