Source Document

European Commission (DG Communications Networks, Content and Technology), "Transparency obligations under Article 50 of the AI Act", official FAQ, last updated 2026-07-24. Companion source: Council of the EU press release on the final approval of the digital omnibus (2026-06-29).

This review covers a regulatory guidance document, not a paper. It is not the text of the AI Act (Regulation (EU) 2024/1689) itself but the Commission's FAQ explaining how it reads and will administer Article 50, so it is not legally binding. It still carries unusual practical weight: it comes from the institution that drafted the law and coordinates its enforcement. That is also the conflict of interest to keep in mind — the author and the enforcement coordinator are the same body, with an incentive toward enforcement-friendly readings.

What Took Effect

The digital omnibus amendment, given final approval by the Council on 29 June, pushed obligations for stand-alone high-risk systems (Annex III) to 2027-12-02 and for product-embedded high-risk systems (Annex I) to 2028-08-02 (the Council press page blocked our fetch, so these dates were cross-checked against multiple reports of the Parliament's 16 June endorsement). Article 50 transparency obligations were excluded from the deferral and became applicable on 2026-08-02 as originally scheduled; this FAQ, updated on 24 July, is the Commission's latest reading just before that date.

ObligationWho is responsibleApplicable from (verified)
50(1) Disclose that users are interacting with AIProvider2026-08-02 · no grace period
50(2) Machine-readable marking of synthetic contentProvider2026-08-02 · systems on the market earlier: grace until 2026-12-02
50(3) Disclose emotion recognition / biometric categorisationDeployer2026-08-02
50(4) Label deepfakes and public-interest AI textDeployer2026-08-02
Annex III high-risk obligationsProvider & deployerDeferred to 2027-12-02 (omnibus)

Fines can reach 15 million euros or 3% of total worldwide turnover for the preceding financial year, with proportionality available for SMEs and small mid-caps, per the FAQ. The extraterritorial reach is explicit as well — providers established outside the EU are covered whenever their system's output is used inside the EU.

Verified Details

The 50(1) notice must come "from the start of the first interaction in a clear and distinguishable manner," and is waived only when the AI nature is obvious to an "average person, who is reasonably well-informed, circumspect, and observant." The 50(2) marking duty carries several operationally important exemptions — machine-to-machine-only outputs, closed-loop industrial environments, standard editing assistance, and short sequences of numbers, symbols, or source code. Text that has undergone human review or editorial control needs no label, and evidently artistic, creative, or satirical deepfakes only require disclosure in a manner that does not hamper the work's enjoyment. On retroactivity, the FAQ states that content generated before 2026-08-02 does not need to be labelled retroactively, while encouraging voluntary labelling where possible.

The enforcement structure differs from the common assumption. Compliance is enforced mainly by national market surveillance authorities; the AI Office is competent only for systems built on general-purpose AI models where the same entity provides both, or systems integrated into very large online platforms or search engines designated under the DSA. Systems used by EU institutions fall to the European Data Protection Supervisor.

Credibility Assessment

Three things support reliance on this document: it is a primary source from the institution that drafted the provision, its 24 July update postdates the omnibus so its timeline is current, and its dates and fine amounts are consistent with reports of the Council (June 29) and Parliament (June 16) approvals. The caveats are equally clear. An FAQ binds neither courts nor national authorities, so open-textured tests like "obvious from context" will only settle once enforcement cases accumulate, and with 27 national authorities as first-line enforcers, divergent readings are built into the structure. We could not open the Council press release itself; that remains a verification gap of this review.

Reviewer's Judgment

First, we judge the under-covered center of gravity of this regime to be the deployer obligations. Most coverage focuses on chatbot disclosure — a provider duty — but paragraphs 3 and 4 impose notice and labelling duties on companies that merely use AI rather than build it. The most common compliance error will be a company embedding a vendor chatbot widget, or a marketing team publishing generated images, assuming the rules do not apply to them.

Second, our same-day blog post's line that "enforcement may reach back to violations since August 2025" is not supported by this FAQ. The source says the opposite about content: nothing generated before 2026-08-02 must be labelled retroactively. The claim likely conflates the staged application of other AI Act provisions, so log retention is better designed as evidence for post-application audits, not as protection against retroactive investigation. In the same vein, "the AI Office's enforcement powers went live" does not directly apply to most chatbot operators — their actual counterpart is the national market surveillance authority.

Practical Levers

  • Settle the role matrix — Determine, system by system, whether you are the provider or the deployer. Companies using vendor widgets still need a 50(3)/50(4) deployer check.
  • Don't lean on the "obvious" exemption — The average-person test is an open standard with real litigation risk; defaulting to disclosure before the first response is the safer design.
  • Count down the marking grace period — Even systems on the market before 2 August only get until 2 December for machine-readable marking. Back-plan the watermarking and metadata work now.
  • Record editorial control — Human-reviewed text is exempt from labelling, so keep review logs that can prove the exemption applies.
  • Identify your enforcement counterpart — The market surveillance authority of your main EU market is the actual regulator. Put its follow-up guidance on your tracking list.

Conclusion

The omnibus did slow the high-risk timeline, but what this FAQ shows is that the transparency obligations arrived on schedule — with their exemptions and grace periods drawn in considerable detail. What operators need now is not fine-driven panic but three concrete tasks: role determination, documentation of exemption conditions, and a back-planned marking timeline. The operational checklist for turning these duties into deployment-gate metrics continues in the blog post below.

References