Existing systems and legacy content: what does and does not apply retroactively

Point (153) of the guidelines of 20 July 2026 states that under Article 113, Article 50 applies from 2 August 2026 and that all in-scope AI systems placed on the market or put into service in the Union must comply on that date, regardless of their date of placing on the market or putting into service. That same point describes that the regulation amending the AI Act provides a targeted transitional rule concerning only the marking and detection obligations under Article 50(2) for generative AI systems placed on the market or put into service before 2 August 2026, giving providers of those existing systems a transitional period to bring them into conformity by 2 December 2026. Point (153) states expressly that systems that are partly interactive and partly generative may benefit from that transitional period only with regard to the marking obligation under Article 50(2), while compliance with the disclosure obligation for AI systems directly interacting with natural persons must be ensured as of 2 August 2026. Point (154) states that AI-generated or manipulated outputs within the scope of Article 50(2) and deep fakes within the scope of Article 50(4), first subparagraph, generated or manipulated before 2 August 2026 do not need to be marked or labelled retroactively, and that the same applies to texts on matters of public interest that were AI-generated or manipulated and published before that date. That same point states conversely that texts generated or manipulated before 2 August 2026 but published on or after that date must be labelled. Point (154) adds that deployers and other actors in possession of or disseminating pre-existing unlabelled deep fakes are nonetheless encouraged to label them, without being expected to engage in disproportionate efforts such as auditing pre-existing content databases or modifying already printed product packaging. A footnote to point (153) clarifies that the special grandfathering rule of Article 111(2) for high-risk AI systems does not extend to Article 50, because the high-risk obligations and the transparency obligations apply cumulatively.

• Locator: Commission Guidelines C(2026) 5054 final, Section 8.4, points (153) and (154)praxikon:eu:ai-act:source:commission-article-50-guidelinesOpen official source

Two things get conflated here. The first is the assumption that the transitional period to 2 December 2026 is a general postponement of Article 50. It is not: it covers only the machine-readable marking and detection of Article 50(2). If your system also runs a chat interface, the disclosure to the user had to be in place on 2 August 2026, even though marking of the output may wait until December. The second is the assumption that legacy content is safe. That holds for already published content, but not for your stock. If you had a batch of AI texts produced last year and are only publishing them now, the labelling duty does apply. The same goes for a campaign visual that has been sitting on the shelf for months. Practically, the distinction is not old versus new, but the moment of publication. The encouragement to label pre-existing deep fakes is not a duty, but it is where a supervisor can look at the reasonableness of your choices.

• Locator: Commission Guidelines C(2026) 5054 final, Section 8.4, points (153) and (154)praxikon:eu:ai-act:source:commission-article-50-guidelinesOpen official source

Split your inventory into three: systems interacting directly with people, which have had to disclose since 2 August 2026; generative systems relying on the transitional period to 2 December 2026 for machine-readable marking; and systems doing both, which therefore carry a different date per obligation. Also walk through your content stock: anything still to be published that was AI-generated or substantially edited gets a label at publication, regardless of when it was generated. Record why you are not retroactively labelling already published legacy content, so you can explain that choice.

• Locator: Commission Guidelines C(2026) 5054 final, Section 8.4, points (153) and (154)praxikon:eu:ai-act:source:commission-article-50-guidelinesOpen official source