Core claim: For most software teams the honest answer is no — the AI Act’s Article 50 transparency rules are aimed at AI systems that interact with people and at synthetic content presented to the public, not at source code written with an assistant — but the comfort in that answer is thinner than it looks, because the question buyers, courts and regulators are actually asking about AI-written code is not answered by Article 50 at all.

What Article 50 actually covers — and its current status

Article 50 of the AI Act (Regulation (EU) 2024/1689) is the transparency article. In broad strokes it requires that: people are told when they are interacting with an AI system rather than a human; providers of AI systems that generate synthetic audio, image, video or text ensure the output is marked as artificially generated in a machine-readable way; and deployers disclose deepfakes and AI-generated text published to inform the public on matters of public interest.

Status matters, because 2026 muddied it for many readers. The “Digital Omnibus on AI” (Regulation (EU) 2026/1744, in force since 27 July 2026) postponed the AI Act’s high-risk obligations — standalone Annex III systems to 2 December 2027, AI embedded in regulated products to 2 August 2028. It did not touch Article 50: the transparency obligations entered into application as planned on 2 August 2026. If your mental model is “the AI Act got delayed,” it is one article too coarse.

Where AI-written code sits: mostly outside — honestly

Read against a normal development workflow, Article 50’s duties attach awkwardly to code at best. The marking duty for synthetic content sits with the provider of the AI system — the maker of the coding assistant or agent platform — not with the team using it; and it is aimed at content categories (audio, image, video, text) in contexts where a person might be deceived about origin. Source code sitting in a private repository is not “published to inform the public,” and your compiled product is not, in any ordinary reading, the AI system’s output presented to an end user. A team using an AI assistant to write software is, in Article 50’s vocabulary, neither marking-obligated provider nor deepfake-publishing deployer in the typical case. This is not just our reading: the Commission’s own guidance names source code among the outputs that fall outside Article 50(2)’s marking obligations.

Honesty requires the edge cases too. If your product embeds a model that generates content for users, you may carry provider or deployer duties in your own right. If you publish AI-generated technical prose — documentation, blog posts — as public-interest information, the text-disclosure duty can reach it. And whether unusual configurations fall in or out of Article 50’s categories is precisely a question for counsel, not a blog post: this piece describes the delimitation, it does not rule on yours.

The wrong comfort: “not covered” is not “not asked”

Here is the trap in stopping at “no.” Article 50 is a transparency rule about deception — its concern is a person not knowing they face a machine. The questions arriving about AI-written code come from a different direction entirely: a due-diligence team asking what share of the codebase is machine-authored and how it was reviewed; a court applying the Product Liability Directive’s disclosure mechanics from 9 December 2026; a market surveillance authority reading CRA technical documentation from September 2026 onward. None of those actors cites Article 50 — and none of them is satisfied by “Article 50 doesn’t apply to us.”

What they want is provenance: who or what wrote this, under whose review, with what verification — recorded when it happened. Teams that treat “we’re outside Article 50” as the end of the AI-authorship question will meet the same question again wearing a different statute, and the evidence that answers it cannot be created retroactively.

So: does AI-written code count as AI-generated content under Article 50? For most teams, honestly, no. Does that make the authorship of your code legally irrelevant? Also no — and the regimes that do reach it are already in motion.

This closes the loop from article 1, “Who wrote your code? The provenance gap in AI-built software”: Article 50 was never the reason to know who wrote your code — the provenance gap is.

Sources