Article 50 takes effect on 2 August — here is what you need to know
Article 50 of the EU AI Act applies from 2 August 2026. Its duties land on the systems that talk to people and generate content — not the retrieval layer that feeds them. Anyone selling you 'Article 50 compliance' for a vector database is fear-selling; here is the honest split.
by David Bunting, Founder
The rule, in one paragraph
Article 50 of the EU AI Act — Regulation (EU) 2024/1689 — is the transparency chapter, and it starts applying on Sunday, 2 August 2026. It is a closed list of four situations in which an AI system has to tell people what it is doing: systems that talk to a person, systems that generate synthetic audio, image, video or text, systems that read emotion or sort people into biometric categories, and systems that produce deepfakes or publish AI-written text on matters of public interest. Miss one that applies to you and the ceiling is €15 million or 3% of worldwide annual turnover, whichever is higher.
The “Digital Omnibus” simplification package, politically agreed in May 2026, moved several of the high-risk deadlines. It did not move this one. Transparency and enforcement stayed on 2 August; only the machine-readable marking of generative systems already on the market got a short runway, to 2 December 2026.
Who each obligation actually lands on
The Act splits its duties between the provider of a system and its deployer, and Article 50 rewards reading with that split in hand, because it decides who owns each line.
The disclosure that you are talking to an AI (Article 50(1)) sits with the provider of the system doing the talking — the chatbot, the assistant, the copilot. Marking generated content as artificial (50(2)) sits with the provider of the generative model that produced it. For most teams that provider is the vendor of the AI client they have adopted, not the team itself — and certainly not the software sitting behind it. The emotion-recognition notice (50(3)) and the deepfake disclosure (50(4)) are deployer duties, and they only bite if you are actually running emotion recognition, biometric sorting, or synthetic-media generation.
Run that list against a retrieval layer and it comes up empty. Laminae holds no conversation with anyone — the AI client does that. It generates no synthetic content — it returns passages copied verbatim from your own documents. It does no emotion recognition and produces no deepfakes. A source-attributed retrieval layer is in none of Article 50’s four situations, because returning evidence is not one of the things Article 50 regulates.
Key dates
“Article 50 compliance” for a vector database is a tell
In the run-up to any deadline, the compliance-flavoured pitch appears: buy this, tick Article 50. Be suspicious when it is pointed at your storage or retrieval layer. A vector database does not interact with a person, does not generate content, and does not recognise a face — there is no Article 50 sticker for it to earn, because there is no Article 50 obligation for it to meet. A vendor selling you one is either misreading the text or counting on you not to read it.
The knowledge layer is not where Article 50 lands.
Where the knowledge layer does the work
This doesn't mean retrieval is irrelevant to the day. Article 50 is a transparency law, and transparency has a substrate: you can only disclose what you can trace. The obligations land on the generation layer, but the thing that makes an answer from that layer defensible — the thing that turns “the AI said so” into “here is the clause it read” — is provenance, and provenance is retrieval’s job.
Two of the obligations make the point directly. The marking rule in 50(2) exists because, once text has been generated, telling authored from synthetic is hard — and a retrieval layer that returns your source passages verbatim hands the AI client real material to quote instead of prose to invent. The carve-out in 50(4) — AI-written public-interest text is exempt where it is “subject to human review and editorial responsibility” — only means something if the human doing the review can see the source. Editorial responsibility without a traceable source is a signature on a page nobody read.
This is the argument the product has been making all along. Laminae does not summarise in the middle: retrieval returns the chunks themselves — each carrying its document, page, and snippet — so the chain from an answer back to an exact span stays intact. And when an auditor asks to see the query logs behind a disclosure, self-hosting keeps them in your own database rather than a vendor’s console.
So: the deadline is real, the fine is real, and most of the work sits with the system that generates the answer, not the one that feeds it. The knowledge layer’s job on 2 August is the job it had on 1 August — to make whatever your AI client says something you can trace back to a document.