EU AI Act Article 50: what applies from 2 December 2026, a developer and CTO checklist

Short version. Article 50 of the AI Act has applied since 2 August 2026. The date 2 December 2026 is narrower than many posts suggest: it is the end of a transitional period of four months for one duty only, the Article 50(2) duty to mark outputs in a machine readable format and make them detectable, and only for providers of generative AI systems that were placed on the market before 2 August 2026. Everything else in Article 50 already applies.

Whether you are a provider or a deployer, and which paragraphs reach your system, is your own assessment.

The dates, with sources

DateWhat happensSource
1 Aug 2024The AI Act, Regulation (EU) 2024/1689, enters into force.Regulation (EU) 2024/1689 in the EU law database
2 Aug 2026Article 50 applies (Art. 113). From this date providers and deployers must meet Art. 50(1) to (5).Regulation (EU) 2024/1689, Art. 113, in the EU law database; Commission Q&A on Article 50
24 Jul 2026The Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, is published (OJ L, 24.7.2026). It adds Art. 111(4) to the AI Act.Regulation (EU) 2026/1744 in the EU law database; European Parliament Legislative Train
2 Dec 2026Providers of AI systems, including general purpose AI systems, that generate synthetic audio, image, video or text and were placed on the market before 2 August 2026 must meet Art. 50(2) by this date (new Art. 111(4); Recital 38 calls it "a transitional period of four months").Regulation (EU) 2026/1744, Art. 1 point (39)(b) and Recital 38, in the EU law database

The Commission's Article 50 Q&A, in summary: the grace period covers only AI systems placed on the market before 2 August 2026, and only the Art. 50(2) obligation to mark content generated by AI and make it detectable. Providers of such systems must meet that obligation only from 2 December 2026. It also says content generated before 2 August 2026 "does not need to be labelled retroactively". (Commission Q&A)

The Commission has also published Guidelines on the transparency obligations under Article 50 in all EU languages. They are not binding.

Other provisions of the Omnibus also carry 2 December dates. They are outside Article 50 and outside this page.

Who carries which Article 50 duty

  • If you are a provider of an AI system that interacts directly with people (a chatbot, an agent, an avatar): design it so people are informed they are interacting with an AI system, unless that is obvious (Art. 50(1)). Applies since 2 Aug 2026.
  • If you are a provider of a system that generates synthetic audio, image, video or text: outputs must be marked in a machine readable format and be detectable as generated or manipulated by AI (Art. 50(2)). Applies since 2 Aug 2026 for systems placed on the market from that date; by 2 Dec 2026 for systems already on the market before it.
  • If you are a deployer of emotion recognition or biometric categorisation: inform the people exposed (Art. 50(3)). Applies since 2 Aug 2026.
  • If you are a deployer that publishes deepfakes, or text generated by AI and meant to inform the public on matters of public interest: disclose it, with the exceptions in Art. 50(4) (for text, human review or editorial control with editorial responsibility). Applies since 2 Aug 2026.
  • All of the above: the information must be clear and distinguishable, at the latest at the first interaction or exposure, and meet accessibility requirements (Art. 50(5)).

Developer and CTO checklist

  1. Inventory. List every place your product calls a model: chat, agents, generated text, images, audio, video. Note when each system was first placed on the market. Only systems on the market before 2 Aug 2026 get the 2 Dec 2026 date for Art. 50(2).
  2. Decide your role per system. Provider, deployer, or both. Write the reasoning down. Spurvik does not classify your use case; its riskTier setting is your own classification.
  3. Art. 50(1) notice. If people chat with your system, show a visible notice from the first message. Options: <AIDisclosureLabel /> from spurvik/react, or the script tag for any HTML site (both have default copy in 24 EU languages; review the wording before use).
  4. Art. 50(2) marking. If you provide a generative system, pick a way to mark each type of output in a machine readable format. Spurvik does not mark content. Its provenance record, signed with HMAC, and marking hooks support, but do not replace, Art. 50(2) marking. Plug a marking tool into the hooks.
  5. Art. 50(3) and 50(4). If you deploy emotion recognition or biometric categorisation, or publish deepfakes or text on matters of public interest, plan the notices. AIDisclosureLabel has notice variants for these ("generated", "deepfake", "public-interest", "emotion-biometric").
  6. Keep an AI event log you can check. Not an Article 50 requirement, but useful evidence of what your system did. Spurvik records each wrapped call as a receipt signed with HMAC SHA256 with keyed fingerprints instead of prompt text. See AI event logging with receipts signed with HMAC.
  7. Gate risky agent actions. If an agent can refund, delete or send, put an approval check in front of it. See an approval check before an agent acts.
  8. Write it down and review it. Keep your reasoning, notice texts and marking choices with your counsel's review.

Code: the notice in two forms

Any HTML site (Webflow, WordPress, Shopify, static): add the script tag before </body>. Copy it from the Try it block below; attributes and CSP notes are on the badge card in the docs.

React / Next.js (spurvik@1.0.8, MIT):

import { AIDisclosureLabel } from "spurvik/react";

export function ChatHeader() {
  return (
    <AIDisclosureLabel notice="interaction" systemName="Support Assistant" locale="en" presentation="inline" />
  );
}

The SDK and MCP server are MIT and run in your own stack.

Try it

Copy the badge
<script src="https://spurvik.com/badge.js" data-system="Your AI" data-attribution="true" defer></script>
  • A small “Art. 50 · Spurvik” link shows after the notice by default. On Pro (€19/mo) you can hide it with data-attribution="false" for a white-label badge.
  • Or install the SDK: npm i spurvik and use AIDisclosureLabel from spurvik/react
Open the sandbox, no login Unlock Pro €19/mo

Free €0 (10k events/mo) · Pro €19/mo · Team €79/mo · excl. VAT · B2B only.

FAQ

Does Article 50 apply now?

Yes. Article 50 has applied since 2 August 2026 (Art. 113 of Regulation (EU) 2024/1689).

What exactly is due on 2 December 2026?

Only the Art. 50(2) marking and detection duty, and only for providers of generative AI systems placed on the market before 2 August 2026 (Art. 111(4), added by Regulation (EU) 2026/1744). Systems placed on the market from 2 August 2026 had no transition.

Does the 2 December date move the Art. 50(1) chatbot notice?

No. The notice duty in Art. 50(1) has applied since 2 August 2026 without a transition.

Do I have to relabel content generated before 2 August 2026?

The Commission's Q&A says such content does not need to be labelled retroactively, though it encourages deployers to do so where possible.

Does Spurvik mark content generated by AI?

No. Spurvik shows notices and records signed receipts. Its provenance record supports, but does not replace, Art. 50(2) marking in a machine readable format.

Does Article 50 require logging?

No. Event logging is optional evidence for your own records; Article 50 does not require it.

Sources

Reviewed 11 November 2026. Not legal advice. Supports parts of your Art. 50 work; responsibility stays with you. EyeTea Solutions ApS · Trekantområdet, Denmark

Your AI Act controls should live in a few lines of code, not in six months of legal memos.