Article 50 of the EU AI Act — the transparency chapter — has been in force since 2 August 2026, and on 2 December 2026 the marking grace period for generative tools that were already on the market closes for good. Most marketing teams now run chatbots, synthetic imagery or AI-drafted copy somewhere in the funnel, which means the ai act article 50 checklist below isn't a future compliance exercise: every line of it describes duties that already apply today.

This page turns Article 50 into four working checklists — chatbots, synthetic media, synthetic text, and the governance that keeps them honest. Work through it with your martech owner in the room; it takes most teams an afternoon.

Before You Start: Which Side of the Duty Are You On?

Article 50 splits its obligations between providers (the vendors who build AI systems) and deployers (the businesses that use them — that's you). The machine-readable marking of synthetic content is a provider duty; the human-visible disclosures are largely yours. If that split is new, read Machine-Readable Marking for Generative Content: What Marketing Teams Must Do first — it covers the marking layer in depth. This checklist covers everything on your side of the line.

One framing note: these are ai transparency obligations marketing inherits from regulation, not optional brand-safety hygiene. They apply to any business marketing into the EU, wherever it is domiciled.

Checklist 1 — Chatbots and Conversational AI (Article 50(1))

  • Inventory every conversational surface. Website chat widgets, WhatsApp and Messenger bots, voice assistants on phone lines, in-app assistants, sales development bots in DMs. Include tools your agency runs on your behalf.

  • Disclose at the first interaction, in plain language. "You're chatting with an AI assistant" — not buried in a terms link. For voice bots, the disclosure must be spoken in the same channel.

  • Never rely on "it's obvious". The exemption is narrow and assumes a reasonably observant person. If a user could plausibly think it's a human agent, disclose.

  • Check the disclosure survives configuration. Many vendors ship compliant defaults that your custom welcome-message editor quietly overwrites. Test the live bot, not the vendor's demo.

  • Script the human handoff. When the bot escalates to a person, say so — undisclosed agent-swapping is the most common complaint pattern regulators see.

  • Log where disclosure appears, with a screenshot and date, per bot per channel. You want evidence, not memory.

Checklist 2 — Synthetic Images, Audio and Video (Articles 50(2)–(3))

  • List every generative image, audio and video tool in use — including "just background music" generators, AI avatars, voiceover cloning and image-to-image editing.

  • Confirm with each vendor that output carries machine-readable marking (signed provenance metadata, watermarks). Ask in writing. This is the provider's duty, but your record-keeping protects you.

  • Disclose deepfakes and inauthentic depictions visibly — in or on the content itself, not in a caption three posts earlier. "Clearly distinguishable" is the standard.

  • Apply platform labelling on top. Meta, TikTok, YouTube and LinkedIn all run AI-content labels with their own advertiser rules; EU obligations and platform policies stack, they don't substitute.

  • Check that marks survive your pipeline. Re-encoding, editing suites and some ad traffickers strip provenance metadata. Preserve original files; verify exports.

  • Keep consent records for synthetic likenesses. Cloning a real person's face or voice drags in GDPR and publicity rights alongside Article 50 — the disclosure label does not cure an unlawful likeness.

  • Flag the creative carve-out correctly. Evidently artistic, satirical or fictional work gets a lighter disclosure touch, but only where manipulation is still acknowledged somewhere it doesn't hamper the piece. Your lawyers, not your art director, decide what is "evident".

Checklist 3 — Synthetic Text (Article 50(3))

  • Decide what counts as public-interest content for you. Market commentary, statements on health, safety, finance, elections, or anything a reasonable reader would treat as factual reporting.

  • Route all such content through documented human editorial review, with a named editor holding responsibility. Keep the review log.

  • Disclose AI involvement on public-interest pieces that skip full review — automated market updates and similar. Small label, visible on the page.

  • Treat platform policies as the real text constraint. Ad platforms and networks increasingly require AI-content declarations regardless of Article 50's scope. Check your campaign settings.

Checklist 4 — Governance, Records and Vendors

Transparency obligations decay fast because tools change fast. The governance model: who owns what in your marketing stack is what keeps this checklist honest six months from now.

  • Name one owner for Article 50 compliance in marketing — with the vendor register, disclosure evidence and re-verification calendar under that role.

  • Maintain a tool register: every generative tool, what it produces, its marking capability, its disclosure features, and the date you last verified both. Shadow tools are the silent failure mode — see our piece on shadow AI compliance risk.

  • Put vendor re-verification in the calendar quarterly. Model swaps and export-path changes break marking quietly.

  • Train the team. AI literacy training under Article 4 is its own legal duty, and it is the practical mechanism that makes checklists like this one survive contact with campaign deadlines.

  • Brief agencies and freelancers. You carry deployer duties for work done on your behalf; make Article 50 expectations contractual, not assumed.

The Two Dates to Circle

Everything above is already in force since 2 August 2026. What changes on 2 December 2026 is that generative systems already on the market lose their marking grace period — the date most teams missed while reading "AI Act delayed" headlines. We've unpacked it in 2 December 2026: the AI Act deadline most marketing teams don't know about, and the full obligation map is in the AI Act compliance deadlines timeline.

The Honest Upside

A chatbot that says what it is converts trust into pipeline. A labelled synthetic video survives the platform-scan purges that unlabelled content won't. Provenance marks are the same infrastructure brand-safety teams have wanted for years. Done properly, Article 50 compliance reads to buyers as operational maturity — the AI Act resource hub collects the rest of the playbook.

Practical guidance, not legal advice — we work alongside your legal counsel.