Article 50, six weeks in: the official playbook is published, zero fines have landed, and December 2 is the next deadline

The transparency obligations of the EU AI Act have been live since 2 August 2026. For six weeks, enforcement has been quiet — zero fines. Now the European Commission has published the official Article 50 guidelines, giving deployers the first authoritative playbook. This is the rare moment when compliance got easier to understand and cheaper to act on. It will not stay that way.

Six weeks in — and the ground just got firmer

The transparency rules of the EU AI Act (Regulation 2024/1689, Article 50) took effect on 2 August 2026. Since then, we have watched the same pattern repeat in client conversations: everyone knows the rules exist, almost nobody has done anything about them, and a quiet hope circulates that “nothing has happened yet, so maybe nothing will.”

Two developments this month should retire that hope — and replace it with a workable plan.

First: the official playbook now exists. The European Commission has published its official guidelines on Article 50 transparency obligations (digital-strategy.ec.europa.eu, page dated 6 August 2026, with an FAQ published 24 July 2026). For the first time, deployers — the companies that use AI systems and publish AI-generated content, not the companies that build them — have an authoritative playbook: what to disclose, how to mark, who is responsible.

Second: the scoreboard is still zero. Six weeks of enforcement have produced no reported fines. That is not a reprieve. It is a head start — the period regulators use to establish interpretation, and the period in which a disciplined SME can become compliant quietly, before the first test cases make compliance expensive and public.

What the official guidelines actually require of deployers

The guidelines, together with the FAQ and the Code of Practice, turn the legal text into a checklist. Here is the deployer-relevant core, stripped of jargon.

1. Explicit AI-interaction disclosure

If a human interacts with an AI system — a chatbot on your website, an assistant in your customer service — they must be told, explicitly, that they are interacting with AI. Not buried in a footer: the guidelines treat this as a disclosure the user must actually perceive. For an SME running a chatbot or an automated email assistant, this is a one-line fix with no excuses.

2. Machine-readable marking of AI-generated content

AI-generated content — text, images, audio, video produced or manipulated by AI — must be marked in a way machines can detect: provenance metadata that travels with the file. This is the technical core of Article 50, and it is where the Commission has set a transition period that ends on 2 December 2026. After that date, machine-readable marking moves from good practice to obligation.

3. Visible labelling for high-impact content

Certain AI-generated content must also be labelled in a way humans can see:

• Deepfakes — synthetic image, audio, or video depicting real people, objects, places, or events convincingly: disclose that the content is artificially generated.
• AI-generated text published to inform the public on matters of public interest — publish it without review, and you must disclose its artificial origin.
• Emotion recognition systems, where lawfully deployed — the people affected must be informed.

The Commission’s guidance is deliberately practical here: the duty follows the content, and the deployer who publishes it carries the duty. If your marketing team publishes an AI-generated image that could pass as a photograph of your product, the disclosure duty is yours — not your vendor’s.

4. Providers vs. deployers — know which one you are

The guidelines draw the line the previous discourse blurred. Providers build the AI systems; deployers use them and publish the output. Most SMEs are deployers — and most of Article 50’s practical duties land on the deployer. The provider’s obligations matter to you too, though: when choosing AI vendors, their compliance posture (transparency, watermarking support, documentation) becomes part of your own compliance story. Notably, Anthropic remains the only major AI provider that has signed the Article 50(2) Code of Practice, committing to watermarking and C2PA content provenance on its output (TNW, 12 August 2026) — a useful signal when you evaluate which models sit inside your content pipeline.

Six weeks, zero fines — why that is a window, not a reprieve

There is a predictable life cycle to EU enforcement: guidance first, interpretation next, test cases after that. The zero-fine scoreboard is not evidence that Article 50 will stay quiet. It is evidence that we are still in the guidance phase — and the guidance has now been published.

The quiet window ends the way these windows always end: a visible test case, an authority looking for an example, or a competitor complaint. When the first fine arrives, it will be covered in the trade press your customers read, and every SME still unlabelled will look either negligent or oblivious.

The calendar gives the real deadline away. The 2 December 2026 machine-readable marking deadline is roughly twelve weeks out. After that, the compliance story shifts from “publish a disclosure line” to “prove the provenance of your content” — a materially bigger lift for anyone who starts late. And the penalty regime attached to AI Act violations runs to up to €15 million or 3% of global annual turnover. For a mid-sized SME, that is not a rounding error.

So: six weeks in, zero fines, official playbook published. This is the cheapest compliance moment you will get.

Three things to do before 2 December 2026

Action 1 — Inventory your AI touchpoints (this week)

List every place AI touches your customer-facing communication: chatbots, automated email sequences, AI-generated images or video on your site and social channels, AI-assisted blog posts and newsletters. For each: who generated it (which tool/model), who published it (you), and whether a customer could currently mistake it for human-made. This inventory is the entire foundation — most SMEs discover in this step that they have AI content they forgot they had.

Action 2 — Put disclosure in place where it is due now

The disclosure duties are live today; the marking deadline is in December. Do the cheap part immediately: add “this is an AI assistant” lines to chatbots and automated senders, add visible captions to AI-generated imagery that could pass as photography, and review AI-assisted text before publication (substantive human review keeps text outside the strictest labelling duty — the EC guidelines state that superficial checks like spell-checking do not qualify, and unreviewed public-interest text must be labelled). None of this requires new tools. It requires an afternoon and a standard.

Action 3 — Prepare for machine-readable marking now, not in November

For content where machine-readable marking applies, provenance metadata needs to become part of your production pipeline — which means your tools and vendors need to support it. This is the action with lead time: test whether your current AI tools emit provenance metadata, ask vendors for their compliance roadmap (their answer is informative — see the Anthropic note above), and build the marking step into your content workflow before it becomes a deadline-driven scramble.

The managed shortcut

If this reads like a second job, that is accurate — and it is the argument for managed delivery. In a managed model, disclosure lines, review-before-publish, provenance-aware tooling, and an audit trail are simply how the service operates: compliance by default, not a project. The obligations do not shrink; the operating burden does.

FAQ

Do the Article 50 rules apply to a Swiss company?

Directly if you operate in the EU market or publish to EU customers — and in practice they set the standard for AI transparency across Europe, including Switzerland. Swiss firms with EU customers, EU-facing websites, or EU advertising should treat Article 50 as their operating baseline.

Does every AI-assisted email or social post need a label?

No. Article 50 is targeted, not total: deepfakes, unreviewed public-interest text, and content that could mislead by appearing human-made. Ordinary AI-assisted internal drafts or marketing copy that receives substantive human review are generally outside the strict labelling duties (the EC guidelines are explicit that spell-checking or grammatical correction alone do not count as human review) — but disclosure of AI interaction (chatbots, assistants) and machine-readable marking obligations still apply on their own schedule.

What happens on 2 December 2026 exactly?

The transition period for machine-readable marking ends. From that date, deployers are expected to ensure AI-generated content carries detectable provenance marking, with enforcement backed by penalties up to €15 million or 3% of global turnover.

We use chatbots on our website — what must we do right now?

One thing, today: tell users they are talking to an AI. That is the explicit-interaction disclosure, it is the cheapest item in the entire regulation, and it is currently the most commonly missed one.

From here

Article 50 is now the best-documented, best-templatized part of the EU AI Act — and the quietest enforcement window you will ever see. Twelve weeks to the marking deadline.

Contact us for a free 30-minute compliance assessment — we will run the inventory in Action 1 with you and tell you exactly what is missing.

Sources

• European Commission — official Article 50 transparency guidelines (digital-strategy.ec.europa.eu, 6 Aug 2026); Article 50 FAQ (24 Jul 2026); GPAI Code of Practice (Jul 2025).
• TNW, 12 Aug 2026 — Anthropic, sole major provider signed to the Article 50(2) Code of Practice (watermarking + C2PA).
• Regulation (EU) 2024/1689 (EU AI Act), Article 50; marking transition period ending 2 December 2026.
• Enforcement status: weekly news sweep, w/e 7 Sep 2026 — no fines reported in the first six weeks of application.

Ai-Fi — AI business solutions for Swiss and Italian SMEs. Compliance by default: disclosure, review layers, provenance-aware tooling, and audit trails built into managed delivery. This article is marketing content, not legal advice; figures and obligations cited as published by the European Commission.