13 days until Article 50(1) enforcement. 4 months until Article 50(2) enforcement for legacy systems. Same article, different clocks.
Most founders are treating Article 50 as a single deadline: August 2. That is half right and half dangerous. Here is where the split is, which deadline applies to your product, and why the December date is the one most people will miss.
The two deadlines, plain and simple
| Deadline | What applies | To whom |
|---|---|---|
| August 2, 2026 (13 days) | Art. 50(1) - Chatbot & AI interaction disclosure | EVERY AI system that interacts with people. No exceptions. No delay. |
| August 2, 2026 | Art. 50(2) - Machine-readable marking | NEW AI systems put on the market after August 2. |
| December 2, 2026 (4 months extra) | Art. 50(2) - Machine-readable marking | EXISTING AI systems already on the market before August 2. |
If your AI feature shipped before August 2, you have until December 2 to add watermarking, provenance metadata, or equivalent machine-readable marking.
If your AI feature ships on August 3 or later, it needs those marks from day one.
And none of that affects the chatbot disclosure obligation. August 2 is August 2 for every chatbot, regardless of when it launched.
Why this split exists
The Digital Omnibus (provisional agreement, May 2026) gave GenAI providers a 4-month transition for the watermarking requirement under Art. 50(2). The reasoning: retroactively marking millions of pieces of pre-existing AI content takes time that a new system does not need.
The EU did NOT extend the same transition to Art. 50(1) - the chatbot disclosure rule. There is no technical burden in adding a one-line "I am an AI" to a chatbot. That is purely a policy decision the Commission saw no reason to delay.
So the split is practical: instant for simple disclosure, 4-month runway for technical marking.
Where founders get tripped up
Three patterns I keep seeing:
"I use GPT to draft blog posts. August 2 is the deadline for everything, right?"
August 2 is the deadline for your chatbot to disclose itself as AI. But if you have been using GPT to draft content since June 2025, the watermarking obligation for those output artifacts lands on December 2. You have until then to add provenance metadata to your content pipeline.
"Our chatbot launched in 2025. Does the December 2 extension apply?"
No. The December extension covers watermarking of AI-generated content, not chatbot disclosure. Your chatbot must disclose itself as AI by August 2 regardless of when it launched.
"We do not generate content publicly. Do we still need to worry about December?"
Only if your AI system produces any output that qualifies as machine-readable content under Art. 50(2). If your AI runs internal analysis or recommendations that never leave your system, you are likely fine. If your AI generates text, images, audio, or video that gets saved or exported (even for internal use), the marking obligation applies by December 2 for pre-existing systems.
What to do this week (before August 2)
For every founder, regardless of launch date:
- Chatbot disclosure. Add "I am an AI assistant" or equivalent to every AI interaction point. First message. Not hidden. Choose the phrasing that fits your product voice, just do not skip the disclosure itself.
- Risk classification. Write down what AI systems you run, what they do, and what risk tier they sit in. Most micro-SaaS products are "limited risk" - the bar is transparency, not the full Annex III regime.
Both take under 10 minutes. Do them now so you are not scrambling on August 1.
What to do by December 2
For pre-existing systems only:
- Content provenance pipeline. Set up machine-readable marking for every AI-generated output your system produces. This is typically metadata headers, watermarking, or equivalent.
- Detection mechanism. Make it possible for downstream users to verify whether content came from your AI system. Documentation of your approach counts.
- Audit trail. Keep a record of when you implemented the marking and what systems it covers.
New systems need this from launch day. For existing systems, you have 4 months.
The risk of getting this wrong
Article 50 penalties run up to EUR 15 million or 3% of global annual turnover. The enforcement agencies across EU member states are actively hiring and training market surveillance teams - the "they won't enforce" argument is wishful thinking.
The more common risk is commercial: enterprise procurement teams are adding AI Act compliance questions to vendor security reviews. A founder who cannot describe their transparency approach loses deals they never knew they were competing for. A founder who confuses the two deadlines looks unprepared.
Where Compliance Copilot fits
The EU AI Act module covers both deadlines:
- August 2 readiness: Art. 50(1) chatbot disclosure checklist, risk classification per system, transparency report export
- December 2 readiness: Watermarking and provenance requirements mapped to Art. 50(2), implementation tracking, audit trail generation
- Deadline calendar: We track which of your systems are pre-existing vs new and tell you which deadline applies to each
Founders Beta at $9/mo for the first 50 customers. Full module at $49/mo after.
Start your EU AI Act compliance at securemymvp.com/eu-ai-act ->
This article is informational, not legal advice. Consult qualified counsel for compliance decisions specific to your business.