If you've read my previous posts on the EU AI Act (Episode 10.8 – No risk, no fun with AI? and the update on the GPAI Code of Practice), you know the roadmap: February 2025 (AI literacy), August 2025 (GPAI obligations) – and now stage three has ignited.
The starting point: “should” has become “must”
Until now, plenty of companies could console themselves with the thought that the regulators weren't cracking down yet. That grace period is over. Since 2 August 2026:
- The transparency obligations under Article 50 are live
- The authorities are actively enforcing – fines included (reminder: up to 3% of global annual turnover or €15 million)
- For many high-risk AI systems, the full obligations now apply (exception: certain regulated products have until 2 August 2027)
The catch: many businesses have been using AI for ages, often without realising that this already puts them under Article 50. The “shadow AI” I've warned about before is now turning into a real compliance risk.
What actually applies
1. Chatbots have to come out
If you use AI in customer contact, you have to tell users they're talking to a machine, not a human. So the friendly “service agent Max” on your website can no longer travel AI-incognito.
Real-world example: an online retailer with an AI chatbot in support now needs a clear notice – ideally right at the start of the conversation, not buried in the terms and conditions.
2. AI-generated content must be labelled
If you use generative AI for text, images, audio or video, you must in certain cases ensure it's labelled – often even in a machine-readable way (think Content Credentials / C2PA metadata).
3. Deepfakes must be recognisable
Artificially generated or manipulated content depicting real people or events must be labelled as such. This doesn't just hit malicious actors, it hits marketing too: the AI-generated “testimonial customer” in your promo video counts.
4. High-risk AI: the full package
For high-risk systems (e.g. in HR recruiting, lending or critical infrastructure), comprehensive requirements now apply: technical documentation, risk management, human oversight, logging, and in some cases conformity assessment and registration.
Three scenarios – where does your company stand?
Scenario A: “We only use ChatGPT & co.”
Even then: review AI-generated content in your external communications, clarify labelling obligations, be able to demonstrate AI literacy. Effort: manageable, but not zero.
Scenario B: “We've integrated AI into customer processes.”
Chatbot disclosure, documentation of deployment contexts and, where relevant, checking whether the integration has turned you into a downstream provider. A structured AI inventory pays off here (if you don't have one yet – see Episode 10.8).
Scenario C: “We run high-risk applications.”
The full compliance machinery: risk management system, human oversight, logging, conformity assessment. If you're only starting now, bring in outside expertise – the authorities are looking closely.
Your options (don't panic 😉)
- Inventory: Which AI runs where? Including AI features in standard software and shadow AI.
- Classify: Which systems fall under Article 50, which are high-risk?
- Label: Implement disclosure notices and content labels – technically often trivial, organisationally the real effort.
- Document: If you have nothing to show when the regulator comes knocking, you lose the “good faith” bonus.
- Prioritise: Not everything has to be perfect. But visible, customer-facing systems (chatbots, generated content) are what regulators and consumers see first.
Bottom line
The EU means business – and, as I argued with the GPAI Code of Practice, this isn't a programme to stamp out the Wild West but an attempt to institutionalise trust in AI. Being transparent with users isn't harassment; it's simply good manners that have now become law. If you've done your homework over the past two years, you just need to fine-tune. If you haven't: the next deadline (August 2027 for regulated products) is coming for sure – but the current one is already binding.
Or, loosely following the motto from 10.8: with great [AI] power comes great responsibility – and since August 2026, a schedule of fines as well 😉
