The Digital Omnibus gave most teams sixteen months of breathing room on high-risk AI obligations. That's real relief, and it's easy to see why compliance leads relaxed. But one deadline was never part of that deal. Article 50 transparency obligations come into force on 2 August 2026, and for anyone running a chatbot, generating synthetic content, or deploying emotion recognition, that date is now days away, not months.
What actually lands on 2 August
Article 50 covers four specific disclosure duties. AI systems that interact directly with people, chatbots, voice assistants, virtual agents, must make clear the user is talking to a machine, not a human. Emotion recognition and biometric categorisation systems need to notify the people they're analysing. AI-generated or manipulated content, including deepfakes, needs machine readable marking. Text published to inform the public on matters of public interest needs a disclosure if it was AI generated, unless a human has reviewed it editorially.
None of this is new law. It's the part of the Act the Digital Omnibus explicitly left untouched. Existing systems get a four month grace period on watermarking specifically, running to 2 December 2026, but the core disclosure duties for new deployments start now.
The gap between "we know about this" and "we've tested it"
Most teams we talk to can recite the August deadline. Fewer can point to where in their product the disclosure actually happens, or confirm it survives a redesign, an A/B test variant, or a third-party widget embedded in the flow. That's the practical risk here. Compliance isn't a policy document, it's a behaviour that has to be verified every time the interface changes.
Before this week is out, it's worth checking three things. First, does every conversational AI surface in your product state clearly, and early, that it's AI, not a person? Second, if you generate images, video or audio, is there a durable, machine readable marker attached to the output, not just a visible watermark that gets cropped or stripped? Third, if any part of your product uses emotion or sentiment detection on user input, is that disclosed before the data is captured, not buried in a privacy policy nobody reads?
The takeaway for QA and product teams
Treat Article 50 the way you'd treat any other functional requirement: write a test for it, run it against every release, and don't rely on a design review from six months ago to still be accurate. The Omnibus bought time on high-risk classification. It didn't buy time here.
If you need help mapping these disclosure requirements to concrete test cases you can run against your own product, that's exactly what EvalSpec is for.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation