The question at hand
Article 50 § 1 of the EU AI Act requires providers to ensure users know they're interacting with an AI system "unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect." This standard creates an immediate fork in the compliance road.
Do you proactively disclose AI use in every interaction, or do you rely on contextual obviousness and disclose only when ambiguity exists? With enforcement beginning August 2, this isn't a philosophical debate. It's a design decision that affects user experience, legal exposure, and operational complexity.
The "reasonably well-informed, observant and circumspect" test sounds clear until you apply it to actual deployments. A chatbot widget labeled "AI Assistant" might seem obvious. But what about an AI-powered phone system that sounds human? An email response drafted by a language model but sent from a human's address? A customer service interaction where the agent uses AI to generate responses in real time?
The case for universal disclosure
Some compliance teams argue you should disclose AI use in every interaction, regardless of context. Their reasoning: the "obviousness" standard is subjective, varies by user sophistication, and creates litigation risk you don't need.
Universal disclosure eliminates judgment calls. You don't need to assess whether each deployment crosses the obviousness threshold. You don't need to train product teams on what counts as "reasonably well-informed" or "circumspect." You build one disclosure pattern and apply it everywhere.
This approach also protects you from regulatory second-guessing. If an authority questions whether your chatbot interface was "obvious enough," you can point to explicit disclosure. You've documented user awareness. You've created an audit trail showing you informed users before the interaction began, not after they complained.
From a user trust perspective, universal disclosure signals transparency. You're not hiding AI use or hoping users figure it out. You're stating it clearly, which can actually reduce friction when users encounter limitations or errors. They know they're working with AI, so they adjust their expectations accordingly.
The implementation is straightforward: add a disclosure banner, include a statement in the first message, or require acknowledgment before the session starts. You incur some UX cost, but you eliminate compliance ambiguity.
The case for context-dependent disclosure
The opposing view: universal disclosure wastes user attention, creates disclosure fatigue, and misreads the regulation's intent. Article 50 § 1 explicitly includes an obviousness exception because the drafters recognized that some AI interactions don't require explicit notice.
If you're using an AI system labeled "AI Assistant" with a robot icon and synthetic voice, adding a redundant disclosure ("You are now interacting with an AI system") doesn't protect users. It just clutters the interface. The regulation's "reasonably well-informed" standard assumes users can recognize AI when the context makes it clear.
This approach also preserves user experience in scenarios where AI augments human work rather than replacing it. Consider a support agent who uses AI to suggest responses but reviews and sends them personally. The user is interacting with a human who happens to use AI tools. Disclosing "This response was generated with AI assistance" in every message undermines the human relationship you're trying to maintain.
Context-dependent disclosure lets you focus transparency efforts where they matter: situations with genuine ambiguity. A voice system that sounds human should disclose. A chatbot that mimics conversational patterns should disclose. But an interface clearly branded as automated doesn't need redundant warnings.
The regulation's law enforcement carve-out reinforces this reading. Systems used to detect, prevent, or investigate criminal offenses are exempt from disclosure requirements (with safeguards). This suggests the rule targets scenarios where users might reasonably misunderstand who or what they're engaging with, not every technical instance of AI use.
Where practitioners actually land
Most organizations we've consulted are adopting a hybrid approach: default to disclosure, but document your obviousness assessments for specific scenarios.
This means building disclosure into your standard AI interaction patterns (chatbots, voice systems, content generators) as the baseline. Then, for deployments where you believe obviousness applies, you document why: the interface design, the branding, the user context, the typical user sophistication level.
You're not skipping disclosure everywhere. You're creating a defensible record of where and why you determined disclosure wasn't required under the regulation's exception. If challenged, you can show you assessed the standard, not ignored it.
This approach also acknowledges that "reasonably well-informed" is a moving target. Users in 2024 have different AI literacy than users in 2019. What seemed novel then is routine now. Your obviousness assessment should reflect current user expectations, not historical assumptions.
For machine-readable watermarking under Article 50 § 2, there's less room for interpretation. If your system generates synthetic content, you mark it. The Digital Omnibus on AI provides a grace period until December 2 for systems already on the market before August 2, but the obligation itself is clear.
Our take
Disclose by default, but document your exceptions rigorously.
The compliance risk of over-disclosing is minimal: you might annoy some users, but you won't face penalties. The risk of under-disclosing is significant: you're betting that regulators, users, and courts will agree with your obviousness assessment after a dispute arises.
That said, universal disclosure without judgment is lazy compliance. The regulation includes an obviousness exception for a reason. If you're running an interface clearly branded as AI, with synthetic voice, robot iconography, and explicit "AI Assistant" labeling, adding a redundant disclosure doesn't serve users or the regulation's intent.
Build disclosure into your standard patterns. Then, for scenarios where you believe obviousness applies, document your reasoning: screenshots of the interface, user research showing awareness, design elements that signal AI use, the typical user's technical sophistication. Treat it like any other compliance judgment call, with evidence and rationale.
And recognize that this decision compounds with other Article 50 obligations. If you're generating synthetic content, you're marking it under § 2. If you're deploying emotion recognition or biometric categorization, you're informing users under § 3. If you're creating deepfakes, you're disclosing under § 4. Your transparency approach should be consistent across these requirements, not fragmented by article number.
The August 2 deadline doesn't leave room for extended deliberation. Pick your approach, document your logic, and implement it. You can refine based on enforcement guidance and user feedback. But you can't wait for perfect clarity that isn't coming.
