Skip to content

Article 50

Five ways Article 50 disclosure fails

On 20 July 2026, the European Commission adopted its Guidelines on Article 50 of the EU AI Act, the transparency obligations that apply from 2 August 2026. Paragraph 38 names five techniques that are not, by themselves, enough to satisfy the disclosure duty. All five are interface decisions. None of them can be fixed by a lawyer or an engineer alone.

Here’s what each one gets wrong, and the pattern that addresses it.

Version 1.0 · Published · Updated

  1. 1. Terms and conditions are not disclosure

    Putting the AI notice only in your terms of service, a linked policy page, or product documentation doesn’t satisfy the obligation on its own. The Guidelines are specific that this kind of disclosure can complement in-context disclosure, but it can’t replace it.

    What to build instead: AI Presence Disclosure: disclosure that lives inside the interaction itself, not one click away from it.

  2. 2. An invisible watermark is not disclosure

    Machine-readable marking (metadata, a watermark) that a person can’t perceive at the point of interaction doesn’t satisfy the disclosure duty either. That doesn’t make the marking pointless: it’s still the right tool for the separate Article 50(2) obligation to mark synthetic content. It’s just not a substitute for something a person can actually see or hear when it matters.

    What to build instead: AI-Generated Content Marking: pair a visible label with embedded metadata. Treat the visible label as the layer that has to hold up on its own, because metadata is routinely stripped by compression, format conversion, or re-upload.

  3. 3. “Assistant” is not disclosure, and neither is a friendly face

    Vague signals like calling something an “assistant” don’t satisfy the obligation. Neither does a human-like representation that could mislead someone about what they’re talking to: a human name, a human profile photo, a natural-sounding voice. The Guidelines treat this as a live design factor: elsewhere in the same document, interface choices like these are named as things that can determine whether an AI system even qualifies for the narrow “obvious interaction” exception in the first place.

    What to build instead: this is where AI Presence Disclosure does its job, but it only works if nothing else on the same screen quietly undermines it. A disclosure label doesn’t survive next to a human avatar.

  4. 4. “This site uses AI” is not disclosure for what this feature does

    A general notice covering a whole platform isn’t specific enough if that platform offers several different services or the AI touches only some interactions. The Commission’s own example: a site that offers a mix of services can’t rely on “services on this website use AI” to cover all of them.

    What to build instead: disclosure scoped to the specific system and interaction a person is actually using, not a blanket statement at the platform level. Part of the same AI Presence Disclosure pattern, applied per-surface rather than once, site-wide.

  5. 5. “This uses LLMs” doesn’t tell anyone anything

    A technical description of the underlying technology, without explaining what that means for the person using it, doesn’t satisfy the obligation. The bar the Guidelines set has three parts: what the system does, what that means for you, and that it’s artificial.

    What to build instead: Model Transparency Cards for the capability and limitation side, paired with AI Presence Disclosure for the artificial-origin side. A technology name is not a substitute for either.

One thing the Guidelines flagged that most pattern libraries haven’t caught up to

Item 3 above is really two different failure modes bundled together: vague wording, and design that actively works against the disclosure sitting right next to it. The wording problem has a pattern. The anthropomorphism problem (avatar design, voice design, choosing a human name for a system) mostly doesn’t, in this library or the major public ones. It’s a real gap, not a solved problem we’re just linking to.

This reflects the Commission’s Guidelines on Article 50, adopted 20 July 2026, and is not legal advice. The Guidelines are non-binding interpretive guidance; the Regulation and any future court interpretation take precedence. Source: European Commission, Guidelines on transparency obligations for providers and deployers of AI systems

Disclosure is the first step. Explanation is the harder one.

Article 50 asks you to tell people they are dealing with AI. From 2 December 2027, high-risk systems in areas like public benefits, employment and credit must also be interpretable for the people who use them, and people affected by a decision gain a right to an explanation. That is where design gets difficult, and where this project focuses.

Article-by-article details