On May 8, 2026, the European Commission published draft guidelines1 interpreting Article 50 of the EU AI Act, with feedback due by June 3 and enforcement of interaction-disclosure rules hitting August 2. The guidelines answer a question that has haunted engineering teams since the Act passed: when is an AI system “obvious” enough that you can skip the disclosure? The Commission’s answer borrows the “average consumer” test from EU consumer protection law, then immediately complicates it with a multi-factor analysis that defaults toward more disclosure the moment vulnerable users or low AI literacy enter the picture. Less than 24 hours earlier, the Council and Parliament had provisionally agreed2 on the Digital Omnibus, deferring Article 50(2) watermarking obligations to December 2, 2026. The result is a split deadline inside the same article: interaction disclosures must ship in under three months, while synthetic-content marking gets a partial reprieve.
What Just Changed: The May 8 Draft Guidelines and the May 7 Omnibus Split
Article 50 contains four distinct transparency obligations. Paragraphs 50(1) (chatbot and AI interaction disclosure), 50(3) (emotion recognition and biometric categorisation), and 50(4) (deepfake disclosure) become applicable on August 2, 20263. Paragraph 50(2), which governs synthetic-content watermarking for generative AI outputs, was deferred to December 2, 20262 by the May 7 Omnibus deal, but only for systems already placed on the EU market before August 2. New systems launched from August 2 onward must comply with watermarking immediately.
That four-month gap is not a holiday. It is a scheduling headache. Engineering teams now face two effective dates for transparency work that sits in the same architectural layer: UI labels and interaction notices must be live by August 2, while content provenance and watermarking infrastructure can wait until December only if the product is already shipping. A new feature rollout in September needs both from day one.
Since May 8: The Code of Practice Lands and Parliament Backs the Omnibus
Two things moved between the draft guidelines and August. First, the consultation closed on June 3 as scheduled, and as of late June the Commission has still not published the final Article 50 guidelines [Updated June 2026]. The draft remains the only interpretive text teams can build against, and the August 2 date does not move with it. Second, on June 10 the Commission finalised a separate instrument: the Code of Practice on marking and labelling of AI-generated content. It is voluntary, it covers Article 50(2), (4), and (5), and it is the closest thing to a safe harbour the regime offers. Signing it buys a presumption of conformity with the marking and labelling obligations, and the deadline to make the AI Office’s initial signatory list is July 22 at 18:00 CEST. Miss it and you are not automatically non-compliant, but you carry the burden of proving adequacy to a market surveillance authority on your own terms.
The distinction between the Code and the guidelines matters more than the names suggest. The guidelines interpret what the law requires, including the obviousness exemption and the average-consumer test. The Code does not touch Article 50(1) interaction disclosure at all; it is a marking-and-labelling framework for synthetic content. A chatbot vendor reading the Code for chatbot-disclosure guidance will find nothing useful, because that obligation lives in the guidelines and the statute, not the Code.
On the Omnibus, the deferral is closer to settled than the May 7 “provisional agreement” framing implied. The European Parliament formally endorsed the deal on June 16, with Council adoption and publication in the Official Journal expected in July, ahead of August 2 [Updated June 2026]. The December 2 watermarking reprieve for pre-August systems is therefore more likely to hold than not, though it is not binding until the Official Journal carries it.
What “Machine-Readable Marking” Actually Demands
Article 50(2) is not satisfied by a visible “AI-generated” caption. It requires that outputs be marked in a machine-readable format and detectable as artificially generated or manipulated, which in practice points at provenance metadata and watermarking. The Code leans on standards work like C2PA content credentials for the metadata side, but it does not mandate a single scheme, and it acknowledges the obvious failure mode: a mark that survives a screenshot, a re-encode, or a crop is still an open research problem. Groundy has documented why dataset watermarks fail to survive fine-tuning and downstream transforms, and the same fragility undercuts a compliance story built on detection. A provider can ship a conformant marker on day one and watch it get stripped by the first social platform that re-compresses the image. The legal obligation is to embed the marker and to make detection tools available, not to guarantee the mark is unremovable, which is the only version of the requirement that is technically honest.
Enforcement Readiness: What August 2 Actually Exposes
Transparency breaches sit in the AI Act’s lower fine tier, but “lower” still means up to EUR 15 million or 3% of worldwide annual turnover under Article 99, whichever is higher. Enforcement runs through national market surveillance authorities, not the Commission directly, so the first real test of the average-consumer standard will come from whichever member-state regulator opens a case, not from a single uniform EU position. For most teams the realistic exposure between August 2 and any first enforcement action is reputational and contractual, since procurement questionnaires already ask for AI Act conformance, rather than an immediate fine. The paper trail from the obviousness analysis is what a market surveillance authority requests first, which is why documenting the multi-factor assessment beats polishing the label.
The Average Consumer Test: How the Obviousness Exemption Actually Works
Article 50(1) contains an exemption: you do not need to disclose that the user is interacting with AI if it is “already obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect”4. The draft guidelines import this “average consumer” standard3 directly from EU consumer protection jurisprudence. It is a legal standard, not a statistical one. You cannot A/B-test your way to compliance. The Act’s other transparency provisions are being read just as literally; Groundy looked at whether word-subset explanations satisfy the AI Act’s transparency rule, and the answer there was similarly unforgiving of minimum-effort compliance.
The guidelines impose a two-step assessment3. Step one: is the AI nature obvious to the average consumer? Step two: even if it is obvious to that hypothetical reasonable person, does the actual audience include vulnerable groups or users with low AI literacy? If the second test triggers, the exemption collapses and disclosure is required. The burden is on the deployer to document why the exemption applies, not on the regulator to prove it does not.
What Counts as Obvious (and What Definitely Doesn’t)
The guidelines provide concrete examples, and the pattern is instructive. Clearly obvious cases3 include “AI-powered code assistance chatbots available only to professional developers” and “AI-enabled Non-Playable Characters (NPCs) in video games.” Both have contextual guardrails: the developer knows they are in a coding tool; the gamer knows NPCs are scripted entities.
Not obvious3: “AI-enabled robotic companion pets designed to mimic natural human-pet interaction,” “AI avatars in immersive environments,” and “chatbots embedded in online helpdesks.” The common thread is anthropomorphic or embedded interaction where the AI is disguised as something else. A user messaging a helpdesk expects a human agent or, at minimum, is not primed to assume automation. A user stroking a robotic cat is not in a frame of mind to inspect for silicon.
The guidelines are equally explicit about what does not satisfy the “clear and distinguishable” requirement when disclosure is needed. Information buried in terms and conditions, manuals, or layered menu options fails3. The disclosure must be “noticeable and easy to understand” and “easy to identify as separate from other information.” If your “AI assistant” label sits inside a settings submenu three taps deep, you are not compliant. You are collecting liability.
The Vulnerable-Group Multiplier: Why the Test Defaults to Disclosure
The multi-factor test3 explicitly weighs three variables: the target audience, potential exposure of vulnerable groups (children, elderly persons, persons with disabilities), and the level of AI and digital literacy among intended users. Any plausible exposure to vulnerable users lowers the threshold for required disclosure. For most consumer-facing products, this means the exemption is practically closed.
The Commission is aware of the tension. The obviousness exemption has drawn criticism5 for risking weakened user protection, particularly for AI companions that form emotional bonds with users. The draft guidelines respond by narrowing the exemption through the vulnerable-group lens rather than removing it. The result is a safe harbour that exists on paper but rarely survives contact with a real user base.
The Dual Deadline Trap: August 2 vs December 2
The Omnibus deferral of Article 50(2) to December 2, 2026 applies only to generative AI systems already on the EU market before August 2, 20262. Systems placed on the market from August 2 onward must comply with watermarking immediately. This creates three compliance buckets:
- Legacy systems, live before August 2: interaction disclosures due August 2; watermarking due December 2.
- New launches from August 2 onward: both interaction disclosures and watermarking due immediately.
- Updates or feature additions that constitute a new “placing on the market”: the December reprieve may not apply; legal review is needed.
The overlap is awkward. Article 50(1) and 50(2) are not independent silos. A chatbot that generates images in the same conversation thread may need interaction disclosure for the chat interface and synthetic-content marking for the images, with different deadlines depending on when the system first shipped. The Commission has not published guidance on how to treat feature additions or model upgrades that change the output modality of an existing product.
Actionable Checklist for Engineering Teams
With the June 3 consultation closing date looming and August 2 enforcement approaching, the practical task list is short and specific:
Audit every AI touchpoint in your EU-facing product. Map which paragraphs of Article 50 apply: 50(1) for chatbots and AI assistants, 50(3) for emotion recognition, 50(4) for deepfakes or synthetic media. Do not assume one label covers all four.
Run the obviousness exemption analysis in writing. Document the target audience, vulnerable-group exposure, and AI literacy assessment. If any factor points toward disclosure, stop there. The analysis itself is your defence in an enforcement action.
Move disclosures to the surface. UI labels must sit at the point of interaction, not in onboarding flows, T&Cs, or settings menus. “Noticeable and easy to understand” means visible without interaction, distinct from surrounding text, and in the user’s language.
Split watermarking from interaction disclosure roadmaps. If your system generates synthetic content, confirm whether the December 2 deferral applies to your specific product version. If you are launching after August 2, budget for immediate watermarking infrastructure.
Review generic labels. Terms like “assistant” or “uses LLMs” are not self-evidently AI to the average consumer. The guidelines reject technical references that do not signal artificiality to a non-specialist. “AI assistant” may suffice; “assistant” does not.
The consultation window has closed; sign the Code if marking applies. [Updated June 2026] Feedback on the draft guidelines closed June 3, so the moment to surface edge cases on immersive environments, companion products, and embedded chat has passed. The live deadline now is July 22: if your system falls under Article 50(2) or 50(4), decide whether to sign the Code of Practice on marking and labelling before the initial signatory list closes, and track the final guidelines, which were still unpublished as of late June.
The Commission has given deployers less than three months to parse a legal standard, redesign UI flows, and ship before August 2. The guidelines are clearer than the statute, but they are not generous. The obviousness exemption is narrower than most provider legal summaries suggest. For engineering teams, the safest assumption is that disclosure is the default, the exemption is the exception, and the burden of proof is yours.
Frequently Asked Questions
Does Article 50(1) disclosure apply to internal enterprise tools used only by trained employees?
Yes, the AI Act’s transparency obligations are not limited to consumer-facing deployments. However, the obviousness exemption is far easier to satisfy when the user base consists of professionals who have been trained on the tool. The “average consumer” benchmark originates from the Unfair Commercial Practices Directive, which historically targets B2C relationships, but the AI Act does not carve out internal B2B use from Article 50.
How does the EU average-consumer test compare to US AI disclosure requirements?
The US has no single horizontal standard equivalent to Article 50. Disclosure obligations are fragmented: California’s SB 1001 requires bot disclosure only in commercial transactions, the FTC enforces against deceptive AI on a case-by-case basis, and state laws such as Connecticut’s SB 5 add their own AI provenance and chatbot-disclosure rules. The EU’s approach replaces that patchwork with one legal benchmark that applies regardless of industry.
What happens to the December 2 watermarking deferral if the Omnibus deal isn’t formally adopted?
The May 7 Omnibus was a provisional political deal that still required formal adoption. That risk has narrowed: the European Parliament endorsed the agreement on June 16, and Council adoption plus publication in the Official Journal is expected in July, ahead of August 2 [Updated June 2026]. The deferral is not binding until the Official Journal carries it, so it remains technically possible that the agreement slips and all four transparency obligations, interaction disclosure, watermarking, emotion-recognition notice, and deepfake labelling, hit simultaneously on August 2 with no phased rollout. After the Parliament vote, that outcome is unlikely rather than open.
How do the guidelines handle AI systems that hand off between automated and human operation mid-conversation?
They don’t. The draft guidelines treat “interacting with an AI system” as a binary state, with no guidance on hybrid sessions where an AI chatbot transfers to a live agent (or vice versa). This creates ambiguity about whether disclosure must appear at session start, persist after human takeover, or re-trigger when automation re-engages. Teams running blended customer-service queues should disclose at session start and treat the entire session as AI-mediated until the handoff is explicitly surfaced to the user.
Will final Article 50 guidelines be published before the August 2 enforcement date?
As of late June 2026 the final guidelines were still unpublished, even though the consultation closed June 3 [Updated June 2026]. A final version is expected before August 2, but the Commission has not committed to a firm date, so enforcement may begin while only draft guidance exists. The separate Code of Practice on marking and labelling was finalised on June 10, but it does not cover Article 50(1) interaction disclosure, so the obviousness exemption and the average-consumer test still rest on the May 8 draft alone. Teams should implement against the current draft rather than waiting for final wording: the legal obligation in the statute is already binding, and the guidelines merely interpret it.
