From 2 August 2026, transparency obligations under Article 50 of the EU AI Act begin to apply. The European Commission published implementation guidelines on 20 July, turning an abstract policy date into a practical consumer question: what will a person actually see when interacting with AI or encountering generated content?
The answer is not one universal badge. Providers of certain interactive systems must inform people that they are interacting with AI. Providers of systems that generate or manipulate content must support machine-readable marking in relevant cases. Deployers have disclosure duties for deepfakes, emotion recognition or biometric categorisation, and certain AI-generated public-interest text without human review or editorial control. The exact duty depends on actor, content and context. A visible label helps with provenance; it does not prove that the content is true, lawful, safe or high quality.
> The short answer: Expect clearer notices around AI interaction and some generated or manipulated content, but do not treat a label—or the absence of one—as a complete authenticity test. Article 50 allocates transparency duties; it does not certify every output.
What is confirmed, and what is not
The Commission’s 20 July guidelines define scope for providers and deployers, while the accompanying quick facts explain current timing and transitional details. The final Code of Practice on marking and labelling is voluntary; organisations outside it still need equivalently adequate compliance where the law applies. The obligations are legal and technical, not a consumer product rating.
This article is a reader’s interpretation guide, not legal advice. Businesses should analyse their own role, location, system and content with qualified counsel. Users should retain the source, date and interface state when a disclosure matters, because labels and product implementations can change.
Article 50 visibility matrix
| Check | Evidence or current signal | Decision use |
|---|---|---|
| Direct AI interaction | The interface tells a person they are interacting with an AI system unless obvious from context | Notice expected |
| Machine-readable marking | Generated or manipulated outputs carry detectable technical provenance where required | Technical layer |
| Deepfake disclosure | AI-generated or manipulated image, audio or video resembling real people, objects, places or events is disclosed in scope | Visible context matters |
| Public-interest text | AI-generated or manipulated text on matters of public interest lacks human review or editorial control | Disclosure may apply |
| Emotion and biometric categorisation | People are informed when exposed to in-scope emotion recognition or biometric categorisation | High-sensitivity notice |
| Provider versus deployer | The party building the system and the party using it have different duties | Do not collapse roles |
| Pre-existing systems | Quick facts describe a grace period for some marking obligations for systems placed on the market before 2 August | Check transition |
| Label limitations | No label can prove truth, consent, copyright clearance, security or professional fitness | Keep verification |
A strong score in one row cannot cancel a hard failure in identity, safety, permission or recovery. Date every fact that can change and keep market, account, device or object configuration attached to the evidence.
Direct AI interaction
A chatbot, voice agent or embodied system may need a disclosure at an appropriate time. The notice should not be buried after the interaction has shaped a decision. For a user, the disclosure answers ‘what am I talking to?’ It does not reveal the complete model, training data, retrieval sources, memory, human escalation or data-retention policy. Ask those questions separately when consequence is high.
Decision use: Notice expected. Evidence to keep: The interface tells a person they are interacting with an AI system unless obvious from context.
Machine-readable marking
Machine-readable marking is designed for tools and platforms, not only the naked eye. It can support detection and downstream handling, but robustness depends on implementation, file transformation and the applicable obligation. A screenshot, crop, re-encode or transcription may lose signals. Users should preserve original files and source links when provenance is important rather than relying on a repost.
Decision use: Technical layer. Evidence to keep: Generated or manipulated outputs carry detectable technical provenance where required.
Deepfake disclosure
A deepfake label should alert the viewer that the apparent event or performance was generated or altered. It does not explain whether a real person consented, whether satire or art is involved, or which parts changed. Evaluate the publisher, original source, date and corroboration. For reputational or fraud risk, save the full context before reporting or sharing.
Decision use: Visible context matters. Evidence to keep: AI-generated or manipulated image, audio or video resembling real people, objects, places or events is disclosed in scope.
Public-interest text
The obligation distinguishes automated public-interest publishing from content under human review and editorial responsibility. A label does not make unreviewed text accurate; human review does not make it infallible. Readers should look for named editorial responsibility, sourcing, correction policy and evidence. Publishers should avoid vague ‘AI assisted’ notices that obscure who made the final decision.
Decision use: Disclosure may apply. Evidence to keep: AI-generated or manipulated text on matters of public interest lacks human review or editorial control.
Emotion and biometric categorisation
The disclosure tells people that a system is making a particular type of inference. It does not validate the inference or make the processing proportionate. In workplaces, events, retail or travel, ask what data is collected, the purpose, legal basis, retention, recipient and challenge route. A transparency notice cannot turn a prohibited or otherwise unlawful use into a lawful one.
Decision use: High-sensitivity notice. Evidence to keep: People are informed when exposed to in-scope emotion recognition or biometric categorisation.
Provider versus deployer
A foundation-model provider, application vendor, employer, publisher and customer may occupy different positions. The user-facing experience can involve several. If a disclosure is missing, identify who controls the interface or publication before attributing responsibility. For procurement, map each actor, contractual handoff and evidence owner rather than assuming the model company supplies every notice.
Decision use: Do not collapse roles. Evidence to keep: The party building the system and the party using it have different duties.
Pre-existing systems
A product’s launch date and later modification may affect timing. Users should not infer that every older system is exempt from every transparency duty. Organisations need the current legal text and guidelines for their exact facts. Record why a transition rule applies, the review date and what would make the conclusion change.
Decision use: Check transition. Evidence to keep: Quick facts describe a grace period for some marking obligations for systems placed on the market before 2 August.
Label limitations
Treat the label as one provenance signal. Verify consequential claims against primary sources; inspect financial, medical, legal or safety advice with qualified expertise; and maintain normal fraud controls. Conversely, an unlabelled item is not automatically human-made. Bad actors may ignore rules, and technical marks can be stripped. The user’s verification burden changes, but it does not disappear.
Decision use: Keep verification. Evidence to keep: No label can prove truth, consent, copyright clearance, security or professional fitness.
Put the framework into a real decision
Build a four-column visibility matrix for any AI-enabled product or publishing workflow. Column one names the interaction or output: chatbot, voice, image, audio, video, public-interest text, emotion recognition or biometric categorisation. Column two identifies provider and deployer. Column three records the user-facing notice and machine-readable mark. Column four records the remaining verification question.
Test the experience as a logged-out user, ordinary customer and administrator where relevant. Capture the date, market, language, interface and original file. If a notice appears only in terms of service, ask whether it reaches the person at the right time. If a technical mark exists, test whether supported downstream tools can detect it after normal publishing steps. Businesses should keep the evidence without storing personal source content unnecessarily. The objective is not a cosmetic badge; it is a reproducible account of how a person is informed.
A decision record another person can audit
Create a dated record containing the exact object, device, system or route; the source URL and access time; the person responsible for verification; and the condition that would change the recommendation. Preserve earlier observations instead of overwriting them when facts move. For product, policy and service claims, record the market and configuration. For private identifiers, store a redacted public copy and a controlled private copy.
End with one of four outcomes: proceed, wait, request evidence or decline. State the reason in one sentence, list the unresolved risk and name the next review date. This makes later performance review more honest: an editor can see what was knowable at publication rather than judging the decision with hindsight.
Action checklist
Classify the interaction or content type.
Identify provider and deployer separately.
Record the visible notice and its timing.
Check for machine-readable provenance where applicable.
Preserve original files and source URLs.
Keep human editorial responsibility visible.
Document transition assumptions.
Verify consequential claims beyond the label.
Related VERTU reading
These links cover adjacent decisions. They do not replace the current primary source or the exact configuration under review.
Sources and verification
Sources were accessed on 21 July 2026. Product availability, software features, laws, official alerts, policies, prices and service terms can change. Recheck the live source before acting. Legal, financial, insurance, health, travel-security and conservation references are general information rather than individual professional advice.
Final view
The 2 August change should make important AI interactions and content easier to recognise, not magically self-verifying. Use the notice to ask better questions: who generated this, who deployed it, what was marked, who reviewed it and what evidence supports it? Transparency is most useful when it leads to accountable verification rather than passive trust in an icon.
