An AI Label Request Arrived. Who Owns the Article 50 Evidence?
Route an EU AI Act Article 50 request by actor, system and publication: provider notice, provider machine-readable marking, deployer disclosure, exception and evidence owner.
Signals to watch
- A provider or deployer has an August 2026 release date but still describes the task only as adding an AI label
- Machine-readable marking, visible disclosure and direct-interaction notice have been assigned to one team without identifying the relevant Article 50 paragraph
- A named product, publication or campaign has an unresolved exception and no owner for the supporting evidence
Since 2 August 2026, an “AI label” request should not be scoped as one universal labelling task. Article 50 of the EU Artificial Intelligence Act assigns different work according to who acts, what the system does, what reaches a person and how it is published. A provider may own a direct-interaction notice or machine-readable output marking; a deployer may own disclosure of an emotion-recognition system, biometric-categorisation system, deepfake or certain public-interest text. Each route has its own conditions and exceptions.
An AI-governance consultancy business-development lead may see this in authorised provider, deployer and synthetic-media Telegram groups. The commercial Signal is a named release, a dated decision and an unresolved responsibility station. Seeing it one day late can mean missing the meeting that assigns the evidence.
What became applicable on 2 August 2026
Article 113 of Regulation (EU) 2024/1689 states that the Regulation generally applies from 2 August 2026, subject to its listed earlier and later dates. Article 50 is within that general application date. The European Commission’s Article 50 guidelines page, updated 6 August 2026, likewise says these transparency obligations apply from 2 August 2026.
Definition: In this article, an Article 50 transparency duty is a requirement to inform a person, mark an output in a machine-readable way or disclose that specified content was artificially generated or manipulated. It is not a general instruction to place identical visible text on every AI-assisted output. “Provider” and “deployer” are legal roles under the Regulation, not synonyms for software vendor and customer in every contract.
The date does not prove a particular request is in scope or a disclosure method has survived authority review. You still need the system, role, output, audience and publication context.
One release passes through four responsibility stations
Follow one release from system design to public exposure. At each station, ask who can produce the evidence.
Station 1: the provider owns the direct-interaction design
Article 50(1) addresses providers of AI systems intended to interact directly with natural persons. The provider must design and develop the system so the person is informed that they are interacting with AI, unless that is obvious to a reasonably well-informed, observant and circumspect person in the circumstances and context. The paragraph contains a bounded law-enforcement exception and a qualification where a system is available for the public to report a criminal offence.
The evidence owner is normally the provider controlling the interaction design. Record the interface, first interaction, notice method, accessibility treatment, system version and reasoning behind any “obvious” conclusion. A later deployer footer does not automatically prove the provider-side duty was addressed.
Station 2: the provider owns machine-readable marking
Article 50(2) addresses providers—including providers of general-purpose AI systems—whose systems generate synthetic audio, image, video or text. Outputs must be marked in a machine-readable format and be detectable as artificially generated or manipulated. The technical solution must be effective, interoperable, robust and reliable as far as technically feasible, taking account of content characteristics, implementation cost and the generally acknowledged state of the art.
This is not a visible deepfake caption. Record covered formats, mark location, export tests, limitations and approver. Article 50(2) also excludes, to the stated extent, systems assisting standard editing or not substantially altering the deployer’s input or its semantics, and contains a law-enforcement exception. “Standard editing” needs facts, not a feature name.
Station 3: the deployer owns disclosure at exposure or publication
Two different deployer routes sit here. Under Article 50(3), a deployer of an emotion-recognition or biometric-categorisation system must inform the natural persons exposed to its operation and process personal data under the applicable EU data-protection instruments. A bounded law-enforcement exception applies. The owner is the organisation operating the system in the actual setting, because it knows who is exposed and when.
Article 50(4) then addresses specified generated or manipulated content. A deployer using AI to create or manipulate image, audio or video that constitutes a deepfake must disclose that artificial generation or manipulation. Article 3(60) defines a deepfake by reference to content resembling existing persons, objects, places, entities or events that would falsely appear authentic or truthful. Evidently artistic, creative, satirical, fictional or analogous works receive a limited disclosure formulation that must not hamper display or enjoyment; that is not a blanket exemption.
The same paragraph separately covers AI-generated or manipulated text published to inform the public on matters of public interest. It provides an exception where content has undergone human review or editorial control and a natural or legal person holds editorial responsibility, plus the stated law-enforcement exception. A glance at a draft is not enough evidence; record the review and responsible person or entity.
Station 4: exception and evidence approval
Article 50(5) requires the information in paragraphs 1–4 to be clear and distinguishable and provided no later than first interaction or exposure, while meeting applicable accessibility requirements. This final station joins legal interpretation to release evidence: which paragraph applies, which exception is claimed, what the person will receive, when they receive it and who signs off.
The Commission’s current Code of Practice on Transparency of AI-generated Content, updated 31 July 2026, covers Article 50(2), (4) and (5). The Commission page says adherence is voluntary and that the Commission and AI Board confirmed the code as an adequate voluntary tool to demonstrate compliance. It does not replace the Regulation or the Commission guidelines. Record whether the organisation is a signatory or uses another method; do not turn voluntary adherence into a claim that the code itself is the legal obligation.
The role–output–audience record
For each release, preserve one row for every distinct route:
| Field | Question the evidence must answer | Likely owner |
|---|---|---|
| Actor | Is this organisation provider, deployer or both for this system and use? | Legal and product owner |
| System and version | Which system produces the interaction or output? | Engineering or vendor owner |
| Trigger | Direct interaction, synthetic output, emotion/biometric exposure, deepfake or public-interest text? | Product and publishing owner |
| Audience and timing | Who first interacts with or is exposed to it, and when? | UX, operations or publisher |
| Method | Visible notice, machine-readable mark or publication disclosure? | Engineering, UX or editorial owner |
| Exception | Which exact condition is relied on, and what facts support it? | Legal owner with technical/editorial evidence |
| Test and approval | What was tested, on which date, with which limitation and approver? | Named release-evidence owner |
This handoff is not a new legal test. It prevents one team accepting another actor’s work.
Example: “Can you add the EU AI label before Friday?”
Consider this composite Telegram thread, written for illustration and not taken from a customer or private group:
“Need the EU AI label sorted before Friday release. Model makes voice + copy.”
“Agency posts the clips. Product team owns the chat flow.”
“Some edits only. Legal asked whether the watermark survives export.”
The fragments omit the model, actors, audience, publication purpose, whether a clip is a deepfake, what “some edits” means and who holds editorial responsibility. They cannot support a legal conclusion or fixed-price proposal.
They are enough to ask a precise first question: “For Friday, are you scoping the provider’s direct-interaction notice, provider-side marking of generated voice and text, the agency’s disclosure for a particular publication, or all three?” The answer locates the responsibility station. A second question can then test the relevant exception and evidence owner. If the agency publishes ordinary generated marketing copy, for example, that fact alone does not make it public-interest text or a deepfake; if a clip portrays an existing person saying something they did not say, the deepfake route may need analysis.
The Article 4 literacy handoff deals with people and capability; the GPAI downstream-documentation handoff separates provider inputs from downstream evidence. Neither replaces Article 50 actor–output analysis. The pricing page explains product access.
TOP Prospect can filter, merge, deduplicate and rank these fragments from Telegram groups a user deliberately connects and is authorised to access, retaining original text, source, time, AI summary and ranking reasons for human review. It cannot decide legal roles, inspect the system, certify an exception, contact the poster or declare compliance. The Telegram business Signal workflow shows the discovery layer; the responsibility record remains a human-owned readiness decision.
Key facts
- Article 50 applies from 2 August 2026 under the general date in Article 113.
- Paragraph 1 addresses provider-designed notice for direct AI interaction; paragraph 2 addresses provider-side machine-readable marking of synthetic outputs.
- Paragraph 3 addresses deployer notice for emotion recognition and biometric categorisation; paragraph 4 addresses deployer disclosure for deepfakes and certain public-interest text.
- Paragraph 5 sets the clear, distinguishable, first-interaction-or-exposure and accessibility conditions.
- As checked on 13 August 2026, the Commission presents the transparency code as an adequate voluntary compliance tool for Article 50(2), (4) and (5), complemented by its Article 50 guidelines.
FAQ
Does Article 50 require the same label on every AI output?
No. The duty depends on the actor, system and use: direct AI interaction, provider-side machine-readable marking, emotion recognition or biometric categorisation, deepfakes, and certain public-interest text are treated differently, with specific exceptions.
Who owns machine-readable marking for synthetic content?
Article 50(2) places that duty on providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text, subject to technical-feasibility language and stated exceptions.
Who discloses a deepfake or AI-generated public-interest text?
Article 50(4) places those disclosure duties on the deployer. Deepfakes and public-interest text have different conditions and exceptions, so the deployer must classify the publication before choosing the disclosure.
What evidence should an Article 50 readiness request name?
Name the legal actor, system and version, interaction or output type, intended audience, publication context, applicable paragraph, exception analysis, disclosure or marking method, timing, accessibility check and accountable evidence owner.
Before quoting, make the requester choose the responsibility station and name the evidence owner. If those two fields remain blank, the honest deliverable is a bounded scoping review—not a promise to “make every AI output compliant.”
Frequently asked questions
Does Article 50 require the same label on every AI output?
No. The duty depends on the actor, system and use: direct AI interaction, provider-side machine-readable marking, emotion recognition or biometric categorisation, deepfakes, and certain public-interest text are treated differently, with specific exceptions.
Who owns machine-readable marking for synthetic content?
Article 50(2) places that duty on providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text, subject to technical-feasibility language and stated exceptions.
Who discloses a deepfake or AI-generated public-interest text?
Article 50(4) places those disclosure duties on the deployer. Deepfakes and public-interest text have different conditions and exceptions, so the deployer must classify the publication before choosing the disclosure.
What evidence should an Article 50 readiness request name?
Name the legal actor, system and version, interaction or output type, intended audience, publication context, applicable paragraph, exception analysis, disclosure or marking method, timing, accessibility check and accountable evidence owner.
Sources and further reading
- Regulation (EU) 2024/1689, Articles 3(60), 50 and 113, official text, accessed 13 August 2026
- European Commission, Guidelines on transparency obligations for providers and deployers of certain AI systems, updated 6 August 2026
- European Commission, Code of Practice on Transparency of AI-generated Content, updated 31 July 2026
How a Signal worth attention is found
See how Top Prospect finds and organizes Signals worth checking, keeps the original Telegram context, removes duplicates, and helps you decide what to review first. You decide whether to follow up and what to do next.
