Transparency obligations
Article 50 disclosure template and checklist
Article 50 creates disclosure and marking duties for particular AI interactions and outputs. The correct control depends on the system and the audience.
Prepared by EU AI Fit editorial team · Published 24 August 2026 · Source review 31 August 2026
Identify the trigger
Review whether people interact directly with AI, whether content is generated or manipulated, and whether the system performs emotion recognition or biometric categorisation. Consider any applicable exceptions and sector-specific duties.
Separate provider and deployer duties
Article 50 contains different duties for providers and deployers and for different output types. Record who creates or places the system on the market, who operates it, what the recipient experiences and whether a specific exception or qualification is relied upon. Do not reuse one disclosure conclusion across every channel.
Design effective disclosure
A disclosure should be clear, accessible and presented at the right time. Burying a generic statement in terms may not achieve meaningful transparency for a direct AI interaction.
Keep evidence of the control
Retain approved wording, interface captures, accessibility checks, machine-readable marking decisions, release records and monitoring evidence. Recheck disclosures when channels or output types change.
- Trigger and applicable paragraph
- Audience, channel and timing
- Approved visible disclosure wording
- Machine-readable marking decision
- Accessibility and interface evidence
- Owner, approval date and change trigger
In this scenario: Alder Customer Assistant
Alder Finance Ltd deploys a web assistant that answers product questions and transfers complex cases to staff. The opening interaction identifies that the user is dealing with AI, the wording is tested with assistive technology, the hand-off remains available and the approved interface capture is retained against the released version. A separate analysis covers any synthetic marketing content because that is a different trigger.
Test the disclosure after release
Verify the control on every relevant device and channel. Check that it appears at the required point, remains understandable, survives interface changes and is not obscured by consent banners or other notices. Monitoring should create evidence and a correction route, not merely confirm that wording exists in a design file.
Article 50 implementation record
Create a separate record for each trigger, channel and audience. The wording, timing, technical marking and responsible operator may differ even within one product.
| Trigger to examine | Implementation question | Evidence to retain |
|---|---|---|
| Direct AI interaction | Is the person informed that they are interacting with AI at the right time, unless this is obvious in the circumstances? | Approved wording, placement, first-interaction capture, accessibility result and release version |
| Synthetic output | Does the provider need machine-readable marking and can it remain detectable through the distribution chain? | Technical design, format tests, limitations, release record and monitoring result |
| Emotion recognition or biometric categorisation | Does the deployer need to inform exposed people and how are applicable data-protection duties addressed? | Audience notice, delivery evidence, system scope, privacy linkage and approval |
| Deepfake content | Is the deployer clearly disclosing that the content was artificially generated or manipulated? | Final disclosure, proximity to content, publication capture and any qualification relied upon |
| Public-interest text | Was the text published to inform the public on matters of public interest, and does an exception or editorial process apply? | Purpose assessment, human review evidence, disclosure decision, approver and source |
| Change and monitoring | Could a new channel, output type, audience or model make the implementation incomplete? | Named owner, test schedule, complaints, change triggers and reassessment history |
Questions teams usually ask
Does every chatbot need the same Article 50 disclosure?
No. The direct-interaction duty includes a qualification where the AI nature is obvious to a reasonably well-informed, observant and circumspect person. Other legal and operational transparency duties may still apply.
Is visible wording enough for synthetic AI content?
Not necessarily. Provider-side duties for certain synthetic outputs concern machine-readable marking, while deployer-side disclosure duties apply to specified uses such as deepfakes and some public-interest text. The responsible operator and applicable paragraph must be identified.
What evidence should be kept for an Article 50 disclosure?
Keep the trigger analysis, approved wording, placement and timing, interface or publication captures, accessibility review, technical-marking decision, release version, owner and change-triggered review history.
Recommended next step
Map each transparency trigger to the exact interface or output where the disclosure will appear, then assign evidence and an owner.
Run the free exposure checkRelated practical guides
Put the guidance into practice