A buyer asks, “Does your product label AI-generated content?” Ask them which output they mean. Article 50 of the EU AI Act gives the provider of a generative AI system a duty to mark certain outputs so a machine can detect their origin. It also requires deployers to disclose certain uses to people. A customer can need both a file that carries a detectable mark and a page that shows a reader a disclosure. European Commission guidance on Article 50
Which systems face the December deadline?
Article 50 generally applies from 2 August 2026. Article 111(4) gives providers of covered generative AI systems placed on the EU market before 2 August 2026 until 2 December 2026 to meet the output-marking obligation in Article 50(2). That transition is specific to provider marking. For a new system, establish when you first placed the system on the market. The release date of the underlying model answers a different question. EU AI Act, Articles 50, 111(4) and 113
Match the output to the duty
Start with one feature in the product your customer will buy. Identify who provides the AI system, who uses it and who encounters the result. A provider that builds a feature into its own product and a customer that publishes the resulting content have different decisions to make. These are three common cases for software teams. European Commission guidance on Article 50
| Situation | Who acts | What to inspect |
|---|---|---|
| Direct interaction with an AI assistant, Article 50(1) | The system provider informs the user they are interacting with AI unless that is obvious in context. | Open the first screen in the customer’s embedded assistant. |
| Synthetic text, images, audio or video, Article 50(2) | The system provider marks covered output in a machine-readable, detectable form. | Use the documented detection method on a sample export. |
| Deepfakes or AI text published to inform the public on matters of public interest, Article 50(4) | The deployer discloses the AI origin, subject to the applicable exceptions. | Read, watch or listen to the published item as a member of the audience. |
Assess exceptions at the feature level. Article 50(2) provides an exception for standard editing assistance. For the public-interest text exception under Article 50(4), check both conditions. The content must have undergone human review or editorial control, and a person or entity must hold editorial responsibility for its publication. EU AI Act, Article 50
For information due to a person, Article 50(5) calls for a clear, distinguishable disclosure by the first interaction or exposure, with applicable accessibility requirements. For a visible notice, test the mobile and screen-reader experience. For an audio experience, check what a listener hears before relying on a notice in the player interface. EU AI Act, Article 50
Follow one draft all the way to publication
Suppose your software drafts a private support reply. A customer also uses it to draft public notices about interruptions to a city water supply. Both features can use the same model and generate text. The support reply goes to one person about their account. The water notice goes to the public about a matter that affects them. The public-interest publication question arises for the second use. Assess provider marking for the generative system separately. European Commission guidance on Article 50
For the water notice, ask your customer to demonstrate a real publishing path with a synthetic incident. Export the draft. Paste it into the publishing system. Inspect the final page and its downloadable copy. If the customer relies on an editorial review, identify the editor, the version they reviewed and the entity responsible for publication. If the publishing system removes a mark or hides a disclosure, the test has found the exact handoff to investigate.
Run the same exercise for the route that customers actually use. An engineer might test a rich-text export while an editor uses a plain-text copy button. Give the reviewer both results and identify the output format, software version and date of each test.
What belongs in the enterprise evidence pack?
Keep an example of the delivered output and the result of the marking or disclosure test. Pair that with evidence of the inference controls your customer requested. With Alectura you can show which customer policy applied, which processing region the request used and what the guardrails found. A customer-specific evidence pack covers that part of the review. Demonstrate the output mark with the detector, and the visible disclosure in the product or publication where the audience encounters it. Explore Alectura’s customer evidence packs
Source documents (2)
Trace a request through Alectura.
Bring a sample prompt and a customer policy. See what reaches the model and which decisions appear in the request record.
