Synthetic content workflows should start with an inventory of output types and channels, assign marking and disclosure steps to named roles at each stage of production and publication, and include review checks before content is shared. The workflow should ensure that machine-readable marking is applied by the provider, visible disclosure is applied by the deployer where required, and both travel with the content as far as possible, with records kept for each item or batch.
How should content be inventoried?
A workflow should start with an inventory of synthetic content: the systems that generate it, the output types, the formats, the channels, and the audiences. The inventory should distinguish between content generated for internal use, content published to inform the public, content that is a deepfake under Article 50(4), and content that is assistive editing outside Article 50(2). Each category should have a named owner and a workflow that matches its duties. An inventory that lists only the main publishing channel, while content is also shared through social media, email, or syndication, will miss the channels where marking and disclosure are most likely to be lost. The inventory should be reviewed when a new system, output type, or channel is added, and should be linked to the provider and deployer records for the systems involved.
Who does what in the workflow?
The workflow should assign each step to a named role: the provider role for machine-readable marking, the production role for applying the mark and checking it, the editorial role for visible disclosure and human review, and the publishing role for checking that disclosure travels with the content. Handover points should be explicit, so that content does not move from production to publication without the marking and disclosure checks having been carried out. A checklist at each handover, signed by the person who carried it out, creates a record that the step was done for that item or batch. Where a team uses an external agency or a platform, the workflow should state who is responsible for each step and how the external party's work is checked, because outsourcing production does not outsource the legal duty.
For the broader regulatory framework, see Transparency duties for AI.
How should review and approval work?
Review and approval should check, before publication, that the content has been marked where required, that visible disclosure is present and correct, that any human review exception is properly recorded, and that the content does not fall within a prohibition. The review should be proportionate to the risk: content that resembles real people or events, or that concerns matters of public interest, needs a more careful review than clearly fictional content for internal use. The reviewer should be competent to assess the content, the disclosure, and the exception relied on, and should record the decision, the date, and any condition. A review that approves content without checking marking and disclosure, or that records only a general approval, does not evidence that the Article 50 checks were carried out for that item.
How should sharing and syndication be handled?
Sharing and syndication are where marking and disclosure are most easily lost. The workflow should include steps to ensure that the mark and the visible disclosure travel with the content when it is shared, downloaded, embedded, or syndicated, as far as the deployer or publisher can ensure. That can include using formats that preserve metadata, attaching disclosure to the content file itself, and including disclosure in the caption or accompanying text where the platform allows it. The workflow should also include a check after sharing, on the channels where the content appears, to confirm that the mark and disclosure remain present. Where a platform strips metadata or disclosure, the limitation should be recorded, the alternative disclosure used should be noted, and the risk should be assessed, rather than assuming the original disclosure still applies.
How should records be kept for content?
Records should be kept for each item or batch, showing the system and version that generated the content, the marking method applied, the disclosure given, the review carried out, and the publication details. Records should be retrievable by content identifier, date, or channel, so that a question about a specific item can be answered from the file for that item. Retention should follow the period the organisation sets for transparency records, taking into account the need to answer authority or audience questions after publication. A record system that keeps only aggregate counts, without item-level detail, cannot show that a particular item was marked and disclosed correctly. The record system should be tested by sampling items and tracing them from generation to publication and sharing.
Continue with Article 50 marking and detection for synthetic content.
What does this guide not decide?
This guide does not decide whether a particular workflow meets Article 50, whether a specific item was correctly marked or disclosed, or whether a human review exception applies. It does not design a workflow for a specific team or platform. Those questions depend on the content, the systems, the channels, and the consolidated legal text. A team should map its own workflow, assign named roles, test the workflow end to end, and confirm the position with a competent authority or legal adviser where the content is high-risk or the workflow spans several parties.
Frequently asked questions
Who is responsible for marking: provider or deployer?
Machine-readable marking under Article 50(2) is a provider duty, while visible disclosure for deepfakes and public-interest text under Article 50(4) is a deployer duty. Both can be relevant to one item of content, and the workflow should assign each duty to the party that controls it, with handover checks between them.
What if a platform strips the marking?
Record the limitation, use the alternative disclosure the platform allows, and assess the risk. Check after sharing that disclosure remains present, and do not assume the original marking still applies where the platform is known to strip it. The limitation and the alternative should be in the content record.
Must every item have its own record?
Records should allow a specific item to be traced from generation to publication, with the marking, disclosure and review for that item. Batch records can be used where items share the same system, version and workflow, provided the batch can be identified and sampled, and high-risk items are recorded individually.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk Article 50 ↗European Commission AI Act overview ↗