EU AI Act obligations for AI-generated assets

What Article 50 requires

Article 50 of the EU AI Act, Regulation (EU) 2024/1689, sets four transparency duties, and two of them reach the assets in a DAM.

This page is orientation for DAM selection teams. It gives no legal advice: confirm how Article 50 applies to your organisation with legal counsel.

Fact: Article 50(5) requires the information to reach people at the latest at the first interaction or exposure, in a clear and distinguishable way that meets accessibility requirements. Fact: Article 50(2) excludes AI systems that perform an assistive function for standard editing or that do not substantially alter the input data.

Application dates and the Digital Omnibus on AI

The Article 50 application date of 2 August 2026 stayed in place, and the Digital Omnibus on AI added a four-month grace period for marking by systems already on the market.

The Digital Omnibus on AI is adopted law: Regulation (EU) 2026/1744 of 8 July 2026, published in the Official Journal on 24 July 2026 and in force since 27 July 2026. The table lists each date with its status as published.

Regulation (EU) 2026/1744 also replaced Article 50(7) on codes of practice for detecting and labelling artificial content. DAMSelection found no further proposed amendment to Article 50 on the EUR-Lex and Commission pages checked on 14 October 2026.

Guidance that sits next to the law

The Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026. Commission guidelines C(2026) 5054 of 20 July 2026 explain Article 50 and state that they are non-binding. The EU icons for labelling AI content are optional, while the Article 50 labelling duties are binding.

Provider or deployer: where a brand sits

A brand that uses generative AI to make marketing assets is usually a deployer, and the company that offers the generative system under its own name is the provider.

The Commission guidelines define a deployer as an entity that uses an AI system under its authority for professional purposes. A company stays the deployer when contractors operate the system on its behalf. A company that only commissions an agency, without control over the agency's AI use, is no deployer under paragraph 14 of the guidelines.

DAMSelection assessment: a DAM vendor that ships image generation under its own brand acts as provider of that feature and carries the Article 50(2) marking duty. The brand that publishes the result carries the Article 50(4) disclosure duty where the image is a deepfake. Ask each vendor in writing which role it claims for each AI feature.

Which marketing assets need a label

Article 50(4) disclosure applies to deepfakes, meaning image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear genuine to a viewer.

The Commission FAQ notes that context and audience expectations shape that assessment. The guidelines list standard editing that needs no marking, such as minor cropping, format conversion and spelling correction. The table gives worked examples, labelled as DAMSelection assessment, for discussion with counsel.

The B2B exemption in paragraph 87 of the guidelines needs three cumulative conditions. The output is strictly technical, a limited group of professionals sees it, and it stays inside the companies involved.

What a DAM must record

A DAM must record, per asset, whether AI generated or edited it, which system did so, and which disclosure the asset needs in each channel.

Store these as structured fields in the metadata schema, so publishing rules can act on them. The Metadata Model Builder helps place the fields next to rights and product data.

• AI involvement, using the IPTC Digital Source Type values such as trainedAlgorithmicMedia and compositeWithTrainedAlgorithmicMedia.
• Generating system, version and provider, with the date of generation.
• Machine-readable mark present on the master file, such as a C2PA manifest or a detected watermark.
• Deepfake assessment outcome, the person who made it and the date.
• Disclosure required per channel, with the label text or icon variant to use.
• Human review and editorial responsibility, for AI-generated text published on matters of public interest.
• Link to the source assets and prompts used, held as related assets.

DAMSelection assessment: make the AI involvement field mandatory on upload. A missing value blocks publication until someone answers it.

How a DAM keeps marks and labels intact

A DAM keeps Article 50 compliance intact when renditions, exports and delivery URLs carry the marking and the disclosure from the master file.

Machine-readable marks often live in embedded metadata. Rendition engines, image CDNs and social exports can strip that metadata unless configured to keep it. Read content credentials and C2PA in DAM for how signed provenance survives resizing.

1. Keep the provider's mark on the master and on every rendition the DAM generates.
2. Add the visible disclosure as a channel rule, because the guidelines state that deployers cannot rely on the embedded mark alone.
3. Block downloads of deepfake assets without the disclosure variant.
4. Log every publication of a labelled asset, with channel and date.

Checklist for DAM owners

DAM owners can close most Article 50 gaps in four to eight weeks with an inventory, a schema change and two publishing rules.

The checklist below is a DAMSelection rule of thumb from PIMvendors practice. Run it with brand, legal and the channel owners in one working session per step.

1. Inventory every generative AI tool in use, inside and outside the DAM, and record provider, version and contract owner.
2. Ask each provider how its outputs carry the Article 50(2) mark, and whether the 2 December 2026 grace period applies.
3. Add the AI fields to the schema and backfill assets created since 2 August 2026 first.
4. Agree a deepfake assessment rule with legal, with three or four worked examples from your own library.
5. Configure channel rules for visible disclosure, using the optional EU icons where they fit the brand.
6. Test renditions and exports for mark survival, then repeat the test after every DAM or CDN release.
7. Review the log each quarter and report labelled assets per channel to the compliance owner.

DAMSelection assessment: retroactive labelling of content created before 2 August 2026 is optional under the Commission FAQ, which still encourages it. Start with assets that remain in active use on public channels.

Questions for vendors and demo tests

Vendors should demonstrate AI labelling on the buyer's own generated assets in a scripted scenario, because generic demos hide where marks get lost.

Add these scenarios to your scripted DAM demo or build them in the Demo Script Generator.

1. Upload an image generated by an external tool and show which AI fields the DAM fills automatically.
2. Generate a variant inside the DAM and show the machine-readable mark on the output.
3. Create a web rendition and a social crop, then show that the mark survives both.
4. Publish a deepfake asset to the website and show that the disclosure label appears.
5. Export a report of all AI-generated assets published in the last 30 days, by channel.

Score the answers under security, hosting and compliance in your scoring matrix. The wider AI picture sits in AI in DAM, and the full list of Learn guides sits on the DAM Learn pillar. Related terms: content credentials and watermarking.