Skip to content

[ Anolilab — Article / EU AI Act ]

Marking AI content under the EU AI Act by 2 December 2026

If you offer AI features, text, images, audio and video need machine-readable marks. Who is a provider, which techniques count, and a checklist.

Article by
Daniel Bannert · Founder · Anolilab

In short

If your product offers a feature under your own name that generates text, images, audio or video, you are usually a provider under Art. 50(2) of the AI Act, even if the model comes from OpenAI or Anthropic via an API. The output must carry a machine-readable mark and be detectable. For images, audio and video, the EU Code of Practice asks for two layers: signed metadata and an imperceptible watermark; for free-form text a watermark is enough. This has applied to new systems since 2 August 2026, and to systems already on the market before then from 2 December 2026. This article is not legal advice.

This article follows on from EU AI Act: what changes on 2 December 2026 for AI content, which lists all the dates. Here the focus is only on marking AI-generated content and what you have to do technically. It is based on three EU sources, all checked in November 2026: the legal text, the Commission's guidelines on Art. 50 and the Code of Practice on Transparency of AI-generated Content.

Which deadline applies to marking?

Article 50 of the AI Act has applied since 2 August 2026. The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, added a new paragraph 4 to Art. 111: providers of generative AI systems placed on the market before 2 August 2026 must comply with Art. 50(2) by 2 December 2026. Recital 38 calls this a four-month transitional period.

The transitional period covers only the machine-readable marking. If your system also has a chat, you have had to tell users they are talking to an AI since 2 August 2026. That is point 153 of the Commission guidelines (C(2026) 5054, adopted on 20 July 2026). Content created before 2 August 2026 does not need to be marked retroactively.

Am I a provider if I only use an API?

Usually, yes. Under Art. 3(3) of the AI Act, a provider develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge. The guidelines give the example of a company offering a generative AI application, such as a chatbot, an image generator or an AI agent, to customers under its own name (point 11). So if your SaaS has a "write it for me" button that calls GPT or Claude, you are the provider of the AI system behind that feature. OpenAI or Anthropic provide the model underneath.

You can be a deployer at the same time (point 15). That applies when you use the feature yourself and publish the output, for example AI images in your own marketing. The labeling duties under Art. 50(4) then come on top; more on those below.

The provider duty has two parts, and you need both: the output must carry a machine-readable mark, and there must be a way to detect that mark (points 69 and 70). A mark nobody can check is not enough.

Which content is out of scope?

Art. 50(2) has three exceptions: AI that only performs an assistive function for standard editing, AI that does not substantially alter the input data provided by the deployer or its meaning, and AI authorized by law for criminal law enforcement. The guidelines (points 68 and 90 to 92) get specific:

No marking neededMarking needed
Spelling and grammar correction, minor stylistic polishingSummaries
AI translations of textRewriting that changes style, structure or meaning
Format conversion, compression, noise removal, minor color correction, croppingRemoving, replacing or inserting objects or people in images and video
Transcripts of conversationsRealistic speech in a specific person's voice
Short sequences: single words, image captions, alt text, UI labelsComposites that change how people, objects, events or facts are shown
Source code, SQL, JSON and YAML configurationNewly generated text, images, audio and video
Machine-to-machine output, such as agent-to-agent messages no human seesOutput of an AI agent that a person perceives as text, image, audio or video

There is a narrow exception for purely B2B and industrial use (point 87). It applies only if all three conditions are met: the output is strictly technical (such as engineering designs or internal documentation), only a defined group of professionals at the provider and deployer sees it, and it is not meant to leave the company, with safeguards such as role-based access. A writing assistant for customer emails does not qualify.

How do I mark text, images, audio and video?

The law does not prescribe a technique. Recital 133 lists watermarks, metadata, cryptographic provenance methods, logging and fingerprints. The Code of Practice on Transparency of AI-generated Content, published in final form by the Commission on 10 June 2026, is more specific. Its Measure 1.1 requires two layers, because no single technique meets all four legal requirements (effective, interoperable, robust, reliable) on its own:

  1. Digitally signed metadata stating that the content is AI-generated or manipulated, time-stamped and tamper-evident (Sub-measure 1.1.1). The open standard for this is C2PA, known as "Content Credentials": a signed manifest is bound to the file by a cryptographic hash.
  2. An imperceptible watermark inside the content itself, in the pixels, audio samples or word choice (Sub-measure 1.1.2). It is meant to survive when metadata is lost during processing.

Fingerprinting or logging is an optional extra under Sub-measure 1.1.3. On its own it is not enough.

ContentWhat the Code asks for
Images, audio, videoSigned metadata and an imperceptible watermark; for video, both picture and sound
Text in a file (PDF, DOCX and similar)Signed metadata and a watermark
Free-form text of 200 tokens or moreA watermark; metadata is dropped because free-form text cannot carry it
Free-form text under 200 tokensNo watermark needed ("very short text" in the Code's glossary)

Detection is part of the duty. Under Measure 2.1, you provide a free way to check the mark, as an open specification, as software or as an API. For text watermarks you may limit access to vetted experts such as authorities, media and researchers, because detection there is less reliable (Sub-measure 2.1.2). According to the guidelines, personal data about the creator does not belong in the mark (point 94).

What do OpenAI and Anthropic already provide?

You may rely on the model provider's marking, but you remain responsible (point 74 of the guidelines, Measure 1.1 of the Code). So check what your provider actually ships. As of November 2026:

  • OpenAI: According to OpenAI's help page, images from the API carry C2PA metadata and a SynthID watermark, and audio carries a watermark. Models planned for imminent deprecation are excluded, including gpt-image-1 and gpt-image-1.5. The text watermark is not on by default in the API: you turn it on in your project or organization settings.
  • Anthropic: According to Anthropic support, text from supported Claude models carries a watermark, including via the API, and generated files such as PNG or JPEG get C2PA metadata. For some older models, the watermark may be missing if you set temperature, top_p or top_k to a non-default value; the API documentation lists the affected models. Detection of the text watermark is in private preview for eligible organizations.

Both providers say you have to assess yourself what Art. 50 requires of your product. The most common gap is on your side: if you resize images, convert them to WebP or serve them through an image CDN, you can strip the metadata. OpenAI explicitly warns that editing and converting files can remove it. Check your pipeline with a C2PA verification tool before the file reaches the user. If a layer is missing, you can add it yourself, for example your own C2PA signature on export or a watermark from a specialized service. The Code explicitly allows third-party solutions.

What must deployers label visibly on top?

Machine-readable marking is the provider's job. If you publish AI content, you have your own duties as a deployer under Art. 50(4), and people must be able to see them without any tool:

  • Deepfakes: images, audio or video that resemble real people, places or events and falsely appear authentic must be disclosed at the latest at first exposure. According to the Commission FAQ, you cannot rely on the provider's machine-readable mark for this. For art, satire and fiction, a disclosure that does not get in the way of the work is enough.
  • Text on matters of public interest: it needs a notice, unless a person has reviewed it and someone holds editorial responsibility. Spelling or grammar checks do not count as review.

For this, the Code proposes an EU icon with the letters "AI" (Section 2, Annex 1): for text, for example at the top or next to the headline; for deepfake videos, ideally for the whole video. Your product can help your customers here: Measure 1.4 encourages providers to offer an option to add a visible label when the content is generated. I would build that in if your customers publish images or text, because the duty then falls on them, not on you.

Do I have to sign the Code of Practice?

No, the Code is voluntary. The Commission assessed it as adequate for Art. 50(2), (4) and (5) on 8 July 2026, and the AI Board on 9 July 2026 (Commission opinion). By the end of July 2026, about 190 organizations had signed it, according to the Commission. Signatories are mainly checked on whether they implement its measures. Non-signatories must show compliance by other means, for example with a gap analysis against the Code, and should expect more requests for information (point 148 of the guidelines).

For a small team, the Code is the best template even without signing, because it spells out what authorities expect. It provides for proportionate implementation for SMEs and small mid-caps, for example in documentation and training (Measures 4.1 and 4.3), and lets you rely on your model provider's test results (Measure 4.2).

Breaches of Art. 50 can be fined up to €15 million or 3% of worldwide annual turnover. For SMEs and start-ups, the lower of the two applies, according to the guidelines (point 152). National market surveillance authorities enforce the rules.

Checklist before 2 December 2026

  1. List your features

    Write down every feature that generates or alters text, images, audio or video, with model, provider and the date it went live. If it was live before 2 August 2026, 2 December 2026 is your deadline; otherwise the duty already applies.

  2. Clarify role and exceptions

    For each feature, check whether you are provider, deployer or both, and whether an exception applies, such as grammar correction, translation, source code or short UI text.

  3. Check the model provider's marking

    Find out which layers your provider delivers. With OpenAI, turn on the text watermark; with Anthropic, check the sampling parameters in your requests.

  4. Test your pipeline

    Generate a file, run it through resizing, conversion and your CDN, and check at the end whether the C2PA metadata and watermark are still detectable.

  5. Close the gaps

    If a layer is missing, add it: your own C2PA signature on export, a watermarking service or both.
  6. Provide detection

    Point to your model provider's verification tool or offer your own, free of charge and with a result people can understand.

  7. Update your terms

    Prohibit the intentional removal of marks in your terms of use or documentation (Measure 1.2 of the Code).
  8. Offer a visible label

    Give your customers an option for a visible "AI" label if they might publish deepfakes or text on matters of public interest.

  9. Document it

    Record which technique you use, how you tested it and whether you sign the Code or do a gap analysis.

Frequently asked questions

Do I have to mark every chatbot reply?

Under the Code, free-form text of 200 tokens or more needs a watermark. Separately, since 2 August 2026 you must tell users they are talking to an AI (Art. 50(1)).

Is a visible "Made with AI" notice enough?

Not for Art. 50(2), because the mark must be machine-readable. A visible notice adds to it and helps your customers with their duties under Art. 50(4).

Does this apply to providers outside the EU?

Yes, if the system is offered in the EU or its output is used in the EU (point 10 of the guidelines).

This article is not legal advice; it summarizes publicly available sources as of 21 November 2026.

[ Anolilab — Blog ]

More hands-on writing on AI, web development and infrastructure.

Related articles.

[ Anolilab — Contact ]

Tell me what you're building. I'll tell you how I can help.

Let's talk.

Quick response.
If you're ready to create and collaborate, I'd love to hear from you.
Clear next steps.
After the consultation, I'll give you a detailed plan.
Alternatively, you can email me directly at hello@anolilab.com