EU AI Act Article 50: what must be marked and disclosed from August 2, 2026, and the Commission guidelines of July 20, 2026

Article 50 of the EU AI Act applies from August 2, 2026. Providers of systems that generate images must mark the outputs in a machine-readable way and make them detectable. Deployers must disclose deep fakes. The Commission's guidelines of July 20, 2026 are non-binding and name no single required technique.

What does Article 50 require for images?

We read Regulation (EU) 2024/1689 on October 30, 2026. Three paragraphs matter for images.

Article 50(2), providers. A provider of an AI system that generates synthetic image content must ensure the outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated". The technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, given the content type, cost and state of the art.
Article 50(2), exceptions. The duty does not apply to the extent a system performs an assistive function for standard editing, or does not substantially alter the input data or its meaning.
Article 50(4), deployers. A deployer of a system that generates or manipulates image content constituting a deep fake must disclose that the content was artificially generated or manipulated. For an evidently artistic, creative, satirical or fictional work, the duty is limited to disclosing that such content exists, in a way that does not hamper the display or enjoyment of the work.

Article 50(5) adds that the information must be clear and distinguishable, no later than first exposure.

When does it apply, and what did the Commission say on July 20, 2026?

Article 113 of the Act says it applies from August 2, 2026. The Commission's guidelines page says the same, and shows the guidelines as published on July 20, 2026 and last updated on July 31, 2026.

We also opened the Communication, numbered C(2026) 5054, which approves the content of the draft guidelines, and the guidelines text annexed to it. Four points:

The guidelines say they are non-binding. Only the Court of Justice of the European Union can interpret the Act authoritatively.
The Communication says the Commission will formally adopt the guidelines at a later date, when all language versions are available. The Commission's page says it adopted them.
The guidelines describe an amending regulation, the AI Omnibus, as recently adopted. They say it envisages a rule giving providers of generative systems already on the market before August 2, 2026 until December 2, 2026 to meet the marking and detection duty. We did not open that regulation; the EUR-Lex text we read is the 2024 original.
Outputs made before August 2, 2026 need not be marked or labelled retroactively. For earlier deep fakes, the guidelines encourage labelling without disproportionate effort.

The guidelines say a provider or deployer that breaks Article 50 may be fined up to EUR 15,000,000 or, for an undertaking, up to 3% of worldwide annual turnover, whichever is higher.

Which technique does the Commission require?

None by name. The guidelines say a provider may use one technique or a combination, as long as the whole solution is machine-readable and meets the four quality tests. They cite recital 133 of the Act for examples: watermarks, metadata identifications, cryptographic provenance methods, logging and fingerprints.

Robustness, the guidelines say, means telling the provider's content from other content under varying conditions, "covering both common alterations and adversarial attacks". Detection is a separate duty: the guidelines say providers must make the means of detection available to people exposed to the content. Until a standard detector exists, a provider may use its own if it works with other providers' detectors.

The Commission's page on the Code of Practice on Transparency of AI-Generated Content says adherence to the Code is voluntary. A Paul, Weiss client memo dated August 4, 2026 reports that the Code asks providers for at least two techniques, such as signed metadata plus an imperceptible watermark. We did not read the Code itself.

Which image edits fall outside the marking duty?

Examples the Commission's guidelines list for images after paragraph 92. Read on October 30, 2026; each case depends on the facts.
Edit	How the guidelines treat it
Minor cropping, minor colour adjustments, sharpening for clarity, format conversion, noise reduction	Standard editing; the exception applies
Removing, replacing or inserting objects or people, where the meaning and substance change	Semantic change; marking required
Face replacement or substantial facial modification; composites that modify how people, objects, events or facts are represented	Semantic change; marking required

Who does the law bind, and does a deep fake need a visible label?

Article 50(2) binds providers. Article 50(4) binds deployers, who use a system under their own authority. The guidelines say a person acting in a purely personal, non-professional capacity is outside the deployer duties. A platform that only hosts or passes on content made by others is not a deployer, though it is strongly encouraged to preserve markings.

A deployer cannot rely on the provider's machine-readable mark for the deep fake duty, the guidelines say, because people cannot read it without tools. Article 3(60) defines a deep fake as content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. The guidelines give a sphinx flying over the Eiffel Tower as an example that is not one, and a manipulated scene with two real footballers as one that is.

How well does an open mark survive saving? Our measurement

Article 50(2) asks for robust marks as far as this is technically feasible. We tested one open mark, the invisible watermark that Stable Diffusion's reference software adds. On October 27, 2026 we marked 7 test files with the real invisible-watermark package and saved them in several ways. The checker found the mark in 7 of 7 PNG files. It found it in 0 of 14 saves as ordinary 4:2:0 JPEGs (the same 7 files at quality 90 and at quality 75). It found it in 7 of 7 JPEGs saved at quality 90 with 4:4:4. So robustness depends on what happens to the file. The full table is on our accuracy page.

This tests one mark in one browser. It says nothing about any provider's compliance or other marks.

What this means when you check an image

Open the checker and choose the file. It cannot test whether a provider complies with Article 50.

"No supported provenance signals found" means none of the signals we read were in the file. The line under it says this does not mean the image is real or human-made. It does not show that a provider failed a duty: a mark may have been lost when the file was shared, resized or re-saved. The note under every result lists what we do not check, including Google's SynthID watermark.

"Content Credentials found; signer trust not established" means the file has a signed record, but no trust list we load vouches for its signing certificate, so the name in it is not verified. It does not say who made the image or how.

Guides on this topic

Frequently asked questions

Do I have to mark an image I made with an AI tool?: Article 50(2) puts the marking duty on the provider of the system, not on the person who prompts it. A professional deployer may have to disclose a deep fake under Article 50(4). The guidelines exclude purely personal, non-professional use.
Are the Commission's guidelines the law?: No. They say they are non-binding. The Act's text is the law.
Does a Content Credentials record satisfy Article 50(2)?: The guidelines list cryptographic provenance methods among possible techniques. They allow one technique or a combination if the whole solution meets the four tests. We do not judge any provider's compliance.
Does the EU rule require a visible watermark?: Providers must add a machine-readable mark; the guidelines do not exclude visible marks as an extra. For deep fakes, deployers must disclose in a form people can perceive.