“AI watermark” covers three different things: a visible label in the image, an invisible watermark in the pixels (such as Google’s SynthID) and provenance metadata such as C2PA. Invisible watermarks and metadata are added by the providers of the generators under Article 50(2) of the EU AI Act; an invisible watermark can usually only be detected by that provider’s own tool, while C2PA metadata can be read by anyone. For your store, only the visible part counts: under the Commission guidelines you cannot rely on the provider’s marking (paragraph 117), you need a visible label on the image.
People searching for “AI watermark” often mean three different things: a visible label in the image, an invisible mark in the pixels, and provenance data in the file. Legally, they belong to two separate obligations with two different addressees. This guide separates the three, shows how each can be detected, and explains which part is your job as a store. The obligations behind it are covered in EU AI Act Article 50.
What is an AI watermark?
| Visible watermark (label) | Invisible watermark | Provenance metadata (C2PA, IPTC) | |
|---|---|---|---|
| What it is | icon or text in the pixels, such as “AI GENERATED” | a mark in the pixels that people cannot perceive, such as Google’s SynthID | signed information in the file about who created the image with which tool |
| Who adds it | you as the deployer, or the provider in addition | the provider of the generator | the provider of the generator, sometimes editing software |
| Legal basis | Article 50(4) and (5) AI Act (disclosure) | Article 50(2) AI Act (marking) | Article 50(2) AI Act (marking) |
| Who can detect it | every viewer | usually only the provider’s tool | anyone with a C2PA or metadata reader |
| Survives a screenshot | yes, if it is part of the image | designed to survive alterations (Code: “robust”) | no |
| Counts as your disclosure (Art. 50(4)) | yes | no (paragraph 117) | no (paragraph 117) |
The Code of Practice of 10 June 2026 defines a watermark in the narrow sense as a technique that embeds a marker in the content in a way that is imperceptible to people and robust, tied to the image pixels or audio samples and explicitly not auxiliary metadata (glossary, “Watermark”).
Who has to add an AI watermark?
The providers of the generators. Article 50(2) of the AI Act requires their outputs to be “marked in a machine-readable format and detectable as artificially generated or manipulated”. The Act does not prescribe a technique; recital 133 mentions watermarks, metadata identifications, cryptographic methods for proving provenance, logging methods and fingerprints, also combined (Guidelines, paragraph 73). Marking always comes with a means of detection; marking alone is not enough (paragraphs 69 and 70).
For signatories, the Code of Practice sets this out in layers:
- Digitally signed metadata stating whether content is AI-generated or manipulated (sub-measure 1.1.1), in practice mostly C2PA.
- An imperceptible watermark that is difficult to separate from the content and backs up the metadata (sub-measure 1.1.2).
- Optionally fingerprinting or logging, which on its own is not sufficient (sub-measure 1.1.3).
Providers whose systems were placed on the market before 2 August 2026 have until 2 December 2026 to comply (Article 111(4) AI Act, added by Regulation (EU) 2026/1744). Until then, gaps are normal. What changes on that date: What changes on 2 December 2026?
How do you detect an invisible AI watermark?
Only with the tool that matches the watermark. According to the guidelines, there is not yet a provider-agnostic standard for detecting invisible watermarks; until there is, providers may use their own or shared detection solutions as long as these work together (paragraph 76). Signatories of the code make their detection tool available free of charge, as an open specification, as software or as an online service (sub-measure 2.1.1), and state whether a result is based on metadata, a watermark or another method (measure 2.3).
Two examples, as of 4 October 2026:
- SynthID (Google): in the Gemini app, you upload an image, video or audio file and ask whether it was made with Google AI. According to Google, the check can currently only recognise content created by Google AI tools; if no watermark is found, the content may still come from another AI system. Gemini also reads Content Credentials. The number of checks is limited, for images to about ten per 24 hours.
- TrustMark (Adobe): the Content Authenticity Initiative describes “durable Content Credentials” made of three parts: metadata, an invisible watermark carrying an identifier, and a fingerprint. The identifier can be used to recover the metadata if a platform has stripped it (post of 22 January 2026). Adobe has released TrustMark under the MIT licence.
Sometimes the metadata also states that a watermark was added. In the measurement in the test bench of Easy AI Act Image Labels on 1 September 2026, the C2PA manifest of a Google Gemini image declared “Applied imperceptible SynthID watermark”, and the manifest of a ChatGPT image contained the action “watermarked”. That is the provider’s statement; only the provider’s tool checks the watermark in the pixels itself.
What about C2PA metadata?
C2PA (Content Credentials) and IPTC DigitalSourceType are data in the file that anyone can read. A marker found is evidence of AI origin; a missing marker says nothing, because many tools do not write one and a screenshot, format conversion or re-encoding deletes it. You can check this free of charge with the AI image checker or the AI metadata checker; both read the metadata in your browser and cannot see a watermark in the pixels. What is inside a manifest: C2PA and Content Credentials.
In Shopify, the marker stays in the stored original, while the delivered copies mostly lose it: in a test store on 30 September 2026, 6 of 38 marked images still carried the marker on the copy, all through an XMP entry; C2PA-only markers did not survive re-encoding there. Whether invisible watermarks survive delivery has not been measured.
Does an AI watermark replace your label?
Not if it is invisible. As a deployer, you must disclose realistic AI content under Article 50(4) in a way people can perceive without technical tools. The Commission guidelines state expressly that deployers cannot rely on the provider’s machine-readable marking for this, because it is not immediately clear and distinguishable for viewers (paragraph 117).
A visible watermark, on the other hand, is a form of labelling: a label such as “AI GENERATED” burned into the file is a visible disclosure under Article 50(4) and travels with the image when it leaves your store. The guidelines see perceptible labels added by providers as a possible complement that helps deployers (paragraph 71). How to decide between an overlay and a burned-in label: AI label: overlay or burned into the image? Examples of correct labels: AI generated label examples.
May you remove an AI watermark?
The AI Act contains no express prohibition for third parties. The Commission does encourage online platforms, search engines and others distributing content to preserve the marks (Guidelines, paragraph 98). Generator terms and platform rules can be binding: Google Merchant Center requires metadata such as IPTC DigitalSourceType in AI images and prohibits removing embedded tags.
Above all, removal does not change your obligation. An AI image without a watermark is still an AI image, and if it looks real, it needs a visible label.
What does this mean for your store?
- Know your obligation: realistic AI images need a visible label on the image, whether or not a watermark is inside.
- Use the markers: provenance data in the original shows you which images come from an AI according to their creator. Check the original files, not the storefront copy.
- Remove nothing: keep metadata and watermarks in the originals, especially for Google Shopping.
- Burn in for channels: images for ads, marketplaces and social media get a visible, burned-in label or the platform’s own labelling route.
Labelling remains your decision and your responsibility as a deployer; no watermark takes it off your hands.
Frequently asked questions
Does every AI image have a watermark?
No. Providers must mark their outputs in a machine-readable format under Article 50(2) of the AI Act, but providers of older systems only from 2 December 2026 (Article 111(4)). In the measurement of 1 September 2026, 34 of 47 image models in an aggregator catalogue wrote a signed provenance marker and 13 did not; whether an invisible watermark sits in the pixels as well can only be checked with that provider’s detection tool.
How can I check for SynthID?
With Google’s Gemini app: upload an image and ask whether it was made with Google AI. According to Google, the check can currently only recognise content created by Google AI tools; if no SynthID watermark is found, the image may still come from another AI system. Gemini also reads Content Credentials (Google help page, accessed 4 October 2026).
Does an AI image detector find watermarks?
No. A probability detector estimates from the pixels whether an image looks AI-made and therefore only gives a hint. A watermark is a deliberately embedded mark that the provider’s matching tool can find again. They are separate methods.
Is the watermark of my AI tool enough as a label?
No. Invisible watermarks and metadata are the provider’s marking under Article 50(2). As a deployer, you must disclose realistic AI images visibly to people; the Commission guidelines state expressly that you cannot rely on the machine-readable marking for this (paragraph 117).
May I remove an AI watermark?
The AI Act does not expressly prohibit third parties from doing so, but the Commission encourages everyone distributing content to preserve the marks (Guidelines, paragraph 98). Generator terms and platform rules can prohibit removal; Google Merchant Center does so for embedded metadata. Removing it does not change your obligation to label visibly.
Does a watermark survive in my Shopify store?
For provenance data in the metadata this has been measured: the stored original keeps it, the delivered copies mostly do not. In a test store on 30 September 2026, 6 of 38 marked images still carried the marker on the copy, all through an XMP entry; C2PA-only markers did not survive on any re-encoded copy there. Whether invisible watermarks in the pixels survive Shopify’s delivery has not been measured.
Sources
- Regulation (EU) 2024/1689 (AI Act), CELEX 32024R1689, Article 50(2), (4), (5), recital 133. publications.europa.eu (accessed 4 October 2026)
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), CELEX 32026R1744, Article 1 point 39(b) (Article 111(4)). publications.europa.eu (accessed 4 October 2026)
- European Commission, Guidelines on the implementation of the transparency obligations under Article 50, Annex to C(2026) 5054 final of 20 July 2026, paragraphs 69 to 76, 98, 117. ec.europa.eu (accessed 4 October 2026)
- Code of Practice on Transparency of AI-generated Content of 10 June 2026, sub-measures 1.1.1 to 1.1.3, 2.1.1, measure 2.3, glossary “Watermark”. ec.europa.eu (accessed 4 October 2026)
- Google, Gemini help “Verify AI-generated images, videos, and audio” (no page date). support.google.com (accessed 4 October 2026)
- Content Authenticity Initiative, “The three pillars of provenance that make up durable Content Credentials”, 22 January 2026. contentauthenticity.org (accessed 4 October 2026)
- Google Merchant Center, help on AI-generated images (IPTC DigitalSourceType, do not remove metadata). support.google.com (accessed 30 September 2026)