Glossary

AI Act

Regulation (EU) 2024/1689, the EU's law on artificial intelligence. It regulates AI systems according to their risk and applies directly in every member state. The transparency obligations under Article 50 have applied since 2 August 2026. It was most recently amended by the Digital Omnibus.

More: Labelling obligation

AI literacy (Article 4)

Providers and deployers of AI systems must take measures to support the AI literacy of their staff (Article 4 AI Act as amended by Regulation (EU) 2026/1744). They do not have to guarantee a particular level. The obligation has applied since 2 February 2025, including to retailers who create AI images.

More: AI images in online stores

Article 50

The transparency obligations of the AI Act. Paragraph 1 covers chatbots, paragraph 2 machine-readable marking by providers, paragraph 4 visible labelling of deep fakes and certain texts by deployers, and paragraph 5 form and timing. Article 50 has applied since 2 August 2026 (Article 113).

More: Labelling obligation

Artistic works

If a deep fake is part of an evidently artistic, creative, satirical or fictional work, a disclosure that does not hamper the work is enough (Article 50(4), third sentence). The guidelines read “evidently” narrowly (para. 122). For product and advertising images this lighter form generally does not apply.

More: What a deep fake is

Bundesnetzagentur

Germany's Federal Network Agency, the market surveillance authority for the AI Act (§ 2(1) KI-MIG). It is also the central complaints office: complaints can be filed free of charge through an online form (§ 8 KI-MIG).

More: Fines and enforcement

Burn-in

The label is written permanently into the pixels of the image. It then travels with the file into feeds, marketplaces and downloads. This creates a new file and invalidates any existing C2PA manifest, so the original should be kept.

More: Label design

C2PA

Coalition for Content Provenance and Authenticity: an open standard for embedding signed provenance information in a file. It records who created a file, with which tool and whether AI was involved. If a program later changes the file without re-signing, the signature becomes invalid.

More: Metadata

Code of Practice

The Code of Practice on Transparency of AI-Generated Content, final version of 10 June 2026. It describes how providers can mark and deployers can label content, for example with “AI” as the main element, visible without interaction. Participation is voluntary, and the Code does not create a presumption of conformity (recital 41 of Regulation (EU) 2026/1744).

More: Code of Practice

compositeWithTrainedAlgorithmicMedia

Value of the IPTC field DigitalSourceType for images edited with generative AI, for example by inpainting or outpainting. In full: http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia.

More: IPTC tool

Content Credentials

The name under which C2PA provenance information is shown to people. It reveals whether and how AI was involved in an image, provided the generating tool wrote it and nobody removed it. Missing Content Credentials do not prove that no AI was used.

More: Check metadata

Deep fake

AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic (Article 3(60) AI Act). According to the guidelines this includes consumer goods, and it is enough that something could plausibly exist (para. 113).

More: What a deep fake is

Deployer

Anyone using an AI system under their authority, except for purely personal purposes (Article 3(4) AI Act). A retailer is a deployer when it, or a contractor acting on its instructions, uses AI to create or change images (guidelines paras. 12, 14). Deployers must label deep fakes (Article 50(4)).

More: AI images in online stores

Digital Omnibus on AI

Regulation (EU) 2026/1744, in force since 27 July 2026. It amends the AI Act: providers get until 2 December 2026 for machine-readable marking of older systems (Article 111(4)). There is no transition period for deployers.

More: Deadlines

DigitalSourceType

IPTC field (Iptc4xmpExt:DigitalSourceType) that states how an image was made: camera, software or generative AI. The value is an address from the IPTC vocabulary. Google reads the field, and Merchant Center requires it for AI-generated product images.

More: IPTC tool

EU icon

Three icons freely provided by the European Commission: “AI”, “AI GENERATED” and “AI MODIFIED”, in black and white. Using them is optional; an equivalent label of your own is allowed, for example for other colours. Using them does not mean signing the Code of Practice.

More: EU icons

EXIF

Exchangeable Image File Format: technical data that cameras and programs write into images, such as camera, exposure, date or software. EXIF is not designed for AI provenance, although some tools note their name there. EXIF data is often lost when images are uploaded or resized.

More: Metadata

First exposure

The moment a person sees content for the first time. The disclosure must be there at the latest then (Article 50(5) AI Act), for every person and every output (guidelines para. 143). In a store this is often the product card in a collection or search, not the product page.

More: Label design

Generative AI

AI systems that create new content from an input: text, images, audio or video. This includes image generators and generative fill in image editors. Their providers must mark the output in a machine-readable way (Article 50(2) AI Act).

More: Labelling obligation

Guidelines

The European Commission's guidelines on the transparency obligations under Article 50, C(2026) 5054 final of 20 July 2026. They explain the terms with examples, including product images. They are not binding (para. 5); only the Court of Justice of the EU can interpret the law bindingly.

More: Guidelines

Inpainting

AI regenerates one area of an image: an object is removed, replaced or added. Whether a label is then needed depends on the result. The guidelines call removing a passer-by from the background a minor change (para. 116); an added object that changes what the image says goes beyond standard editing (para. 92).

More: Labelling check

IPTC

International Press Telecommunications Council. Its Photo Metadata Standard defines fields for image information, including DigitalSourceType for provenance and, since version 2025.1, fields on the AI system used. The fields are usually stored in the file's XMP block.

More: Metadata

KI-MIG

The German act on AI market surveillance and innovation (KI-Marktüberwachungs- und Innovationsförderungsgesetz) implements the AI Act in Germany (Federal Law Gazette 2026 I No. 223). In force since 29 July 2026, it makes the Bundesnetzagentur the market surveillance authority and central complaints office.

More: Fines and enforcement

Legacy content

Deep fakes generated or manipulated with AI before 2 August 2026 do not need to be labelled retroactively. The guidelines encourage it, though (para. 154). What counts is the date of the last AI edit, not the upload date.

More: Deadlines

Manifest

The data block in which C2PA stores provenance information in a file. It contains statements such as the editing actions performed and the DigitalSourceType, signed with a certificate. A valid manifest shows that the information has not been changed since signing.

More: Metadata

Market surveillance

Authorities check whether the AI Act is complied with. They can ask deployers about their labelling practice (guidelines para. 148), and anyone can file a complaint (Article 85 AI Act). In Germany this is the Bundesnetzagentur.

More: Fines and enforcement

Merchant Center

Google's platform for product data in Google Shopping. For AI-generated product images, Google requires the IPTC field in the image file and prohibits removing embedded metadata. A visible label in the store does not replace this, and the field does not replace a visible label.

More: Google Merchant Center

Outpainting

AI extends an image beyond its edges, for example for a different aspect ratio. The guidelines count edge completion as standard editing (para. 92) and background extensions in product advertising as minor changes (para. 116). If a new scene is created, the assessment can differ.

More: Labelling check

Overlay

The label sits on top of the image, placed by the website's code. The image file stays unchanged. The Code of Practice explicitly allows such an overlay (measure 1.2.1 c). It only appears where the code runs, not in downloads, feeds or marketplaces.

More: Label design

Paragraph number (para.)

The guidelines are numbered paragraph by paragraph. “Para. 113” refers to paragraph 113 of guidelines C(2026) 5054. We cite this way so that you can check every statement in the original.

More: Guidelines

Provider

Anyone who develops an AI system or has it developed and places it on the market or puts it into service under their own name (Article 3(3) AI Act), such as the maker of an image generator. Providers of generative AI must mark its output in a machine-readable way (Article 50(2)). A retailer that merely uses such tools is generally not a provider but a deployer.

More: Labelling obligation

Second layer

Additional information on the label that opens on click, tap or keyboard: whether an image was generated or modified with AI and what was changed. The Code of Practice encourages it. The second layer must be accessible, for example closable with Esc (WCAG 1.4.13).

More: Label design

Standard editing

Editing that only prepares existing content for publication: cropping, colour and brightness correction, sharpening, background removal, edge completion (guidelines paras. 90 to 92). If AI only does this, a label is generally not required under the guidelines. If the product then looks different from reality, this no longer applies.

More: What a deep fake is

Synthetic performer

A term from New York State law (General Business Law § 396-b): an AI-generated, human-like figure that does not correspond to any recognisable real person. Advertising with such figures must disclose this conspicuously, since June 2026. EU stores selling to customers in the USA may be affected as well.

More: Worldwide

SynthID

An invisible watermark by Google DeepMind for content from Google's AI tools. It is embedded in the pixels themselves, not in the metadata, and can only be checked with Google's tools. Like any machine-readable marking, it does not replace the visible label (guidelines para. 117).

More: Metadata

trainedAlgorithmicMedia

Value of the IPTC field DigitalSourceType for images created entirely with generative AI. In full: http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. This is the value Google Merchant Center expects for AI-generated product images.

More: IPTC tool

Upscaling

AI enlarges an image and adds detail in the process. The guidelines call re-scaling in product advertising a minor change (para. 116), and enhancement for clarity counts as standard editing (para. 92). If the AI invents visible details on the product, it is more than upscaling.

More: Labelling check

Watermark

A mark in the image itself, visible or invisible. Invisible watermarks are among the machine-readable markings of providers (Article 50(2)). Deployers cannot rely on them because people cannot see them (guidelines para. 117).

More: Metadata

XMP

Extensible Metadata Platform: a format for writing metadata as text into an image file. It usually holds the IPTC fields, including DigitalSourceType. C2PA uses a data block of its own.

More: Metadata

Not legal advice. This content explains the law based on primary sources; it does not replace an assessment of your individual case.