What is AI image metadata, and which platforms read it?

When an AI tool makes an image, it may write a note about that into the file. Platforms that read the note can label the image, and a shop that keeps the note intact can show where its pictures came from. This guide explains the three kinds of note, which platforms say they read them, what we measured about how often they exist and where they get lost, and how to write one without damaging the file. It is general information, not legal advice, and we checked every source on October 21, 2026.

Metadata is one layer and a visible label is another. The Commission's Guidelines say deployers cannot rely on the provider's embedded marking for a deep fake, as we explain below. For the legal side, start from our AI disclosure laws hub and the platform rules page.

The three markers

Invisible watermarks such as Google's SynthID are a different mechanism. They sit in the pixels, not in the file's metadata, so a metadata reader does not see them. The EU's Code of Practice on transparency of AI-generated content asks its provider-side signatories for "at least two layers of machine-readable marking" (Section 1, Measure 1.1) and names no file format. A text search of the Code finds neither C2PA nor IPTC.

What real AI images carry

We measured this on October 21, 2026. The samples are small and the files may have been re-saved, so read them as observations and not as rates.

• 65 AI images from Wikimedia Commons, across 15 generator categories: 15 carried a C2PA manifest, 2 carried an IPTC DigitalSourceType in plain XMP, 4 carried only a vendor-specific hint, 15 had other metadata with no AI marker, and 29 (45%) had no metadata at all. A reader that looks only at IPTC XMP misses the Adobe Firefly, ChatGPT and Microsoft Image Creator files entirely.
• In shops, almost nobody tags yet. None of 38 first product images from 13 Shopify stores and none of 62 Amazon images (12 listing images, 50 A+ Content images) carried an IPTC AI source type or a dc:subject keyword. One Amazon A+ image carried a C2PA manifest with an Adobe signer and the compositeWithTrainedAlgorithmicMedia value.
• In news, IPTC's Metawatch run of October 1, 2026 on 425 publisher sites found C2PA on 0.5% of 6,580 lead images and any IPTC metadata on 11%.

Who says it reads what

"Reads" is not "requires by law". These rows are what each platform says on its own pages, with the page's date where it shows one.

A visible label is a separate duty. Paragraph 117 of the Commission's Guidelines on Article 50 (C(2026) 5054 final, non-binding) says "deployers cannot rely on the machine-readable marking embedded in the content by the provider", because those markings are "not immediately clear and distinguishable" to people. Our guide to accessible AI labels covers the visible side.

Does the marking survive your store?

The file your customers receive is often not the file you uploaded. We fetched public images and requested several variants of each. We did not upload test files to stores, so whether a tag you upload survives each pipeline is untested.

The practical rules follow. Check the file your customers' browsers receive, not the master. Ask for both a JPEG-friendly and a WebP version when you test. Keep the tagged original. And read every manifest in a file: in a ChatGPT sample the generative claim sat in the parent manifest, not the active one. Our free AI image metadata checker reads what a file declares in your browser and has a mode that checks whether the markers survive your store.

Writing a tag without breaking the file

• It can be done without re-encoding. We wrote a DigitalSourceType and a keyword into 70 files (JPEG, PNG and WebP, including animated, transparent and gain-map files). All 70 read back in three independent readers, and the pixel data was byte-identical. Our AI image metadata writer works this way. It does not handle GIF, AVIF, HEIC, TIFF or SVG.
• Do not add XMP to a file that has C2PA. In all 3 tested files, adding XMP changed a valid signature to invalid (assertion.dataHash.mismatch). A file that already carries a manifest already declares its origin in signed form, so leave it alone.
• Never strip a marking. Section 1, Measure 1.2 of the Code of Practice says signatories will not make available or promote tools "whose purpose is to circumvent the machine-readable markings", and Google says "Don't remove embedded metadata tags". Our writer has no remove function. To strip location data for privacy, remove those fields only. A tool that clears all metadata clears the AI declaration too.
• Pick the value that matches the edit. IPTC's vocabulary gives trainedAlgorithmicMedia for media created with a generative model and compositeWithTrainedAlgorithmicMedia for augmentation such as inpainting or outpainting. Merchant Center's help page names only the first, and we found no Google page that says which value a real product photo with an AI-generated background needs.
• Leave the prompt field empty. IPTC's 2025.1 standard added fields for the AI system, its version and the prompt. A prompt can hold text you would not publish.

What metadata cannot do

• It is not a label people see, as paragraph 117 says.
• Presence is not trust. Unsigned XMP can be typed by anyone. A C2PA file raises three questions: what it claims, whether it is intact, and whether the signer is trusted. Of 15 real AI images with C2PA that we validated with the official browser library, 8 came back as trusted and 7 as invalid on credential grounds such as an untrusted or expired signer certificate, with no sign of tampering.
• Absence proves nothing. 45% of our AI sample had no metadata, and a pipeline may have stripped the rest. A file with no marker is not evidence that an image is not AI-made.
• It cannot tell whether an image is AI-generated. A reader reports what the file declares. Anthropic's documentation for the vision model we use says it "cannot determine whether an image is AI-generated", so we read declarations and do not guess.