What Metadata Do AI Image Generators Put in Your Images?

Every AI image generator embeds metadata into the files it produces. The type, depth, and location of that metadata varies dramatically by tool. DALL-E 3 buries a cryptographically signed certificate deep in the binary. Stable Diffusion writes your entire prompt in plain text into a PNG chunk. Adobe Firefly signs with Adobe's root certificate and timestamps the creation.

Platforms like Instagram, LinkedIn, and TikTok read this metadata automatically at upload time. You cannot see it. They can. Here is what is actually inside your AI images and which generators put it there.

C2PA Content Credentials

C2PA (Coalition for Content Provenance and Authenticity) is an open technical standard for embedding signed provenance data into image files. It was built by a consortium that includes Adobe, Microsoft, BBC, Intel, and Arm. When an AI tool embeds C2PA credentials, it creates a cryptographically signed manifest that declares the image as AI-generated and identifies the specific tool that created it.

The following tools embed C2PA by default: Adobe Firefly, DALL-E 3 and ChatGPT image generation, Microsoft Copilot Designer, Google ImageFX, and DreamStudio (Stability AI's hosted Stable Diffusion). These are all founding or early members of the C2PA standard.

In JPEG files, C2PA lives in the APP11 marker segment at byte offset 0xEB. This segment holds a JUMBF (JPEG Universal Metadata Box Format) container with content blocks including the generator assertion, signer certificate, a timestamp from the signing authority, and a soft binding hash tied to the image pixels.

In PNG files, C2PA appears in dedicated chunk types: caBX (Content Authenticity Box), caMs (CA Manifest), and caSt (CA Store). These chunks sit in the PNG chunk sequence before the IEND marker.

C2PA is the primary trigger for "Made with AI" labels on Instagram, LinkedIn, YouTube, and Pinterest. These platforms are C2PA member organizations and read the manifest automatically at upload time.

XMP metadata (AI tool signatures)

XMP (Extensible Metadata Platform) is an XML packet embedded in most image files. It was created by Adobe and is widely supported across tools and platforms. Nearly every AI image generator writes identifying information into XMP fields, even tools that do not embed C2PA.

Common XMP fields written by AI generators include the creator tool name and version, a creator identifier, software identifiers, job or session IDs, and a standardized digital source type field from the IPTC standard that explicitly marks the image as algorithmically generated.

The IPTC trainedAlgorithmicMedia value is a secondary trigger for platform AI labels. It is separate from C2PA and read independently by some platforms.

Stable Diffusion's parameters chunk

Stable Diffusion running through Automatic1111 (A1111) does not embed C2PA. Instead, it writes a PNG text chunk named parameters that contains your complete generation workflow in plain text.

This chunk includes your full positive prompt, the full negative prompt, number of steps, sampler name, CFG scale value, seed, model hash and name, and any LoRA weights applied. This is not a summary. It is every parameter used to generate the image.

This chunk is read natively by A1111's own PNG Info tab. Anyone who receives your PNG file can paste it into A1111 and extract your entire workflow instantly. There is no C2PA involved, but the exposure is arguably more severe.

EXIF and GPS data

EXIF (Exchangeable Image File Format) metadata is most commonly associated with camera-captured photos. It records the device model, lens, exposure settings, and timestamp. Some AI generators embed partial EXIF including a software field that names the generating tool and a creation timestamp.

The GPS risk is most relevant in a hybrid workflow scenario: if you photograph something on your phone, apply AI enhancement through a camera app or filter, and export the result, the original EXIF, including GPS coordinates from where the photo was taken, may be preserved in the output. The AI layer sits on top but the location data from the original capture remains in the file.

Pure AI generators (DALL-E, Firefly, Midjourney) do not embed GPS because there is no source photo with a location. The risk is specific to AI-enhanced photography pipelines where device metadata is preserved through the editing process.

Why you can't see this metadata

None of this metadata is visible inside the image itself. It lives in binary segments that image viewers ignore entirely. File managers on Windows and macOS do not surface XMP or C2PA fields. Right-click, Get Info on a Mac shows basic dimensions and a handful of EXIF fields at most.

To actually read C2PA JUMBF blocks, XMP packets, or PNG text chunks, you need a tool that specifically parses the raw binary structure of the file and knows what to look for. Most metadata viewers focus on EXIF because that is what photographers care about. AI metadata lives in segments those tools were never built to surface.

StripShot's detection panel reads all binary segments before stripping, surfacing what is actually present in the file before you decide what to remove. You see exactly what platforms like Instagram are reading when you upload.

Which generators embed what

Frequently asked questions

Related guides