What Is AI Metadata? The Four Layers Inside an AI Image in 2026

An AI-generated image looks like any other JPEG or PNG. Inside, it can carry a surprising amount of information about how it was made. In 2026 that information comes in four layers, and they behave very differently: some are plain text anyone can read or delete, some are cryptographically signed, and some are woven into the pixels themselves.

This guide explains each layer, who uses it, and what a metadata cleaner like BlankAI can and cannot change. The short version: file metadata is where your personal details live and it is easy to remove; invisible watermarks are where AI companies now put their durable marks, and they are not.

Layer 1: plain metadata (EXIF, XMP, IPTC, PNG text)

The oldest layer is ordinary file metadata. Cameras use EXIF for device and GPS data. Editing apps use XMP for history and IPTC for captions and credits. PNG files have free-form text chunks. AI tools reuse all of these.

• Midjourney writes the prompt, parameters and Job ID into the Description field, your username into Author, and the IPTC digital source type trainedAlgorithmicMedia.
• AUTOMATIC1111 and Forge save a “parameters” PNG chunk with prompt, seed, sampler and model. ComfyUI saves the whole workflow graph.
• The FLUX.1 reference code writes “AI generated;txt2img;flux” into EXIF Software and your prompt into ImageDescription.
• Photoshop exports that used Generative Fill carry the IPTC value compositeWithTrainedAlgorithmicMedia.

The IPTC digital source type is worth knowing by name. It is a controlled vocabulary: trainedAlgorithmicMedia means fully AI-generated, compositeWithTrainedAlgorithmicMedia means a real image with AI-edited parts. Instagram and Pinterest read it.

Layer 2: C2PA Content Credentials

C2PA manifests are signed records of how a file was made and edited. OpenAI, Google's Gemini app, Adobe Firefly, Microsoft Copilot, Black Forest Labs' API, TikTok's AI tools and Anthropic now sign their image output, as do some cameras and phones. The signature proves the record has not been altered, but the record itself sits in the file like any other metadata, so re-encoding the image leaves it behind.

Layer 3: invisible watermarks

Watermarks are the layer that changed most in 2026. Google's SynthID is embedded in the pixels of every image Google's models produce, and OpenAI has added it to ChatGPT and API images since 19 May 2026. Apple has announced SynthID for Apple Intelligence images in a later iOS 27 update. Meta uses its own Content Seal on Muse Image output, and TikTok and ByteDance use their own marks. Some local pipelines, including Stability's SDXL scripts and the FLUX.1 reference code, add a simpler DWT-DCT watermark.

Watermarks are designed to survive compression, resizing, cropping and screenshots. Removing metadata does not affect them.

Layer 4: visible marks

Some services still draw a mark on the image: Grok Imagine's watermark cannot be turned off, Seedream's API adds an “AI generated” corner label by default, and the Gemini sparkle can now be switched off by most users for new images. Older Meta Imagine images carry “Imagined with AI”.

Why the metadata layer still matters to you

Metadata is where the personal details are. A prompt is often a creator's working method. A username links a file to an account. Photos edited with AI still carry the GPS coordinates and device serial numbers of the original. Editing histories reveal the tools and steps behind client work. None of that is needed to show the picture, and all of it can be removed.

What metadata removal cannot do

It cannot remove watermarks, it does not affect classifiers that judge an image by its appearance, and it does not change any platform's disclosure rules. If you publish realistic AI content on TikTok, YouTube or Meta's apps, their rules expect a label. Some laws also address removal directly: China's labeling measures prohibit deleting AI labels, and several generators forbid removing provenance data in their terms.

Frequently asked questions

Does every AI image contain metadata?

No. It depends on the tool and the path the file took. xAI documents no metadata for Grok Imagine, and platforms such as Discord re-process uploads, which often drops the original metadata. Check the actual file.

Can AI metadata be faked?

Plain metadata can be written by anyone. C2PA manifests are signed, which makes forgery harder, but in August 2026 a researcher showed that a rooted Pixel 10 could sign arbitrary images with valid camera credentials. Treat provenance data as evidence, not proof.

Is it legal to remove AI metadata?

Removing metadata from your own files is generally legal in the US and EU, where current rules place duties on AI providers and large platforms. China prohibits removing AI labels, and generators' terms may forbid it. This is general information, not legal advice.