What is AI metadata in an image?

Four families people search under “AI metadata”

C2PA / Content Credentials are a signed manifest (often called C2PA metadata) embedded in JPEG APP11 / JUMBF, PNG caBX, or WebP credential chunks, sometimes with XMP that points at the manifest. The Coalition for Content Provenance and Authenticity defines the format; Google Cloud’s Content Credentials documentation describes it as cryptographically signed provenance that can record AI generation. The dedicated map for that manifest alone is what Content Credentials are; the tool owner is remove C2PA.

IPTC DigitalSourceType is a standard vocabulary for how media was sourced. For generative AI, IPTC recommends trainedAlgorithmicMedia — “digital media created algorithmically using an Artificial Intelligence model trained on captured content” ( IPTC NewsCodes). You may see the full URI in XMP-iptcExt:DigitalSourceType or inside a C2PA IPTC assertion. Midjourney exports sometimes carry unsigned IPTC instead of C2PA — see Midjourney export AI metadata.

EXIF and XMP “AI” fields cover camera-style tags plus software strings. A Stable Diffusion JPEG may store the whole recipe in EXIF UserComment; ComfyUI may use XMP on WebP. Those are still file metadata, not pixels. The PNG-specific case is Stable Diffusion PNG AI metadata.

PNG text chunks are another metadata layer: Automatic1111’s parameters key, ComfyUI prompt / workflow, and similar keys are plain text beside the image data. AI provenance mode removes named generation keys; a leftover workflow chunk may need strip-all.

AI metadata tagging and extraction

The query ai metadata tagging usually means one of two steps. Tagging is the write: a generator, an editor, or a camera app stores C2PA, IPTC DigitalSourceType, XMP, or a PNG text chunk beside the pixels. Extraction is the read: open the file and list those fields. Neither step is the picture, and neither step is a detector score.

The check on this site is extraction of the fields the scanner knows. Inspect lists a C2PA line when a manifest is in the container, and an AI line when it sees a known hint such as DigitalSourceType or a PNG parameters chunk. An empty list means this scan did not find that category. It does not mean a model did or did not make the picture. Pixel watermarks such as SynthID are not extracted here, because they are not metadata. The direct page is can you remove SynthID.

Tagging that already happened can be removed only when it lives in those file fields. The steps are in the section below. A platform that stored its own label after upload does not give that label back as a field you can delete. This section answers ai metadata tagging. It does not publish a search volume for the phrase.

What is not “AI metadata” on this site

SynthID (Google) and similar pixel-domain watermarks are designed to survive many edits. Google’s Gemini help distinguishes SynthID from Content Credentials: SynthID is an invisible watermark in the content; Content Credentials are C2PA metadata that documents origin ( Gemini Apps Help). OpenAI documents the same split for supported ChatGPT downloads — metadata plus SynthID — on ChatGPT image AI metadata. This cleaner strips metadata; verifiers can still detect watermarks after a clean After report. The direct answer is can you remove SynthID from an image?

A platform badge may read file metadata, a cloud credential, or some other signal. The page to remove the Instagram AI info label covers a feed post, a Story, a carousel, the older Made with AI label on Instagram, and a badge that remains after the file report is already empty. What LinkedIn Content Credentials are is the CR icon on a still you have not uploaded yet. To take that icon off the next file, use remove the CR icon on LinkedIn. Cleaning the file does not edit a live post. When the report is already empty and a badge remains, do not run another strip pass.

How to check an image for AI metadata

1. Save or export the exact file you will upload — chat preview pixels are not the download.
2. Open the tool above and press Inspect only. Read chips for C2PA, AI, EXIF, GPS, and PNG text.
3. Match the producer: ChatGPT and Adobe Firefly often show C2PA; Gemini adds C2PA on newer model exports per Google’s blog; Midjourney may show IPTC without C2PA; local SD PNGs show text chunks.
4. If you need a written checklist for Instagram mismatches, see how to check photo AI metadata — it focuses on inspect-first workflow, not repeating this vocabulary list.

External validators (Adobe Inspect, OpenAI Verify, C2PA conforming products) read the same manifest bytes. They can also report watermarks this browser tool does not remove.

How to remove AI metadata you confirmed

Pick a mode from what the Before column listed. AI only (AI provenance) removes C2PA, Content Credentials XMP, DigitalSourceType when flagged as AI, and known generator PNG keys. All metadata (strip-all) also drops ordinary EXIF, IPTC, and remaining PNG text — use it when a prompt survives in EXIF or a workflow chunk remains. Press Clean this file and download only when the After report matches your goal. The AI metadata remover is the general cleaner. Generation tags on their own are the AI metadata cleaner. How that differs from AI Metadata Cleaner is the alternative page. Generator pages below spell out what each product tends to embed.

Generator-specific guides

• ChatGPT / DALL·E / API downloads — C2PA metadata strippable; SynthID not.
• Google Gemini / Imagen — C2PA on supported exports; SynthID not.
• Adobe Firefly — Content Credentials on exports; cloud copy separate.
• Midjourney — often IPTC / EXIF, often no C2PA chip.
• Stable Diffusion PNG — parameters and workflow text.

Once you know what is inside, the removal guides are listed by file type, device, and tool at remove image metadata.

Questions

Is AI metadata the same as a watermark?

No. AI metadata lives beside the pixels: C2PA JUMBF, XMP, IPTC, EXIF comments, PNG text chunks. SynthID and TrustMark are embedded in the image data. This tool removes metadata it can list. It does not remove pixel watermarks.

What is IPTC DigitalSourceType trainedAlgorithmicMedia?

It is a vocabulary term for synthetic media created with a model trained on captured content. IPTC publishes the URI http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. It can appear in XMP-iptcExt:DigitalSourceType or inside a C2PA IPTC assertion.

If the inspect report is empty, is the image free of AI?

No. An empty report means this scanner did not find metadata in the categories it reads. Invisible watermarks, remote credentials, and platform classifiers are outside that list.

Does checking or cleaning upload my photo?

No. Inspect and clean run in your browser on the bytes you selected.

Who operates the cleaner, and what it refuses to do, is on the about page.