Which AI Image Generators Add Metadata?

The table

This table shows what each company or project says about its image output. We checked the sources on November 8, 2026. These policies change fast, so treat the table as a snapshot.

How to read the table

• C2PA Content Credentials are signed records inside the file. See What Are Content Credentials (C2PA)?
• IPTC or XMP markers are plain metadata fields, such as DigitalSourceType. They are not signed, so any metadata editor can remove them.
• PNG text chunks are free text in the PNG file. Stable Diffusion tools use them for the prompt and the settings.
• Pixel watermarks are invisible signals in the image itself. WipePic cannot remove them.

Notes on single tools

OpenAI. OpenAI says that images from ChatGPT, Codex and the API include both Content Credentials and SynthID. It also says coverage may vary by model, export path and file type. Read ChatGPT Image Metadata for details.

Google. In November 2025 Google said that images from Nano Banana Pro would carry C2PA metadata in the Gemini app, Vertex AI and Google Ads, and that it planned to expand this to more products. We found no statement that covers every Google image tool, so we mark those as not verified.

Adobe. Adobe lists the content of its credentials: the issuer, the date, the app, the AI tool and the action. Adobe does not state a pixel watermark on that page.

Midjourney. Midjourney says every image it delivers has a small ID tag hidden in the file, and that screenshots, edits and most social media sites remove it. In May 2023 IPTC announced that Midjourney would use the IPTC DigitalSourceType marker for generated content. We found no current Midjourney statement about C2PA or a pixel watermark.

Microsoft. In September 2024 Microsoft said it had added Content Credentials to images from Bing Image Creator, Microsoft Designer and Copilot. Microsoft Learn (2026) says its AI systems also use invisible watermarks, and its detection tool checks for them. It does not list which product adds which signal.

Stable Diffusion tools. The AUTOMATIC1111 wiki lists a feature that adds the generation parameters to the PNG as a text chunk. ComfyUI's documentation says its save nodes write the workflow and the API prompt into the file. The prompt text in these chunks can be long and personal.

Meta. Meta says that for photorealistic images from Meta AI it adds visible markers, invisible watermarks and metadata in the file. Meta labels content on its apps when it detects industry-standard indicators such as C2PA and IPTC data, or when a person says the content is AI-generated.

What this means when you share an AI image

• To see what your file holds, use the AI metadata checker.
• To remove Content Credentials, use Remove Content Credentials.
• To remove the prompt and settings of Stable Diffusion tools, use Remove Stable Diffusion Metadata.
• To remove all AI labels in one pass, use Remove AI Metadata.

WipePic removes the metadata in the file and does not change the pixels. Pixel watermarks such as SynthID stay in the image. Read Does Removing Metadata Remove SynthID? for the reason.

Sources

OpenAI Help Center, Provenance signals (2026); OpenAI, Advancing content provenance (May 2026); Google, image verification in the Gemini app (November 2025); Google, SynthID Detector (May 2025); Adobe Help Center, Content Credentials overview (2026); IPTC, Midjourney and Shutterstock AI (May 2023); Midjourney, Content Authenticity; Microsoft, written testimony of Brad Smith (September 2024); Microsoft Learn, Content provenance (2026); AUTOMATIC1111 wiki, Features; ComfyUI docs, Workflow metadata; Meta, Labeling AI-generated images (February 2024).