ExifZero What metadata do AI image generators add?
By the ExifZero team · Updated
AI image tools differ a lot in what they write into the file. Hosted services tend to add a formal label; local tools tend to save everything needed to reproduce the image — including the prompt. This guide summarises what to expect. Behaviour changes over time, so check a fresh download with our AI image checker.
Hosted generators
ChatGPT and the OpenAI API
Images carry a C2PA manifest signed by OpenAI with a c2pa.created action and the trainedAlgorithmicMedia source type. Edits made
in ChatGPT add further actions.
Adobe Firefly and Photoshop generative features
Firefly output and Photoshop files that use Generative Fill or Generative Expand can carry Content Credentials naming Adobe Firefly, with AI source types on the relevant actions. Photoshop also writes extensive XMP, including edit history.
Google Gemini and Imagen
Google adds an invisible SynthID watermark to the pixels, and for many products also IPTC metadata and C2PA credentials. The metadata can be removed; the pixel watermark cannot be removed by stripping metadata.
Midjourney
Downloads carry XMP with the IPTC digital source type, and a description that includes the prompt and a Job ID linking back to the job.
Meta AI, Microsoft Designer, Leonardo, Ideogram and others
Most label output with the IPTC digital source type, C2PA or both. Some add only a visible watermark, and some add nothing at all.
Local tools
| Tool | Where | What |
|---|---|---|
| Stable Diffusion WebUI (A1111, Forge, SD.Next) | PNG parameters; EXIF UserComment in JPEG/WebP | Prompt, negative prompt, steps, sampler, CFG, seed, model hash, LoRAs |
| ComfyUI | PNG prompt and workflow | The full node graph, including model file names |
| InvokeAI | PNG invokeai_metadata | Prompt and generation settings |
| NovelAI | PNG Software, Comment | Prompt, undesired content, sampler and seed (JSON) |
| Fooocus | PNG parameters / fooocus_scheme | Prompt, styles, performance mode, seed |
Regional AI labelling rules
Since 1 September 2025, China's measures for labelling AI-generated content and the national standard GB 45438-2025 require services to add an implicit label — metadata identifying the content as AI-generated, with the producer and a content id — alongside visible labels. The EU AI Act requires providers of generative systems to mark output in a machine-readable way from August 2026. Many platforms (TikTok, YouTube, Instagram) also require creators to disclose realistic AI content when they post, whatever the file's metadata says.
Why it matters for privacy
Prompts can include names, client projects or unreleased ideas; ComfyUI workflows reveal model files and folder names. If you publish AI artwork, check what the file contains before you share it — and keep any disclosure your platform requires.