INITIALISATION DU SYSTÈME...
Can you tell if an image was generated by AI? Can C2PA be removed? Can a photo reveal where you live? Here are direct answers to the most asked questions — each one links to the deeper guide or tool.
| Generator | Embedded marker | Location | Removable by re-encoding? |
|---|---|---|---|
| ChatGPT / DALL-E (OpenAI) | C2PA manifest + IPTC trainedAlgorithmicMedia assertion | caBX (PNG) / APP11 (JPEG) | Yes |
| Sora (OpenAI) | C2PA Content Credentials | APP11/JUMBF (JPEG frame) | Yes — but not the server-side soft binding |
| Midjourney | tEXt/iTXt chunks (prompt, job ID) + caBX (Pro/Mega) | PNG | Yes |
| Adobe Firefly / Photoshop | C2PA Content Credentials | JPEG / PNG | Yes |
| Stable Diffusion (A1111 / ComfyUI) | Prompt, seed, full workflow | PNG "parameters" chunk / iTXt | Yes |
| Google Imagen / Gemini | C2PA + IPTC — and SynthID watermark | JPEG (metadata) + pixels (SynthID) | Metadata yes — SynthID NO (pixels) |
| Smartphone photo | EXIF, GPS, date, device model | JPEG / HEIC | Yes |
Yes, often. DALL-E/ChatGPT, Sora, Adobe Firefly and Microsoft Copilot sign their files with a C2PA manifest (Content Credentials), and many generators write the prompt and tool name into PNG chunks or XMP. But not always: those metadata vanish in a screenshot, and some tools embed none. Pixel-level markers (SynthID watermark, AI classifiers) also exist, independently of metadata. The simplest reflex: inspect the file itself.
Source: C2PA Viewer — Content Credentials adoption (2026)
Yes. The C2PA manifest is stored inside the file — a JUMBF box holding a claim, assertions and a cryptographic signature — and is not bound to the pixels. A clean re-encode from a bare canvas removes the whole structure, with no leftover fragment and no visible quality loss. It can be done 100% in the browser, without ever uploading the image to a server.
Source: Official C2PA specification
Screenshotting: yes, for metadata — the manifest doesn't propagate into new pixels, but you lose the original quality. Cropping: no — most editors copy the APP11 manifest over unchanged, and since the hash no longer matches the new pixels, a C2PA verifier flags the image as "modified" rather than clean. To truly erase the metadata layer, the JUMBF box itself must be removed through re-encoding.
Source: Sesamedisk — AI Content Provenance 2026 (metadata is not bound to pixels)
No. SynthID is a steganographic watermark embedded in the pixels by Google DeepMind, designed to survive standard re-encoding — over 100 billion pieces of content have carried it since 2023 (Presenc, 2026). Removing metadata (C2PA, EXIF, XMP) does not touch it — be wary of any tool claiming otherwise. GhostMeta states it plainly: it removes the metadata layer only, never pixel watermarks.
What GhostMeta removes from Google images
Source: Google DeepMind — SynthID · Presenc — AI Content Watermarking Adoption 2026
Yes, by default. Every image carries a C2PA manifest with the IPTC digitalSourceType assertion set to trainedAlgorithmicMedia and a content_binding hash of the file's bytes. That is exactly the field automated marketplace admission filters read. Privacy-wise, your account identity is not embedded in the file. Once the manifest is removed, the assertion no longer exists.
Source: IPTC NewsCodes — digitalSourceType (official definition)
Yes. OpenAI attaches Content Credentials to Sora outputs; on a JPEG frame, the manifest lives in an APP11 segment in JUMBF format. Removing it from the file works. However, a "soft binding" — a fingerprint Sora can compute from the pixels and keep server-side — is not in your file and therefore survives any local cleaning.
Source: C2PA Viewer — OpenAI joins the C2PA steering committee (May 2026)
Yes. Midjourney PNGs embed tEXt/iTXt chunks named Description, parameters, prompt and Software, often XMP, the job ID and version — plus a C2PA caBX box on Pro and Mega tiers. These are the chunks stock and print-on-demand platforms scan at upload, triggering automatic AI flags or rejections.
Yes. Automatic1111 writes the entire generation chain — prompt, negative prompt, seed, sampler, CFG, model hashes — into a PNG tEXt chunk named "parameters", and ComfyUI can embed the whole workflow as JSON. Serious cleaning doesn't trust the chunk name: it enumerates every tEXt and iTXt chunk in the stream, compressed or renamed included, and strips them all on re-encode.
Clean a Stable Diffusion image
Yes. A smartphone photo's EXIF embeds by default GPS coordinates accurate to a few meters, the capture date and time, and your device's exact model. Posted as-is on a classified ad or forum, it can literally pin your home on a map. The location can be stripped in seconds, before publishing.
Guide: understanding EXIF and GPS data
Not reliably. Some platforms strip EXIF at upload, many don't — and either way, the original is stored on their servers before any processing. Email attachments, cloud links and marketplaces often keep metadata intact. Cleaning it yourself before sending is the only dependable guarantee.
Guide: selling safely on Vinted
Both systems offer partial native options: iOS lets you strip location at share time ("Options" > "Location"), Android varies by version and gallery app. Neither cleans everything (date, device, embedded thumbnails). For a complete clean, a dedicated tool that re-encodes the image is still needed — preferably local, with no upload.
No. There are two architectures: server upload (your file travels and may be stored) and 100% local in-browser processing via the Canvas API (the file never leaves your device). Anyone can verify the difference: open your browser's Network tab during cleaning — with local processing, no request carries your image.
Generally yes, for your own images and non-deceptive use: Adobe's, OpenAI's and Google's terms usually allow removing their metadata for downstream use. Context is tightening though: the EU AI Act, effective August 2026, mandates disclosure of AI-generated content. So check the rules of the platform you publish on — removing metadata is not a license to pass off an AI image as authentic.
Source: Regulation (EU) 2024/1689 — AI Act (EUR-Lex) · AI Act Explorer
Because its admission filters read metadata, not pixels. A growing number of stock, microstock and print-on-demand platforms scan the C2PA manifest's IPTC digitalSourceType assertion (which ChatGPT sets to trainedAlgorithmicMedia), the Software/parameters PNG chunks or the caBX box — and flag, demote or reject the upload on that field alone. Without those markers the file presents as neutral; pixel-based detectors remain independent.
Understand and remove those markers
Source: IPTC NewsCodes — digitalSourceType
Free for the essentials: cleaning a photo, inspecting its metadata and viewing GPS location cost nothing and require no account. Paid covers professional comfort: batch processing up to 50 images, one-click ZIP export and priority processing, at €5/month or €40/year. The core privacy function stays free.
By re-scanning it — never on trust. GhostMeta re-analyzes the cleaned file and shows a measured report: "N metadata tags detected → 0", GPS removed, C2PA removed. You can then cross-check on the official verifier at contentcredentials.org/verify, or reopen the file in any EXIF reader. A tool that won't let you verify it doesn't deserve your trust.