An image made by an AI generator looks like any other picture. The file, though, often tells on itself. Most generators write AI metadata into the image - records that name the tool that produced it, and frequently the exact prompt, seed and settings that were used. Sometimes that is useful provenance. Sometimes it is your rough draft, your prompt research, or simply none of anyone's business.
This guide covers what the major generators actually embed, how to remove it step by step with an AI metadata remover, and - just as important - what cleaning cannot do. Everything described here runs 100% client-side in your browser: your images are never uploaded to a server.
The four layers of AI metadata
There is no single "AI flag" inside a generated image. Depending on the generator, the traces land in up to four different places inside the file. A cleaner that only wipes classic EXIF fields will usually miss three of them. These are the four layers, in plain English.
1. C2PA content credentials
C2PA is the industry provenance standard backed by Adobe, Microsoft, OpenAI and others. A C2PA-enabled tool appends a signed manifest store to the file: claims about what created it, assertions describing the actions taken (such as "created with a generative AI model"), the ingredients used, and a cryptographic signature from the vendor. The manifest lives inside dedicated bytes - JPEG APP11 "JP" JUMBF segments, or a PNG caBX chunk - not in ordinary EXIF. Removing the visible metadata without touching the C2PA manifest leaves the most formal record behind. You can inspect exactly what a manifest says with our C2PA Credentials Checker.
2. IPTC DigitalSourceType
IPTC is the news-industry metadata standard for captioning and rights. One specific dataset, DigitalSourceType, uses a controlled vocabulary to describe how the image was produced - values such as trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMedia explicitly flag synthetic content. Some generators and editing apps write this value into the IPTC record (and C2PA references the same vocabulary). It is a small but formal "AI" marker that platforms and rights systems read.
3. PNG generation parameters
This is the most talkative layer, and it only affects PNG files. Interfaces built around Stable Diffusion write a plain-text PNG chunk - usually the tEXt chunk named parameters - containing the full prompt, the negative prompt, Steps, Sampler, CFG scale, Seed, image size, model hash and model name, sometimes down to the VAE and upscaler settings. ComfyUI goes further: its prompt and workflow chunks can embed the entire node graph as JSON, including every prompt and model file name on your machine. None of this is hidden or encrypted; any PNG parser can read it in milliseconds.
4. XMP packets with AI tool signatures
XMP is Adobe's metadata framework, embedded as XML in JPEG, PNG, TIFF and other formats. Editing apps and generators record the software name, edit history and document ID in XMP. When an XMP packet carries an AI tool signature - a generator or editor name that identifies the file as synthetic - the whole packet is the fingerprint. A selective (AI-only) clean drops that packet wholesale while leaving ordinary XMP alone.
A useful shorthand: layer 1 is a signed certificate, layer 2 is a controlled-vocabulary tag, layer 3 is a plain-text recipe, and layer 4 is the software name. Different generators hit different layers.
What each generator embeds (2026)
The details shift with every product update, but the patterns are stable. The table below maps the major generators to the layers they touch as of mid-2026. Only write-down what you can verify in a file - export paths change, so scan a real image rather than trust any summary blindly.
| Generator | C2PA manifest | IPTC DigitalSourceType | PNG parameters | XMP AI signatures | Notes |
|---|---|---|---|---|---|
| ChatGPT / DALL·E | Yes | Possible | No (JPEG) | Yes | Images delivered through ChatGPT and the OpenAI API ship with C2PA content credentials that mark the file as AI-generated. OpenAI backs the C2PA standard, so the manifest is the primary, formal signal. |
| Adobe Firefly | Yes | Yes | No | Yes | Firefly and Photoshop generative exports attach a C2PA manifest and set DigitalSourceType to a trained-algorithmic value. The most fully labeled of the major generators. |
| Midjourney | [unverified] | [unverified] | No | Yes | Midjourney exports identify the generator in XMP and software fields. Whether job identifiers and prompt fragments travel along depends on the export path - Discord download, web gallery or upscale - and changes as the product evolves. Scan the file rather than assume. |
| Stable Diffusion (Automatic1111 / Forge) | No | No | Yes (parameters) |
Sometimes | The most verbose generator. A single PNG tEXt chunk carries the full prompt, negative prompt, seed, sampler, CFG, steps, model hash and model name. Any PNG parser reads it in milliseconds. |
| ComfyUI | No | No | Yes (prompt + workflow) |
Sometimes | Goes further than Automatic1111: its prompt and workflow chunks embed the entire node graph as JSON, including every prompt and model file name on the machine. |
| Google Gemini / Imagen | [unverified] | [unverified] | No | Yes | Gemini-generated images identify the generator in XMP/software fields. Google also applies its own SynthID pixel watermark (see the limits section) - that is a separate, pixel-level signal and is not metadata. Whether C2PA is attached varies by surface and is not independently verified here. |
Verified entries reflect what the MetadataRemover scanner can parse in real export samples; [unverified] marks behavior we could not verify from a file or public spec and that may differ by export path.
How platforms label AI images
The reason most people want AI metadata gone is the label. Major platforms have begun flagging AI-generated uploads, and a "Made with AI" tag on a portfolio or product shot is something many creators want to avoid. How each platform decides to label is worth understanding, because it determines whether cleaning metadata even helps.
- Instagram (Meta). Meta applies "AI info" labels to images it believes were made with AI. The label can be triggered by C2PA content credentials and by Meta's own visual classifiers. Stripping metadata removes one trigger; it does not remove the other.
- Pinterest. Pinterest surfaces AI disclosures on pins it detects as synthetic. Detection relies on platform-side signals, not solely on file metadata, so a clean file is not a guarantee of any particular treatment.
- LinkedIn. LinkedIn marks AI-generated or AI-modified content with a label, and C2PA content credentials are one of the inputs it reads. As with the others, classifiers and heuristics also play a role, and behavior changes over time.
The common thread: platforms use both metadata and their own classifiers. Removing metadata clears the file's written record. It does not promise any particular label outcome - and the platforms change their rules faster than any guide can track.
Step by step: remove AI metadata
The whole process takes under a minute and happens entirely on your device. You have two options: strip only the AI provenance and keep the rest, or remove everything.
- Open the AI Metadata Remover and drop your image in - or paste from the clipboard. The file loads into browser memory only; no upload request is ever made.
- Read the scan. Findings are grouped by category, so AI metadata shows up explicitly: look for "AI Parameters", "C2PA", "XMP" and "IPTC" groups to see exactly what the generator left behind.
- Pick a mode. "AI-only strip" performs surgical, byte-level removal of AI provenance - C2PA/JUMBF, PNG
parameters/promptchunks, XMP packets with AI tool signatures, and IPTCDigitalSourceType- while preserving camera EXIF, GPS, copyright, ICC and the original pixels byte for byte. "Remove all metadata" re-renders the image on a fresh canvas and drops every metadata segment by construction. - Clean and verify. The output is scanned again so you can confirm zero AI findings remain, then saved. For AI-only mode the output bytes are otherwise identical to the input; for remove-all mode, choose PNG output for a mathematically lossless re-encode.
If your concern is specifically the provenance manifest, the Remove C2PA tool focuses on that single layer. If you want a full scan before deciding, the Image Privacy Analyzer reads EXIF, GPS, C2PA and hidden data in any file.
Verify before you publish
Trust, but check. Run a cleaned file back through the scanner, or drop it into the C2PA checker if provenance credentials were the concern - it should report no manifest found. If a platform still labels your upload as AI-generated after a clean scan, the label is coming from pixels, not metadata. Which brings us to the honest part.
What you CANNOT remove
- Pixel-level watermarks such as Google SynthID. These live inside the image's pixel data, not in its metadata. Stripping metadata leaves them fully intact.
- Visible watermarks and burned-in text. Anything drawn into the picture itself is part of the image, not metadata, and is not touched.
- A platform's own judgment. Platforms can label content using their own classifiers, regardless of metadata. Removing AI metadata does not promise any particular treatment by any platform.
Invisible watermarks are not metadata. Systems like Google's SynthID weave a statistical pattern into the pixel values themselves. It survives cropping and recompression, and no metadata cleaner can remove it - be skeptical of any tool that claims otherwise. Separately, AI-image detectors classify the pixels, not the file's records, so scrubbing metadata does not change their verdict either way.
There is also a judgment call that no tool can make for you. Removing production details from your own images before publishing is ordinary privacy hygiene - photographers have stripped EXIF for decades. But stripping provenance credentials in order to pass AI content off as human work can breach platform rules, and disclosure requirements for synthetic media are beginning to appear in legislation. Clean for privacy, not for deception.
When to KEEP AI metadata
- When disclosure is required. Some platforms, marketplaces and regulations require AI-generated content to be labeled. Removing the label there can break those rules.
- When you need to prove origin. Content credentials and AI tags are evidence of how an image was made - useful for licensing, copyright questions and disputes.
- When credibility matters. In journalism, research and legal contexts, provenance is the chain of custody. Stripping it undermines the file as evidence.
Legal and responsibility
For your own files, removing metadata is generally lawful - metadata itself is not specially protected in most jurisdictions. What can break the law or platform rules is what you do next: misrepresenting the origin of content where disclosure is required.
Removing C2PA content credentials is not the same as removing a watermark. C2PA Content Credentials are metadata attached to the file. Watermarks such as SynthID are embedded in the pixel data itself. A metadata tool removes the former; it cannot and does not touch the latter. Stripping your copy also changes nothing elsewhere - platforms that already ingested the file keep whatever they recorded.
Remove credentials only from files you own or have the right to modify. Where a platform, marketplace or law requires provenance disclosure, keep it. See the terms of use for the full picture. The file is your responsibility.