What is the AI Metadata Cleaner?
The AI Metadata Cleaner is a free AI label remover and AI metadata remover for photos. It removes the AI info, AI tags and C2PA Content Credentials stored inside a photo, the records platforms such as Instagram read for their AI info label. It runs in your browser, so the photo is never uploaded.
Use it to remove AI metadata from a photo before you post it: a real photo that picked up an AI tag after a small edit in Photoshop or Google Photos, or an image made with ChatGPT, Gemini or Midjourney. As an AI tag remover it shows every tag first, so you see what the photo holds before anything is removed. It takes JPEG, PNG, WebP, HEIC files on a computer or a phone, with no sign-up and no daily limit.
What does the AI Metadata Cleaner remove from a photo?
The AI Metadata Cleaner removes the AI info and C2PA Content Credentials stored inside a photo: the metadata platforms such as Instagram read for their AI info label. It also removes the EXIF, IPTC and XMP details of where, when and with what the photo was taken, and keeps only what the photo needs to look the same.
What it removes
-
AI info
The IPTC digital source type (trainedAlgorithmicMedia), AI generator names in the software and creator fields, and the prompts and settings that tools such as Stable Diffusion and Midjourney save in the file.
-
Content Credentials
The signed C2PA manifest that records which app or AI model made the image and how it was edited. The whole manifest goes, signature included.
-
GPS location
GPS coordinates and altitude, plus any place names written into the file.
-
Camera
Make, model, lens, serial numbers and maker notes. On an iPhone photo, also the lens calibration data that can identify the camera.
-
Date taken
When the photo was taken, digitised and last changed, with the time zone.
-
Software
The app that edited the photo, its version and the edit history saved in XMP.
-
Author
Artist, copyright, contact details, captions, keywords and comments.
-
Hidden extras
In a JPEG: the preview thumbnail, a Motion Photo's video clip, other images stored after the picture and the Ultra HDR gain map, so an HDR screen shows the standard picture.
-
Everything else
Any other EXIF, IPTC or XMP tag and any PNG text, known or unknown. The cleaner keeps a short list and removes the rest.
What it keeps, and why
-
Colour profile
So the colours display as they did before.
-
Orientation
One number that keeps the photo the right way up. It says nothing about you.
-
Pixels
Copied byte for byte. Nothing is re-encoded, so there is no quality loss.
-
File structure
The format details a viewer needs to open the file, such as its width and height.
Not removable: a watermark hidden in the pixels, such as SynthID, because it is part of the image, not its metadata.
To see what your own photo holds, pick it above and open View metadata. Every tag shows before anything is removed.
How do I remove AI metadata from a photo?
To remove AI metadata from a photo, add it to the AI Metadata Cleaner, check what it holds, press Remove and save the clean copy. The tool deletes the AI info and C2PA Content Credentials that platforms such as Instagram read for their AI info label, along with the EXIF, IPTC and XMP data, and leaves every pixel as it was. It is free, needs no sign-up and runs in your browser, so the photo is never uploaded.
-
Add your photos to the AI Metadata Cleaner
Add up to 5 photos at once, 256 MB in total, by dropping them on the AI Metadata Cleaner or pressing Choose Photos. It accepts JPEG, PNG, WebP, HEIC files. To see how it works first, press Try a sample photo: it runs the same steps on a photo deliberately tagged with AI metadata and Content Credentials.
-
Check each photo for AI metadata
The AI Metadata Cleaner reads each photo's metadata and gives it one of three verdicts: “AI metadata detected”, “Content Credentials · no AI marker” or “No AI metadata detected”. It looks for AI markers such as the IPTC digital source type trainedAlgorithmicMedia, AI generator names, saved prompts and C2PA Content Credentials. The check reads metadata only, not the pixels. Press View metadata to see every C2PA, EXIF, IPTC and XMP tag the photo holds before anything is removed.
-
Remove the AI metadata and Content Credentials
Press the Remove button under your photos to delete the AI info, the C2PA Content Credentials and every other EXIF, IPTC and XMP tag. Only the colour profile and orientation stay, so the photo looks the same. The pixels are copied unchanged, with no re-encoding and no quality loss. Each cleaned copy is checked again before you can save it, and a photo that still holds a removable tag shows an error instead of a download.
-
Save the cleaned photos
Press Save cleaned photo, or Download on each photo when you cleaned several. Where your browser can share files, Save all sends every copy to the share sheet at once. Each copy keeps its format and its file name, with -cleaned added, such as photo-cleaned.jpg. The cleaned copies live only in this browser tab, so save them before you close it.
Remove AI Metadata From Images: Delete C2PA, IPTC, and Prompt Data in Your Browser
AI metadata is provenance data that AI generators and editing apps write into an image file to record that AI made or changed it. The AI Metadata Cleaner on Remove AI Metadata finds those records, removes them in your browser, and leaves the pixels untouched. Provenance means the history of a file: what made it and what changed it.
Last tested: 14 November 2026 by Ipathy Srinu, photographer and builder of Remove AI Metadata.
What Is AI Metadata? The Hidden Records Behind the AI Label
AI metadata is information stored inside an image file that says AI made or changed the picture. AI generators and editing apps write it on their own, and platforms such as Instagram read it when they decide whether to show an "AI info" label. It sits in the same part of the file as the camera model, the date, and the location, apart from the pixels you see. Nothing in the picture shows it, but any metadata reader can print it.
People call it by several names. Content Credentials is the public name for a signed C2PA manifest, the most detailed AI record. "AI info" and "AI tags" are the everyday names for the fields platforms read. Provenance data is the formal term: the history of a file, what made it and what changed it.
Why do AI tools add metadata to images?
AI tools add metadata so that apps and platforms can tell an image was made or edited with AI without guessing from the picture. IPTC, the news industry body behind the photo metadata standard, recommended a value for AI generators in May 2023, and C2PA built a signed record for the same job. OpenAI has attached Content Credentials to its images since 2024, and Google adds them to Gemini images. Editing apps write a record too: Photoshop and Google Photos mark a photo when a generative tool changes any part of it, even a small one.
How does AI metadata lead to the AI info label on Instagram?
Instagram reads the C2PA and IPTC records in a photo when you upload it, so a real photo whose records say AI edited it can get an "AI info" label after a small retouch. Meta said in February 2024 that it reads both standards and named Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock as sources of labelled images. For a photo labelled only because an editor wrote an AI record, removing that record is what helps.
Images made entirely by AI are a different case. Instagram also detects those from the picture itself, for example from a SynthID watermark in the pixels, so removing the metadata does not change their label.
Is AI metadata the same as an AI watermark?
No. AI metadata is a record stored beside the pixels, while a watermark such as SynthID is a signal hidden in the pixels themselves. A metadata reader prints the record, and a metadata remover deletes it without touching the picture. A pixel watermark survives that, because the AI Metadata Cleaner copies the pixels unchanged. Google adds SynthID to Gemini images, and OpenAI added it to ChatGPT images on 19 May 2026.
AI Metadata Types: The Four Records That Label an Image as AI Made
An image can carry four kinds of AI metadata: a C2PA manifest, an IPTC Digital Source Type value, generator fields in EXIF or XMP, and prompt text written by AI tools. The first two are standards built to declare AI use. The last two are side effects: a tool writing its own name into the file, or saving its settings for later.
What is a C2PA manifest (Content Credentials)?
A C2PA manifest is a signed record, stored inside the file, that lists who made an image and which actions changed it. C2PA stands for the Coalition for Content Provenance and Authenticity, the group that writes the standard. Adobe, Arm, the BBC, Intel, Microsoft, and Truepic founded it in 2021. Content Credentials is the public name for the same record.
Signed means the tool sealed the record with its private key, so a checker can tell if the record or the picture changed afterwards. An AI generator declares itself inside the manifest: it records a created action and tags it with the source type value from the next section. The C2PA specification site lists version 2.5, released in October 2026, as the newest. Version 2.2 came out in May 2025 and version 2.4 in April 2026.
What does IPTC Digital Source Type "trainedAlgorithmicMedia" mean?
trainedAlgorithmicMedia is the IPTC value that tells software an image was made by a trained AI model. IPTC is the International Press Telecommunications Council, the news industry body behind the photo metadata standard. Digital Source Type is the one field in that standard that says how an image came to exist: from a camera, a scanner, or a model.
The value is stored in XMP, a text format from Adobe that holds metadata as XML inside the file. ExifTool shows the field as XMP-iptcExt:DigitalSourceType, and the full value is this URI:
http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia
IPTC recommended this value for AI generators on 9 May 2023. A sibling value, compositeWithTrainedAlgorithmicMedia, joined the vocabulary on 26 July 2023. It marks a real image with AI parts, such as a photo changed with Photoshop Generative Fill. In late 2024 IPTC gave both values plain display names: "Created using Generative AI" and "Edited using Generative AI". The IDs stayed the same. C2PA reuses this vocabulary inside its actions, so one image can state the same fact twice: once in XMP and once in the manifest. The Google Gemini files I tested store it in both places.
Which EXIF fields and PNG text chunks carry AI generator names and prompts?
Generator names usually sit in the EXIF Software or IPTC Credit field, and many AI tools write the full prompt into PNG text chunks. EXIF is the block where cameras record their settings, and it has a Software field that any app can fill. A chunk is one labelled block in a PNG file, and a text chunk holds a keyword and a string.
The Gemini images I tested carry "Made with Google AI" in the IPTC Credit field. Automatic1111, a widely used Stable Diffusion interface, writes one tEXt chunk keyed parameters. That chunk holds the prompt, the negative prompt, the seed, the sampler, and the model. For JPEG and WebP it writes the same text into the EXIF UserComment field (source code). ComfyUI writes two tEXt chunks: prompt holds the inputs that ran, and workflow holds the whole node graph.
Midjourney uses plain text chunks too. This is the Midjourney image you can download further down, with the prompt trimmed:
$ exiftool -a -G1 -s ai-generated-before.png
[PNG] CreationTime : 2026:09:22 13:41:21+00:00
[PNG] Author : Elizabeth
[PNG] Description : Young adult blonde woman walking a lively Manhattan street, carrying a wooden violin. [...] --ar 9:16 --raw --hd --v 8.2 Job ID: 5dd747f8-47d6-4c71-ace7-80f2aa1ab80e
[XMP-iptcExt] DigitalImageGUID : 5dd747f8-47d6-4c71-ace7-80f2aa1ab80e
[XMP-iptcExt] DigitalSourceType : http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia
All 4 Midjourney files I tested had the same layout: four tEXt chunks keyed Creation Time, Author, Description, and XML:com.adobe.xmp. The Description chunk held the full prompt, the settings, and the job ID, and the Author chunk held the account name. The longest Description chunk was 1,134 bytes.
Where does AI metadata sit inside JPEG, PNG, WebP, and HEIC files?
Each file format stores AI metadata in its own container block, separate from the compressed pixels, which is why it can be removed without touching the picture. The C2PA manifest travels in JUMBF, a box format the JPEG committee built to carry extra data. Each format gives that box a different home.
| Record | JPEG | PNG | WebP | HEIC |
|---|---|---|---|---|
| C2PA manifest (JUMBF) | APP11 segments | caBX chunk |
C2PA chunk at the end of the file |
C2PA uuid box |
| IPTC Digital Source Type (XMP) | APP1 XMP packet | iTXt chunk keyed XML:com.adobe.xmp |
XMP chunk |
XMP item in the meta box |
| Generator fields (EXIF) | APP1 Exif block | eXIf chunk |
EXIF chunk |
Exif item in the meta box |
| Prompt text | EXIF UserComment or COM marker |
tEXt, iTXt, or zTXt chunk |
EXIF UserComment |
Rare |
A JPEG segment holds at most about 64 KB, so a larger manifest spans several APP11 segments in a row. WebP is built on a container layout called RIFF and HEIC on one called ISO BMFF, and those layouts name their blocks chunks and boxes. The C2PA row follows the embedding appendix of the C2PA specification. Midjourney puts its XMP in a tEXt chunk rather than iTXt, and the cleaner removes both.
The diagram shows a test JPEG I tagged with AI metadata and a signed C2PA manifest. The manifest took 17,351 bytes in one APP11 segment, ahead of the picture data. After cleaning, the 54,415 bytes of picture data were the same as before.
AI Metadata by Generator: What Each AI Tool Writes Into Your Image
Every AI image I tested from ChatGPT, Google Gemini, and Midjourney carried at least one AI metadata record, and each generator writes a different mix. The table also shows which files declare a pixel watermark, a faint pattern mixed into the picture itself. SynthID is the one Google DeepMind built.
OpenAI has attached Content Credentials to its images since 2024 and added SynthID on 19 May 2026 for images from ChatGPT, Codex, and its API. Google began embedding C2PA metadata in Nano Banana Pro images in November 2025, on top of SynthID. Meta writes IPTC metadata into Meta AI images and adds its own invisible watermark. Local Stable Diffusion tools write neither standard. They leave only the prompt chunk.
What each file showed when I read it on 14 November 2026:
| Generator | C2PA | IPTC Digital Source Type | Generator field or prompt | Pixel watermark |
|---|---|---|---|---|
| ChatGPT (OpenAI gpt-image) | Yes, claim generator "OpenAI Media Service API" | trainedAlgorithmicMedia, inside the manifest only | Software agent "gpt-image" in the manifest | Declared in the manifest as a c2pa.watermarked.unbound action |
| Google Gemini (2 files) | Yes, claim generator "Google C2PA Core Generator Library" | trainedAlgorithmicMedia, in XMP and in the manifest | Credit "Made with Google AI" | Declared in the manifest: "Applied imperceptible SynthID watermark." |
| Midjourney v8.2 (4 files) | No | trainedAlgorithmicMedia in XMP | Full prompt, settings, and job ID in Description, account name in Author | None declared |
The watermark column shows what each file declares about itself. After cleaning, all 7 files had no C2PA manifest, no source type, and no prompt, and their pixel hashes matched the originals.
You can download one of the Midjourney pairs and compare them:
- Midjourney image, original (4,978,491 bytes)
- Midjourney image, cleaned (4,976,717 bytes)
The original holds the prompt, the account name, the job ID, and the AI source type.
Why does a real photo get an "AI info" label after light editing?
Editors such as Photoshop and Google Photos write AI metadata when a generative tool touches any part of a photo, and platforms read that record as AI use. The record is the "Edited using Generative AI" source type. Google Photos adds it, with the Credit "Edited with Google AI", after Magic Editor, Magic Eraser, and Zoom Enhance edits.
Meta renamed its "Made with AI" label to "AI info" on 1 July 2024 after labels based on these records landed on photos with minor AI retouching. In September 2024 it moved the label for posts with AI edits into the post menu. Fully generated images kept the label on the post itself.
Photoshop has a further issue. Users on the Adobe Community forum have reported since October 2023 that the AI credential stays in the export after they undo the edit. A bug report from April 2026 says the label stays even after the AI layers are deleted. These are user reports, not Adobe documentation.
A June 2026 guide from the Canadian Association for Photographic Art gives matching advice for Photoshop 27.9. Undo is not enough, so close the file without saving and reopen the last clean save. The guide lists the Remove tool with generative AI turned off, the Spot Healing Brush, and the Clone Stamp as tools that add no AI declaration. The Remove tool left on Auto can add one.
Do AI prompts and workflows leak through image metadata?
Yes. Midjourney, Automatic1111, and ComfyUI save the prompt inside the PNG, and the two local tools also save the seed, the model, and the workflow unless you turn it off. The text stays in the file when you email it or share it from a drive. A client brief pasted into a prompt, an unreleased product name, or the name of a private LoRA (a small extra model file) can be read by anyone who opens the PNG in a metadata reader. The ComfyUI workflow chunk is enough to rebuild the whole workflow.
Both local tools have an off switch. In ComfyUI Desktop, open Settings, then Server-Config, and turn on "Disable saving prompt metadata in files", or start ComfyUI with --disable-metadata. In Automatic1111, turn off the setting "Write infotext to metadata of the generated image". For files that already exist, removal is the fix.
AI Metadata Cleaner FAQ
How do I remove the AI info label from an Instagram post?
Remove the AI metadata from the photo before you post it, because the AI Metadata Cleaner cannot change a post that is already on Instagram. For a post that already shows AI info, delete it, clean the photo here and post the clean copy. Instagram reads the C2PA and IPTC records in a file when you upload it, and a cleaned copy carries neither. Removal is not a guarantee: Instagram also detects images made entirely by AI from the picture itself, such as a SynthID watermark.
Can the AI Metadata Cleaner remove AI detection from an image?
It removes the AI metadata that detection tools read from the file, such as C2PA Content Credentials and the IPTC digital source type, but not what a detector sees in the pixels. The pixels are copied unchanged, so a detector that judges the picture itself, or a watermark hidden in it such as SynthID, gives the same result as before.
Can I remove only the AI metadata and keep the rest?
No. The AI Metadata Cleaner removes all metadata except the colour profile and orientation. It has no mode that keeps copyright, camera or location fields. Those details stay in your original photo, so hold on to it.
Does cleaning change my original photo?
No. The cleaner works on a copy in your browser, so the original file on your device stays exactly as it was. Its Content Credentials stay in it too. Keep the original if you may need that record later, because a cleaned copy cannot get it back.
Why does an AI image show "No AI metadata detected"?
The file no longer holds any AI metadata. A screenshot leaves none behind, and so does an app or website that saves a new copy without the metadata. The verdict comes from the file's records, so it cannot tell from the picture alone that AI made it.
Is there a daily limit?
No. There is no account, quota or daily cap. After a batch, press Clean more photos and add the next one.
Does it work on iPhone and Android?
Yes. Open this page in your phone's browser, tap Choose Photos and pick photos from your library. The photos are cleaned on the phone itself.
Disclaimer
The AI Metadata Cleaner removes metadata stored inside the file, in your browser, and never uploads the photo. It cannot remove a watermark hidden in the pixels (such as SynthID), a copy of the Content Credentials kept outside the file, or a label a platform has already applied. It also does not change what an AI detector sees in the image. We do not promise how any platform or detector will treat a cleaned photo: you remain responsible for each platform's disclosure rules and the law where you live.