How to remove AI detection from image?

You can remove most embedded AI metadata from an image in a few minutes. EXIF and XMP fields, PNG generation parameters, IPTC source tags, and C2PA Content Credentials all sit in the file’s data containers, apart from the visible pixels. A metadata cleaner deletes those containers. It does not remove invisible pixel watermarks such as Google’s SynthID, and it does not change what a visual classifier sees.

TL;DR:

You can use online AI metadata removal tools like https://deletefootprints.ai/ to quickly clean image metadata for free. This site doesn’t ask for an account, and you can clean the metadata of 5 images at once.

Knowing which signals are metadata and which are pixels tells you what cleaning can achieve. Clean metadata keeps your prompts, workflow settings, and GPS location private. It also stops many false “AI info” labels on photos you edited lightly with generative tools. It cannot make an AI image impossible to identify, and any tool that promises guaranteed undetectability is making a claim you cannot verify.

Which AI signals can actually be removed?

AI signals: editor reviewing portrait's removable metadata layers

Metadata layers are removable; pixel-level signals are not. AI-generation data can sit in five parts of a file, and a typical ChatGPT output carries identifying data in four of those five layers at once.

Signal	Where it lives	Removable with a metadata cleaner?
EXIF (Software, Make, Model)	File header	Yes
XMP and IPTC DigitalSourceType	XMP packet	Yes
PNG text chunks (prompt, seed, sampler)	tEXt, zTXt, iTXt chunks	Yes
C2PA Content Credentials	JUMBF box, APP11 (JPEG), caBX (PNG)	Yes, if the tool finds all of it
SynthID and similar watermarks	Pixels and frequency data	No
Generator patterns detected by classifiers	Pixels	No

The IPTC tag is small and easy to miss when you strip only EXIF. Its value, trainedAlgorithmicMedia, is a plain statement that generative AI made the file, and platforms read it alongside C2PA.

Stable Diffusion PNGs from AUTOMATIC1111 or ComfyUI expose the most. The parameters chunk holds the full prompt, negative prompt, seed, sampler, steps, and model hash. If your goal is protecting your creative recipe, that chunk is the first thing to remove.

How do you inspect and remove AI metadata?

Install Exiftool on your PC or Mac and run it from the command line.

Start with an inspection, clean a copy, and keep the original untouched. A three-step routine works for most files:

Inspect. Run exiftool -G1 -a -s image.png to list every tag with its group name. For PNGs, exiftool -PNG:all image.png shows the text chunks.
Choose selective or full removal. Selective removal deletes AI fields and keeps camera EXIF, copyright, and the color profile. Full removal deletes everything.
Export to a new file and leave the master alone.

For a selective clean, delete the JUMBF and C2PA groups, the PNG parameters chunk, and any XMP prompt fields. A full strip that keeps color looks like this:

exiftool -all= -tagsFromFile @ -ColorSpaceTags -overwrite_original image.jpg

In a first hypothetical case, a Stable Diffusion PNG with a prompt in its parameters chunk needs the PNG text chunks removed; camera data is irrelevant there.

In a second hypothetical case, a real photo touched with Photoshop Generative Fill carries Content Credentials, so selective removal suits it best because you keep the lens, date, and copyright fields while dropping the AI record. If that photo has GPS coordinates, decide on purpose whether to keep them.

If this process feels too technical, use the free server-side Exiftool cleaner at https://deletefootprints.ai/. It cleans metadata on its server and auto-deletes your files after 30 minutes, so it’s safe. I have personally verified that by checking the upload URLs.

Always use Exiftool-based metadata cleaners instead of browser-based ones. Depending on the website, browser-based cleaners can mess up the pixels, lower the quality, or cause other problems.

How can you verify the cleaned image metadata?

Verify by running the same inspection on the cleaned copy using the Exiftool command line and comparing it with the original. A cleaner’s “success” message tells you little; the tag dump tells you what is left.

If you can’t use the command line, use this page to inspect ai metadata online.

Check these points on the output:

No JUMBF, C2PA, or APP11 entries remain.
No DigitalSourceType field with trainedAlgorithmicMedia.
No parameters or prompt chunk in a PNG.
Width and height match the original exactly.
File size dropped only slightly, roughly the size of the removed blocks.

A large drop in file size points to recompression. Removing metadata alone does not change visible pixels when the tool avoids new lossy compression, so a 100% zoom comparison should show identical detail.

Partial removal can also break a credential. A C2PA manifest is cryptographically signed, and a tool that deletes part of it can leave a file with a broken credential, which gives a platform more reason for suspicion. Some tools also clean EXIF and XMP thoroughly while skipping the JUMBF box. Your recheck catches both problems. For a second opinion on C2PA, the Content Credentials verifier at contentcredentials.org/verify reads manifests, but it uploads your file.

Why can watermarks and AI detectors still flag an image?

Watermarks and detectors read the pixels, and metadata tools never touch the pixels. Google embeds SynthID into Gemini and Imagen output, and the watermark is built to survive resizing, compression, and re-encoding. Google’s SynthID Detector portal checks for it, and from May 19, 2026, SynthID detection began rolling into Google Search and Chrome.

Some services claim to remove SynthID by regenerating the image through a diffusion model. Treat those claims as unverified. Regeneration also rewrites your image, which defeats the purpose if you wanted the original preserved.

Classifiers are a separate problem. Detection firms train models on large sets of real and synthetic images so they learn the patterns each generator leaves behind. Hive states that its detection still works when metadata has been stripped or the content has been edited or re-uploaded.

Stripping can also add a signal of its own. A JPEG with no camera data at all reads as a file someone cleaned, since most cameras and phones write basic EXIF. Missing metadata therefore proves nothing in either direction: it does not show human authorship, and it does not show AI origin.

Will removing metadata change platform AI labels?

Metadata removal changes labels that come from metadata, and leaves other triggers in place. Meta has said its “AI info” label relies on technical metadata standards such as C2PA and IPTC. That explains why a real photo with one Generative Fill edit can get labeled, and why a clean selective strip often clears that label.

Platforms draw on several inputs:

Provenance metadata such as C2PA
Embedded watermarks such as SynthID
Your own disclosure toggle at upload
In-house classifiers that ignore metadata

TikTok auto-labels content when Content Credentials or its own watermark indicate AI, and YouTube relies on creator self-disclosure alongside provenance and classifiers. A file with no metadata can still be labeled through the other routes.

Conclusion

Metadata cleaning leaves pixel watermarks like SynthID and classifier signals in place. Before you upload, inspect the file, pick selective or full removal, clean a copy, and recheck the output for leftover JUMBF, IPTC, or PNG chunks. Then disclose AI use wherever the platform asks.