Direct answer
An AI tag on a photo is almost always metadata, not pixels: an EXIF/XMP field or a C2PA Content Credential that a camera app, editor, or generator wrote into the file, which platforms like Google Photos then read and display as an "AI-generated" label [1]. You remove it by stripping or rewriting that metadata — re-exporting through an editor, running a metadata cleaner, or taking a fresh screenshot — and by cropping away any visible watermark baked into the image [3]. What you cannot do is make the picture stop looking AI-made: metadata removal changes the label, not the pixels [4].
What Actually Creates the AI Tag on a Photo?
The tag is a data record, and it is written at three different layers — which is why removing one layer often leaves the label intact.
The first layer is ordinary EXIF and XMP. Fields such as Software, CreatorTool, and the IPTC DigitalSourceType property carry the AI signal; trainedAlgorithmicMedia is the value that explicitly declares an image was generated by a model [2]. These are plain text fields inside the file container, editable by nearly any metadata tool, and they are what most "AI generated" badges ultimately read [1].
The second layer is C2PA Content Credentials. These are cryptographically signed manifests attached to the file that record which tool produced it and what edits followed [2]. Because the manifest is signed, it cannot simply be edited without invalidating it — the signature check fails, and a broken manifest is itself a signal that something was changed [2]. This is the layer most people miss, and it is the reason a photo can still show an AI label after a normal "save as" [1].
The third layer is the visible watermark: the sparkle glyph, the "Made with Google AI" strip, or a corner badge composited into the pixels. That one is not metadata at all, which is why metadata cleaners do nothing to it [3].
How Do You Remove the AI Tag From a Photo Step by Step?
Work from the least destructive method outward, and check the result in the app that was showing the label before you assume it worked.
1. Re-export through an editor. Open the file in a desktop editor and export a new copy rather than saving over the original. A full re-export typically rebuilds the file container and drops most XMP and EXIF fields, though it does not reliably remove a signed C2PA manifest [3]. Flattening layers and exporting as a fresh JPEG or PNG is the single highest-yield step.
2. Strip metadata directly. A dedicated metadata tool removes EXIF, XMP, IPTC, and embedded thumbnails in one pass and is far more thorough than an editor's export [3]. Batch-capable command-line utilities handle whole folders, and GUI wrappers exist for people who do not want to touch a terminal. Verify afterwards by reopening the file's properties panel and confirming the fields are empty.
3. Crop or patch the visible watermark. If a badge is composited into a corner, cropping it out is clean and lossless to the rest of the frame; if it sits over important content, content-aware fill or a clone tool is the alternative [3]. Cropping changes dimensions, so re-check the composition before you post.
4. Screenshot as a last resort. A screenshot captures rendered pixels and carries no inherited metadata, which makes it the most reliable single-step removal [3]. The cost is real: you lose resolution, you lose the original color profile, and on high-DPI displays the result may be softer than the source.
5. Re-check in the original app. Labels are cached. Force-close the gallery app, clear its cache, or re-upload the new file under a different name, then confirm the badge is gone [1].
Does Removing the Tag Mean the Image Is No Longer Detected as AI?
No. Metadata removal changes what a platform displays; it does not change what a detector measures.
Detectors that work from pixels look for statistical fingerprints left by the generator — upsampling artifacts, frequency-domain regularities, and color-noise patterns that survive re-encoding [4]. Those signals live in the image data itself, so stripping EXIF or invalidating a manifest leaves them untouched [4]. An image can carry a perfectly clean metadata profile and still be flagged.
Detection accuracy also degrades in ways that cut both directions. Re-compression, resizing, and cropping measurably reduce detector confidence, which means the same edit that removes a tag can also push an image below a detection threshold — not because it became human-made, but because the evidence got noisier [4]. That is a reliability limit, not a loophole, and it is why no single detector should be treated as ground truth [4].
The practical distinction matters most in graded or professional contexts. A platform label is a disclosure convention; a detector verdict is an inference. Clearing the first does not settle the second, and in academic or publishing settings the disclosure expectation usually survives regardless of what the file's metadata says [2].
If the AI question you are actually trying to settle is whether a piece of work will be flagged, a photo tag is the wrong thing to chase — the same metadata-versus-detection gap applies to written submissions, where the AI score is computed from the text and not from anything you can strip out of the file. That is the gap turnitin0 closes: you upload your draft and get back the actual Turnitin AI detection report and similarity report, the same PDFs your institution sees, so you know the real number before you submit rather than guessing from a label.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Is it legal to remove an AI tag from a photo?
Removing metadata from your own file is generally lawful, but removing an AI disclosure in order to misrepresent origin can breach platform terms or publishing standards. The safer rule is to strip metadata when the tag is a nuisance and to keep the disclosure when the image's origin is material [2].
Why does the AI label come back after I upload the photo again?
The label is regenerated from the file's metadata each time it is read, so re-uploading the original restores it. Upload the cleaned copy instead, and force-close or clear the gallery app's cache, since labels are often cached locally [1].
Does a screenshot really remove the AI tag?
It removes the inherited metadata, because a screenshot is a fresh capture of rendered pixels. It does not remove a visible watermark that sits inside the frame, and it reduces resolution and color fidelity compared with the original [3].
Can I check whether a photo still carries AI metadata before sharing?
Yes — open the file's properties or info panel and look for Software, DigitalSourceType, or a Content Credentials entry. If those fields are populated, the tag will likely surface on platforms that read them [2].
Will a humanizer or rewriter help with an AI-flagged photo?
No. Rewriting tools operate on text, and pixel-level detection signals in an image are not something a text rewrite can touch [4]. For written work, the equivalent check is running the draft through an AI detection report before submission.