Your photo is real. You took it yourself, edited it for two minutes in Photoshop or Canva, posted it to Instagram, and now it wears a grey "AI content" badge. Facebook calls it "Made with AI." Your followers think you faked it.

The badge has nothing to do with the picture. It comes from hidden data inside the file. Once you understand what Instagram is actually reading, you can strip it out in seconds, in your browser, without uploading the photo anywhere.

What Instagram actually reads

Instagram, Facebook and TikTok do not judge your photo by looking at it. They read three hidden markers buried in the file:

1. C2PA Content Credentials. Photoshop, Canva, Lightroom and most phone editors embed a C2PA manifest (a JUMBF block) every time you export, even after a tiny crop or a color filter. The manifest records the edit history, including which AI tools touched the file. Instagram trusts it.

2. IPTC DigitalSourceType. A small tag in the photo header that records how the image was made. Editors set it to an AI value on export, and the platforms read it as a confession.

3. SynthID. Google's invisible pixel watermark. It lives in the pixels themselves, not the header, so it survives cropping, resizing and re-export. Images made or touched by Google AI tools carry it.

Any one of these is enough for the label. A real photo edited for two minutes can carry all three.

Why "real" photos get flagged

Most flagged photos were never AI-generated. A background cleanup, an object removal, a color filter — the editor writes a C2PA manifest into the export, and Instagram reads that manifest as proof of AI. The label is about the edit trail, not the image. That is why the same photo, exported from a different editor or with metadata stripped, posts clean.

Stripping them locally

Every EXIF and PDF scrubber I tried wanted me to upload my files to someone else's server. So I built one that runs entirely in the browser: CleanMyShot.

It does three things the naive tools miss:

PDFs lie about being clean. An incremental save appends a new revision but leaves the old /Info objects and metadata sitting in the file. Most scrubbers only touch the newest revision. CleanMyShot finds every occurrence of each object across all revisions and rebuilds the file without them.

Compressed object streams hide metadata. /ObjStm streams pack objects in Flate-compressed form, so a plain text scan misses them. The tool decompresses those streams in the browser, scrubs the objects inside, and recompresses.

PNG stripping breaks transparency. A naive chunk strip kills the tRNS chunk and your transparent PNG comes out with a broken or black background. This one uses a whitelist — it keeps tRNS, sBIT, cICP and APNG animation chunks, and drops the metadata chunks (C2PA, EXIF, IPTC, XMP, text chunks).

For photos, it strips the C2PA manifest and IPTC AI tags, scrambles the invisible in-pixel watermarks including SynthID, and rebuilds the image with natural camera grain and a real camera EXIF profile. The visible picture stays the same. The hidden signals that trigger the label are gone.

It also cleans ChatGPT text (zero-width characters, smart quotes, grey paste boxes) and scrubs Word and PDF metadata. Single HTML file, works offline once loaded, no signup, nothing uploaded.

The honest limit

Stripping metadata is not magic. A pixel-baked watermark that is truly burned into the visible image cannot be removed by any metadata tool, and no tool should promise otherwise. What this removes is the hidden trail: the headers, the manifests, the invisible pixel patterns. That is what the platforms read, and that is what gets your real photo posted without a badge it did not earn.

Try it on a flagged photo and see: CleanMyShot. If you find a file it does not clean properly, I want to hear about it.