If you have ever used Photoshop's Generative Fill to remove a stray signpost from a vacation photo, you might have been surprised to see Instagram slap an "AI info" label on your post. This isn't a glitch. It is a feature. The signal triggering the label is rarely in the visible pixels of the image itself; it is buried in the file's metadata. Meta reads industry-standard markers embedded by editing tools, treating a minor generative fix the same as a fully AI-generated image. Understanding where this data lives and how to clean it is essential for photographers who want their work judged on its visual merit, not its edit history.

What Triggers the Label

Meta explicitly states that it looks for three specific types of metadata markers. The first is C2PA Content Credentials, a signed manifest describing the image's creation process. In JPEG files, this lives in JUMBF boxes inside APP11 segments, and tools like Adobe Firefly and Photoshop generative features write this automatically. The second is the IPTC DigitalSourceType field, which categorizes the image as either fully generated or a composite of real and generated parts. Finally, XMP history records the specific tools and steps used, leaving a clear trail for editors like Generative Fill. If any of these are present, Instagram assumes the image needs an AI disclosure.

Checking and Stripping Metadata

Before uploading, you should inspect the file using ExifTool. Running the command exiftool -G1 -a -s photo.jpg | grep -i -E 'c2pa|jumbf|DigitalSourceType|Credit' will reveal if any markers are present. If you find them and wish to remove them, you can strip all metadata with exiftool -all= -o clean.jpg photo.jpg. However, this is a blunt instrument that also removes copyright info, GPS data, and ICC color profiles, which can cause color shifts or orientation issues. A safer approach is to copy the color profile back using -tagsfromfile @ -icc_profile. For those who want to keep most metadata but remove only the C2PA manifest, dedicated on-device tools exist that target just that specific data structure.

Troubleshooting Persistent Labels

If the label still appears after cleaning, run the check command on the exact file you uploaded. If markers remain, you may have uploaded the wrong copy. If the output is empty but the label persists, Meta likely used another signal, such as pixel-based detection. Additionally, be aware that saving or exporting an image often brings metadata back. Phone galleries and many editing apps rewrite XMP data on export, and Photoshop or Lightroom can write a new manifest if the edit history still includes a generative step. Therefore, you must strip metadata as the final step before uploading, or re-check after every save. Screenshots can remove metadata but also resample the image and may not remove pixel watermarks, so they should be treated as a last resort.

Limitations and Best Practices

It is crucial to remember that stripping metadata is not a magic bullet. Some generators, such as Google's SynthID, embed invisible watermarks directly into the pixel data, which metadata cleaning cannot remove. Meta also employs its own detection algorithms that may identify AI content regardless of the file tags. Therefore, if an image is substantially AI-generated, ethical disclosure is still the best practice. Cleaning metadata is most appropriate when the AI label is a false positive caused by a minor touch-up on a real photograph.

Key Takeaways

  • Instagram's "AI info" label is triggered by metadata markers like C2PA, IPTC, and XMP, not just visible AI content.
  • Use ExifTool to check for specific tags before uploading to see if your file is flagged.
  • Stripping all metadata with -all= can remove important color profiles and copyright info, so selectively restore ICC profiles.
  • Invisible pixel watermarks (e.g., SynthID) and Meta's own detection methods may still trigger labels even after metadata cleaning.
  • Always strip metadata as the last step before upload to avoid editors rewriting the data during final saves.

The Bottom Line

While stripping metadata is a necessary technical fix for minor generative edits, it is not a license to hide substantial AI generation. The label exists for disclosure, and ethical practice demands honesty when the AI contribution is significant, regardless of what the file tags say.