Calabi Labs · Guide · 2026-06-19
Ai text remover
When creators type "AI text remover" into Google, they're usually not looking for a tool to erase words pixel-by-pixel from an image. They're looking for a way to strip the invisible "made by AI" fingerprint that gets attached to their AI-generated files — the metadata layer that makes Instagram, TikTok, YouTube, and Reddit flag their content as synthetic. Calabi does exactly that: it removes the detection signals, injects authentic phone-capture identity, and shows you a forensic proof card before you download. It does not erase visible words from images or videos.
What Actually Gets Your AI File Flagged
Platforms don't read your content to decide it's AI-generated. They scan invisible metadata embedded in your file. The primary signal is a C2PA manifest — also called Content Credentials — stored as JUMBF atoms in the file structure. This cryptographic manifest says exactly which AI model generated the content, what version, and when. Tools like Midjourney, Sora, DALL-E, and Runway embed these manifests automatically, and platforms like Instagram and Reddit have been scanning for them since mid-2025.
Beyond the C2PA layer, there's an XMP metadata tag called DigitalSourceType: trainedAlgorithmicMedia — a plain-text flag that says the file came from a generative model. That's separate from C2PA and gets read by different scanners. Then there are encoder fingerprints: when you export an AI video, it carries a Lavc (libavcodec) or x264 SEI header that identifies it as machine-encoded rather than phone-captured. A file with no GPS coordinates, no capture timestamp, and one of these encoder fingerprints is a red flag that gets caught by perceptual hash comparisons too.
All of this exists before the file ever reaches a platform. The "AI text" you're trying to remove isn't visible text — it's a metadata layer embedded by the AI tool during export.
Why Cropping, Screenshots, and Re-Uploading Don't Work
Most creators try the obvious fixes first. Cropping removes the visible area but leaves all the metadata intact — the JUMBF atoms, the XMP tags, the encoder fingerprints survive a crop because they're stored at the file level, not in the pixel region. Screenshots re-encode the image, which strips some metadata, but platforms still catch the Lavc/x264 fingerprint in video, and the remaining XMP flags often survive a screenshot re-encode. Re-uploading through Telegram or Discord changes the file container but doesn't strip C2PA — many platforms re-attach the manifest from the original upload, and any missing GPS/timestamp is still a signal.
The visible "AI text" people sometimes see — a small Midjourney logo, a Sora sparkle icon in the corner — is a separate problem. Cropping removes it. Calabi doesn't erase logos or visible marks. What Calabi removes is the invisible detection layer that survives cropping and re-encoding — the layer that gets you flagged even when the visible content looks clean.
How Calabi Actually Cleans an AI-Generated File
Calabi runs a one-pass pipeline against your upload. Here's what happens:
- Upload your file — any AI-generated image or video, direct from Midjourney, Sora, DALL-E, Runway, Kling, or any other model. The free trial gives you 10 cleans with no credit card required.
- Automatic strip — the pipeline removes all JUMBF/C2PA atoms, reduces C2PA references to zero, strips the
DigitalSourceType: trainedAlgorithmicMediaXMP tag, removes generator/tool tags, and clears encoder fingerprints like Lavc and x264 SEI headers from video. A raw AI export carrying 144 metadata tags typically drops to around 94 neutral structural tags. - Automatic inject — the pipeline writes authentic phone-capture identity into the file: a real device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), realistic software version, GPS coordinates, capture timestamp, and a real-phone encoder name. This is what makes the file look like a normal phone recording to platform scanners.
- Forensic proof card — before download, you receive the same ExifTool scan that newsrooms and platform investigators use, showing exactly what was stripped and what was injected. You can verify it yourself.
- Download the cleaned file — ready to post without the AI detection layer attached.
FAQ
Does Calabi remove visible watermarks or logos from AI images?
No. Calabi doesn't edit pixels, use inpainting, or remove visible marks. If your AI export has a visible logo or sparkle icon in the corner, cropping removes it. Calabi removes the invisible detection metadata that survives cropping — the signals platforms scan for that you can't see.
Can re-encoding my video through HandBrake strip the AI metadata?
Partially. Re-encoding removes some metadata but often leaves the encoder fingerprint (Lavc/x264) and any remaining XMP flags. Platforms also use perceptual hashing — comparing your file's visual characteristics against known AI outputs — which a re-encode doesn't fully defeat. Calabi removes the metadata layer and injects device identity, addressing both the metadata scan and the encoder fingerprint problem.
Will this guarantee my post won't get flagged?
No tool can guarantee that. Platform detection systems evolve, and some use perceptual hashes that operate on the actual image/video content rather than metadata. Calabi removes the metadata and encoder signals that automated scanners rely on — the same layer ExifTool reads — which handles the majority of automated flags. Results vary by platform and source model.
Try Calabi free at calabilabs.com — 10 cleans, no card.