Trend report · lobsters · 2026-06-17
Developers building AI content tools are learning a hard lesson: the generated images and videos their tools export carry a invisible fingerprint that platforms like Instagram and TikTok flag within seconds. It's not about the visual content — it's about the metadata layer underneath.
When you export an image or video from an AI generator, it arrives wrapped in a forensic trail that platform scanners read like a confession. Here's what's actually in there:
C2PA / Content Credentials — The "Made by AI" manifest stored as JUMBF atoms. This is a cryptographic manifest that explicitly declares the file was AI-generated, listing the model, version, and generation parameters. A single AI export can contain 18+ JUMBF atoms and 16+ C2PA references that all point back to the generation source.
XMP AI flags — Fields like DigitalSourceType: trainedAlgorithmicMedia embedded directly in the XMP metadata. Some exporters also include Generator, Software, or CreatorTool tags that name the AI model outright.
Encoder fingerprints — Video files carry encoder signatures in the bitstream. Lavc (FFmpeg's libavcodec) and x264 SEI messages are dead giveaways. A phone recording doesn't have these. An AI export does.
Missing phone identity — Real phone captures include Make, Model, Software version, GPS coordinates, and capture timestamp. AI exports have none of this. Platforms treat "unknown device" as a soft signal — suspicious but not damning on its own. Combined with C2PA flags and Lavc fingerprints, it's a confirmation.
A raw AI export from Midjourney, Sora, or Runway can carry 144+ metadata tags. Platform scanners parse this in under a second, compare it against known AI-generation patterns, and flag before the content ever appears on your profile.
Calabi runs a three-stage pipeline that strips the incriminating signals and injects authentic phone-capture identity — all in one pass.
Stage 1: Strip — Calabi removes every C2PA / Content Credentials manifest, reduces JUMBF atoms from 18+ to 0, strips DigitalSourceType: trainedAlgorithmicMedia and all generator/tool tags, and clears encoder fingerprints like Lavc and x264 SEI from video bitstreams. The result: 144 tags down to about 94 neutral structural tags.
Stage 2: Inject — Calabi writes real phone identity into the file: iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra profiles with authentic Make, Model, Software version, GPS coordinates, and capture timestamp. The device profile is complete enough to pass forensic inspection.
Stage 3: Verify — Before download, you receive a forensic proof card — the same ExifTool scan platforms use — showing exactly what was stripped and what was injected. You see what Instagram, TikTok, and YouTube will see.
What about visible watermarks like Sora's sparkle or a corner logo?
Cropping removes the visible mark. Calabi handles the invisible layer — the C2PA manifest, AI metadata, and encoder fingerprints — that survives cropping and is what platforms actually scan for.
Can platform detection be fully guaranteed?
No tool can guarantee a platform won't flag you. Results vary by platform and source model. Calabi removes the documented forensic signals — C2PA, XMP AI flags, encoder fingerprints — that automated scanners rely on. A clean file with phone identity and no AI manifest is significantly harder to flag than the raw export.
Does this work on video and images?
Yes. Both formats carry C2PA manifests, XMP metadata, and video-specific encoder fingerprints. Calabi processes both through the same strip-and-inject pipeline.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.