Trend report · gnews_detection · 2026-06-19
DeepDect: an explainable AI platform for face swapping and face generation DeepFake detection - Springer Nature Link
DeepDect Just Made AI Detection Better — So Did the Platforms
Springer Nature published DeepDect, an explainable AI platform that classifies face-swapping and face-generation DeepFakes with confidence scoring. That's real progress on the detection side. But here's what it means for creators: the detection tools are getting sharper, and so are the platforms that automatically scan uploads. In 2026, Instagram, TikTok, YouTube, and Reddit don't need a DeepDect to flag your file — they read the metadata layer underneath.
What Actually Flags Your File
When you upload an AI-generated video, platforms run an automated pipeline that looks for three things before your content ever reaches a human moderator or a tool like DeepDect.
C2PA / Content Credentials. This is the cryptographic manifest embedded in your file — stored as JUMBF (JPEG Universal Metadata Box Format) atoms. It says, in machine-readable form, "this was generated by an AI model." A raw Sora or Runway export typically contains 18+ JUMBF atoms and 16+ C2PA references. Platforms parse these automatically and flag the upload within seconds.
XMP AI flags. Beyond the C2PA block, XMP metadata fields like DigitalSourceType: trainedAlgorithmicMedia are embedded directly into the file. This flag was added by the generator software and survives re-encoding in many cases. A raw AI export can carry 144 metadata tags; a significant subset are AI-origin signals that forensic tools read as a "made by AI" fingerprint.
Encoder fingerprints. The encoder used to compress your video leaves traces. Lavc (FFmpeg's libavcodec) and x264 SEI (Supplemental Enhancement Information) messages are common in AI exports. These aren't visible to the viewer, but they're invisible signals that tell platforms "this was processed by a specific software pipeline, not a phone."
Missing device identity. A real phone recording has Make, Model, Software version, GPS coordinates, and a capture timestamp. An AI export has none of that. The absence of these fields is itself a signal.
How Calabi Handles It
Calabi is a one-pass web tool that strips the detection layer and injects authentic phone-capture identity in its place. It works on the invisible signals, not the pixels. You won't see a before/after change in how your video looks — but forensic tools will read the output as a normal phone recording.
Stage 1 — Strip. Calabi removes all JUMBF/C2PA atoms, XMP AI flags including DigitalSourceType: trainedAlgorithmicMedia, generator/tool tags, and encoder fingerprints like Lavc and x264 SEI. A raw AI export's 144 metadata tags compress down to roughly 94 neutral structural tags — no AI origin signal.
Stage 2 — Inject. Calabi writes real phone-capture identity into the file: a device profile (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), corresponding software version, GPS coordinates, and a capture timestamp. It also adds a legitimate phone encoder name. The file now reads identically to a direct phone upload.
Stage 3 — Verify. Before download, Calabi generates a forensic proof card — the same ExifTool scan that platforms use. You see exactly what was stripped (18 JUMBF atoms → 0, 16 C2PA references → 0, the trainedAlgorithmicMedia flag) and what was injected (device profile, GPS, encoder). You verify before you trust.
The Steps
- Upload your AI-generated video or image to calabilabs.com. No account required for the free trial.
- Automatic pipeline runs. Strip → Inject → Verify happens in one pass. No manual editing, no selecting regions, no inpainting.
- Review the forensic proof card. See the exact metadata changes before you download — what was stripped, what was injected, verified with ExifTool.
- Download the cleaned file. Use it on Instagram, TikTok, YouTube, or Reddit. The file-level identity reads as a phone recording.
FAQ
What about visible watermarks like Sora's sparkle or a corner logo?
Calabi doesn't erase pixels or remove visible marks. Cropping the visible watermark removes what viewers see — and Calabi removes the invisible detection metadata that survives cropping and re-encoding. The visible mark is a viewer problem; the metadata is the platform's automated-detection problem. Both matter.
Can platforms still detect my content after processing?
No tool can guarantee a platform won't flag you — detection methods evolve. Calabi fully removes C2PA, XMP AI flags, and encoder fingerprints. Re-encoding disrupts some invisible pixel patterns but results vary by platform and source model. Calabi removes the metadata layer that automated systems read, which is what triggers the initial scan.
Does this work on images or just video?
Both. The strip-and-inject pipeline works on AI-generated images and video files. The forensic proof card shows the before/after metadata for either format.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.