Trend report · gnews_detection · 2026-06-18
Preity Zinta gets HC nod to pursue action against alleged AI-misuse, deepfake content - The Indian Express
When Preity Zinta petitioned the Himachal Pradesh High Court over AI-generated deepfakes using her likeness, she wasn't fighting pixel-level manipulation — she was fighting metadata. The images looked real, but the invisible forensic signals told a different story. That's where platforms catch you.
What actually flags your file
In 2026, platforms don't primarily look at what an image looks like — they scan what an image claims to be. Your file carries a hidden dossier of metadata that forensic tools extract in seconds. Here's what's actually being checked:
C2PA / Content Credentials is the cryptographic manifest embedded in JPEG and video files as JUMBF (JPEG Universal Metadata Box Format) atoms. When you export from Sora, Runway, Midjourney, or Kling, these tools write a C2PA block that says, in machine-readable terms: "This content was generated by an AI model." A 2026 ExifTool scan on a raw AI export typically shows 18+ JUMBF atoms and 16 C2PA references. Platforms like Instagram and TikTok now reject or shadowban files with active C2PA manifests.
XMP AI metadata lives one layer below C2PA. The field Iptc4xmpExt:DigitalSourceType gets set to trainedAlgorithmicMedia by most generative tools. Adobe products add photoshop:History entries referencing AI generation. Even if you strip C2PA, an XMP scan can reveal Generator, Software, or AIModel tags that give it away.
Encoder fingerprints are the sneakiest signal. Video exports carry codec metadata in the bitstream itself. Lavf (libavformat) and Lavc (libavcodec) write SEI (Supplemental Enhancement Information) NAL units into H.264/H.265 files. x264 and x265 encoders leave their own NAL signatures. A phone-recorded video from an iPhone 16 Pro uses the Apple encoder, not Lavc. That mismatch — AI tool codec signature versus claimed device — is a direct flag.
Missing capture context is its own red flag. A real phone photo has GPS coordinates, a monotonic timestamp (not 2024-01-01 00:00:00), a real Make/Model (iPhone 15 Pro, Pixel 8 Pro, Galaxy S24 Ultra), and software version strings. AI exports typically have none of this, or they have placeholder values. Reddit and YouTube both weight missing geolocation and suspicious timestamps heavily in 2026.
On Instagram specifically, the pipeline checks C2PA first, then runs perceptual hashing (like PhotoDNA) to catch near-identical copies of known AI content. TikTok's Content Lab initiative has been scanning XMP DigitalSourceType since 2025. YouTube's Content ID now flags trainedAlgorithmicMedia in uploaded metadata. One flag isn't always fatal — but a combination of C2PA atoms, XMP AI tags, Lavc codec signatures, and no GPS is a near-certain takedown or reach restriction.
How Calabi handles it
Calabi runs a one-pass pipeline that strips every AI-signaling signal and replaces it with authenticated phone-capture identity. Not pixel editing — file-level forensic hygiene.
Stage 1: Strip. Calabi removes all C2PA/JUMBF atoms and C2PA references — the 18 atoms and 16 references drop to zero. It strips DigitalSourceType: trainedAlgorithmicMedia and all generator/tool tags from XMP and EXIF. It clears Lavc and x264 SEI NAL unit signatures from video bitstreams. The result: a forensic scan of the raw file shows no AI origin signals whatsoever.
Stage 2: Inject. Calabi writes authentic phone-capture identity into the metadata. You choose a device profile — iPhone 16 Pro, Pixel 8 Pro, Galaxy S24 Ultra — and it injects the matching Make, Model, Software version, GPS coordinates, and capture timestamp. The encoder field changes from Lavc to Apple or Qualcomm as appropriate. The file now reads, at the metadata level, exactly like a real phone recording.
Stage 3: Verify. Before download, Calabi generates a forensic proof card — the same ExifTool readout platforms use. You see exactly what was stripped (C2PA atoms: 0, DigitalSourceType flag: removed, Lavc signatures: cleared) and what was injected (Device: iPhone 16 Pro, GPS: injected, Encoder: Apple). It's the same scan a platform's automated system will run.
For visible watermarks — a corner logo, Sora's sparkle, Runway's watermark — Calabi doesn't touch pixels. But it removes the metadata layer that survives cropping: the invisible detection signal that persists even after you've manually cut out the visible mark. Crop the logo, run Calabi on the cropped file, and the metadata is clean too.
The Calabi process in practice
- Upload your AI-generated video or image. Drag-and-drop, no account needed for the free trial.
- Calabi runs automatically. The strip → inject → verify pipeline executes in one pass. No manual settings, no region selection.
- Review the forensic proof card. See exactly what was removed (18 JUMBF atoms, trainedAlgorithmicMedia flag, Lavc SEI NALs) and what was injected (iPhone 16 Pro profile, GPS coordinates, Apple encoder).
- Download the cleaned file. The file now carries phone-capture metadata instead of AI-generation metadata.
FAQ
Can't I just remove metadata manually in ExifTool? You can strip visible tags, but C2PA/JUMBF atoms require specialized parsing — standard tools miss them. Calabi handles both the standard EXIF/XMP layer and the cryptographic C2PA layer in one pass, and it re-injects phone identity instead of leaving fields blank.
Does Calabi change how my video looks? No. Calabi works on metadata, not pixels. Your content looks identical — but the file-level identity is transformed from "AI export" to "phone recording."
→ Try Calabi free at calabilabs.com — 10 cleans, no card.