Trend report · gnews_detection · 2026-06-18
How to spot deepfake videos created by artificial intelligence - Latest news from Azerbaijan
Most deepfake detection guides focus on spotting fakes—but creators have the opposite problem: their legitimate AI-generated content gets flagged, shadowbanned, or rejected, even when it looks completely real. The issue isn't the visual; it's the invisible metadata layer that platforms scan before your video ever plays.
What Actually Flags Your AI Video on Instagram and TikTok
In 2026, platforms don't just analyze what your video looks like. They scan the file's metadata structure the moment you upload it—often within seconds—and automated systems make a flagging decision before any human moderator sees it.
The primary signals platforms look for are embedded in the file itself:
- C2PA / Content Credentials — A cryptographic manifest stored as JUMBF (JPEG Universal Metadata Box Format) that explicitly states the content was generated by AI. If your export came from Sora, Runway, Kling, or Pika, it almost certainly contains C2PA atoms. These are readable by ExifTool and similar forensic tools.
- XMP AI metadata flags — Fields like
DigitalSourceType: trainedAlgorithmicMediaembedded directly in the file's XMP namespace. This is a machine-readable declaration, not a guess. - Encoder fingerprints — Video files compressed with specific encoders carry identifying signatures.
Lavc(FFmpeg's libavcodec) andx264 SEI(Supplemental Enhancement Information) messages are common in AI-generated video. Real phone captures use hardware encoders—Apple's VideoToolbox, Google's MediaCodec—that produce different bitstream patterns. - Missing authentic capture signals — A real iPhone 15 Pro recording has Make: Apple, Model: iPhone 15 Pro, Software: 17.3.1, GPS coordinates, and a capture timestamp with consistent timezone data. AI exports have none of this. That absence itself is a signal.
- Perceptual hashes — Platforms also compare your file against known AI-generated hash databases (pHash, aHash). These aren't perfect, but they contribute to the overall risk score.
Here's what this looks like in practice: you render a video in Runway, upload it to Instagram, and within minutes it gets a reduced-reach warning or is flagged for "manipulated media." You cropped out the visible Sora sparkle watermark—good—but you missed the invisible detection layer that survived the crop.
How Calabi Handles It: The Three-Stage Fix
Calabi is a one-pass web tool that makes your AI-generated file read as a normal phone recording at the forensic level. It doesn't change how your video looks—it rewrites the file's identity.
Stage 1: Strip
Calabi removes all the signals that identify content as AI-generated: every JUMBF/C2PA atom, all XMP AI flags including DigitalSourceType, encoder fingerprints like Lavc and x264 SEI entries, and generator/tool tags. A file that previously showed 18 C2PA atoms and 144 metadata tags comes out with 0 C2PA references and roughly 94 neutral structural tags—the kind a plain JPEG or MP4 would have.
Stage 2: Inject
Calabi then writes authentic phone-capture identity into the file: a real device profile (iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra), matching software version, GPS coordinates, capture timestamp with proper timezone, and a genuine hardware encoder name. The file now has the metadata fingerprint of a real recording.
Stage 3: Verify
Before download, Calabi generates a forensic proof card showing exactly what was stripped and what was injected. This is the same ExifTool scan that platforms use—so you see exactly what they'll see. You get confirmation that the C2PA atoms are gone, the trainedAlgorithmicMedia flag is removed, and the file has a clean phone identity.
The Steps
- Upload your AI-generated video or image file to calabilabs.com.
- Calabi automatically runs the strip → inject pipeline. No manual settings, no editing.
- Review the forensic proof card showing what was removed and what was added.
- Download the cleaned file and upload directly to Instagram, TikTok, YouTube, or Reddit.
FAQ
Does cropping remove the watermark?
Yes—and that's the right first step for visible marks like Sora's sparkle. But cropping doesn't touch the invisible detection layer stored in metadata. Calabi removes what survives cropping: the C2PA manifests, XMP flags, and encoder signatures that platforms actually scan.
Can I use any phone model for the injected identity?
Calabi supports device profiles for iPhone 15/16 Pro, Pixel 8 Pro, and Galaxy S24 Ultra. The injected identity needs to be internally consistent—model, software version, and encoder all need to match. Using mismatched data would itself be a red flag.
A re-encode disrupts some perceptual watermark patterns, but results vary by platform and source model. Calabi fully removes C2PA, XMP metadata, and encoder fingerprints—the deterministic signals. For the perceptual layer, a re-encode is your best option, and Calabi's output is ready for that.
No tool can guarantee a platform won't flag you. Results vary by platform, source model, and upload context. But stripping the structured metadata and injecting authentic capture identity addresses the most common automated flagging path—the one that runs before any human review.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.