Trend report · gnews_detection · 2026-06-19
Russian and Chinese Influence Actors Use AI to Evade Bot Detection and Mimic Human Behavior - CyberSecurityNews
When influence actors use AI to generate content that looks human, platforms don't just react to what the content says — they scan the file itself for signs it was machine-made. In 2026, that scan runs deep: C2PA manifests, XMP metadata flags, encoder fingerprints, and the absence of typical phone-capture signals. If any of those signals are wrong or missing, your upload gets flagged or suppressed — even if nobody can see why.
What actually flags your file
Platforms like Instagram, TikTok, YouTube, and Reddit run automated forensic checks on every upload. They're not reading your caption — they're reading the metadata layer underneath.
C2PA and Content Credentials are the biggest trigger. When you export from Midjourney, Sora, Runway, or Kling, the file embeds a JUMBF (JPEG Universal Metadata Box Format) block containing a C2PA manifest. This manifest is a cryptographically signed record that says exactly which model generated the content, when, and with what parameters. Platforms parse this manifest in seconds. A file that left a C2PA atom saying digitalSourceType=trainedAlgorithmicMedia is an automatic flag — no human took that photo or video.
XMP metadata flags are the second layer. Adobe-added XMP packets include fields like xmpMM:CreatorTool, DIG:DigitalSourceType, and stf:SoftwareAgent. A raw AI export from Stable Diffusion typically carries 144 metadata tags. Many of those tags directly reference the generation tool. Platforms maintain blocklists of known AI tool signatures — when ExifTool or a platform's scanner finds generator, SDWebUI, or ComfyUI in the metadata, that's a flag.
Encoder fingerprints are subtler but equally damning. Video files carry codec metadata in the bitstream — H.264 and H.265 files include SEI (Supplemental Enhancement Information) messages. An AI video export from FFmpeg using libx264 or libx265 leaves encoder fingerprints like Lavc (FFmpeg's encoder library) and x264 core version numbers. Human phone recordings use hardware encoders — Apple QTX, OMX.qcom, or Samsung's MFC — which look nothing like a software transcoder. A video file that was generated, re-encoded, or processed by FFmpeg is readable in its structure, even if it looks fine visually.
Missing phone identity signals are the final tell. A real iPhone 15 Pro capture includes Make (Apple), Model (iPhone 15 Pro), SoftwareVersion (iOS 17.x), GPS coordinates, and a capture timestamp with the device's native timezone. It also carries specific EXIF tags: LensMake, LensModel, FocalLength, and ShutterSpeed that correspond to actual hardware. AI-generated files have none of this — no GPS, no device identity, and timestamps that either don't exist or are clearly fabricated.
Perceptual hashes like pHash or aHash compare your file against known AI-output fingerprints. If your video's hash clusters near known Sora or Kling outputs, it gets flagged regardless of metadata. This is why just stripping metadata isn't enough — the encoder structure and bitstream patterns have to match a real phone capture too.
How Calabi handles it
Calabi runs a three-stage pipeline that rewrites the file at the forensic level — not just the visible pixels. The result is a file that looks, to platform scanners, like it came off a real phone.
Stage 1 — Strip. Calabi removes every signal that identifies a file as AI-generated. That means nulling the JUMBF/C2PA atoms (18 atoms reduced to 0), stripping C2PA references entirely, removing XMP fields like DigitalSourceType and CreatorTool, and wiping encoder fingerprints from the video bitstream. The platform scanner looks for those signals and finds nothing.
Stage 2 — Inject. Calabi writes in authentic phone-capture identity. It assigns a real device profile — iPhone 15 Pro, Pixel 8 Pro, or Galaxy S24 Ultra — complete with Make, Model, Software version, GPS coordinates (approximate city-level), capture timestamp, and the correct hardware encoder name. A video gets the right bitstream structure: H.264 with a hardware encoder identifier instead of Lavc. The file now carries the EXIF fingerprint of a real device recording, not an AI pipeline.
Stage 3 — Verify. Before you download, Calabi runs an ExifTool scan — the same tool newsrooms and platform trust systems use — and shows you exactly what was stripped and what was injected. You see: C2PA atoms 18 → 0, C2PA references 16 → 0, trainedAlgorithmicMedia flag removed, metadata tags 144 → ~94 neutral structural tags. The forensic proof card is your record that the file passed through the full strip-and-inject cycle.
The steps
- Upload your AI-generated video or image to calabilabs.com. No account needed to start.
- Automatic pipeline runs. Strip C2PA and XMP flags, remove encoder fingerprints, inject phone device profile with GPS, timestamp, and hardware encoder signature.
- Review the forensic proof card. See exactly what was removed and what was injected — the same ExifTool output platforms use to scan uploads.
- Download the cleaned file. A 1:1 visual copy with forensic identity replaced. Upload to Instagram, TikTok, YouTube, or Reddit with a file that matches a real phone capture.
FAQ
Does removing the visible watermark Sora or Kling adds fix the flagging problem?
Partially. Cropping removes the visible sparkle or logo, but the metadata layer — C2PA manifests, XMP AI flags, encoder fingerprints — survives cropping. Platforms scan that metadata layer, not just the pixels. Calabi strips all of it so the file-level signal matches a real phone recording.
Re-encoding disrupts some perceptual watermark patterns, but results vary by platform and source model. Calabi fully removes the C2PA/metadata/encoder signals — the invisible pixel patterns are a separate layer with inconsistent removal guarantees.
Can I pick which phone profile gets injected?
Yes. Calabi supports iPhone 15 Pro, Pixel 8 Pro, and Galaxy S24 Ultra profiles. The device profile determines the EXIF tags, encoder name, and software version written into the file.
Platform scanners in 2026 are not guessing. They run ExifTool, parse C2PA manifests, and compare encoder fingerprints against known AI-output patterns. If your file carries any of those signals, it gets flagged — often within seconds of upload, before any human review. Stripping the metadata without replacing it leaves a file that looks like a camera without a memory card: structurally wrong. The only durable fix is replacing the forensic identity entirely.
→ Try Calabi free at calabilabs.com — 10 cleans, no card.